diff --git a/src/.vuepress/public/img/LargeModel-moirai2.png b/src/.vuepress/public/img/LargeModel-moirai2.png new file mode 100644 index 000000000..837a5a693 Binary files /dev/null and b/src/.vuepress/public/img/LargeModel-moirai2.png differ diff --git a/src/.vuepress/public/img/LargeModel-toto.png b/src/.vuepress/public/img/LargeModel-toto.png new file mode 100644 index 000000000..d6a7b9a35 Binary files /dev/null and b/src/.vuepress/public/img/LargeModel-toto.png differ diff --git a/src/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md b/src/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md index 907aa13b6..fc7d36253 100644 --- a/src/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md +++ b/src/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md @@ -417,7 +417,7 @@ State Flow Explanation: **Viewing Example:** ```SQL -IoTDB> SHOW MODELS +IoTDB> show models +---------------------+--------------+--------------+-------------+ | ModelId| ModelType| Category| State| +---------------------+--------------+--------------+-------------+ @@ -432,7 +432,12 @@ IoTDB> SHOW MODELS | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -440,22 +445,24 @@ IoTDB> SHOW MODELS | Model Name | Core Concept | Applicable Scenarios | Key Features | |-----------------------------------------|------------------------------------------------------------------------------|-----------------------------------------------------------|------------------------------------------------------------------------------| -| **ARIMA** (Autoregressive Integrated Moving Average) | Combines AR, differencing (I), and MA for stationary or differenced series | Univariate forecasting (stock prices, sales, economics) | 1. For linear trends with weak seasonality2. Requires (p,d,q) tuning3. Sensitive to missing values | -| **Holt-Winters** (Triple Exponential Smoothing) | Exponential smoothing with level, trend, and seasonal components | Data with clear trend & seasonality (monthly sales, power demand) | 1. Handles additive/multiplicative seasonality2. Weights recent data higher3. Simple implementation | -| **Exponential Smoothing** | Weighted average of history with exponentially decaying weights | Trending but non-seasonal data (short-term demand) | 1. Few parameters, simple computation2. Suitable for stable/slow-changing series3. Extensible to double/triple smoothing | -| **Naive Forecaster** | Uses last observation as next prediction (simplest baseline) | Benchmarking or data with no clear pattern | 1. No training needed2. Sensitive to sudden changes3. Seasonal variant uses prior season value | -| **STL Forecaster** | Decomposes series into trend, seasonal, residual; forecasts components | Complex seasonality/trends (climate, traffic) | 1. Handles non-fixed seasonality2. Robust to outliers3. Components can use other models | -| **Gaussian HMM** | Hidden states generate observations; each state follows Gaussian distribution | State sequence prediction/classification (speech, finance) | 1. Models temporal state transitions2. Observations independent per state3. Requires state count | -| **GMM HMM** | Extends Gaussian HMM; each state uses Gaussian Mixture Model | Multi-modal observation scenarios (motion recognition, biosignals) | 1. More flexible than single Gaussian2. Higher complexity3. Requires GMM component count | -| **STRAY** (Search for Outliers using Random Projection and Adaptive Thresholding) | Uses SVD to detect anomalies in high-dimensional time series | High-dimensional anomaly detection (sensor networks, IT monitoring) | 1. No distribution assumption2. Handles high dimensions3. Sensitive to global anomalies | +| **ARIMA** (Autoregressive Integrated Moving Average) | Combines AR, differencing (I), and MA for stationary or differenced series | Univariate forecasting (stock prices, sales, economics) | 1. For linear trends with weak seasonality
2. Requires (p,d,q) tuning
3. Sensitive to missing values | +| **Holt-Winters** (Triple Exponential Smoothing) | Exponential smoothing with level, trend, and seasonal components | Data with clear trend & seasonality (monthly sales, power demand) | 1. Handles additive/multiplicative seasonality
2. Weights recent data higher
3. Simple implementation | +| **Exponential Smoothing** | Weighted average of history with exponentially decaying weights | Trending but non-seasonal data (short-term demand) | 1. Few parameters, simple computation
2. Suitable for stable/slow-changing series
3. Extensible to double/triple smoothing | +| **Naive Forecaster** | Uses last observation as next prediction (simplest baseline) | Benchmarking or data with no clear pattern | 1. No training needed
2. Sensitive to sudden changes
3. Seasonal variant uses prior season value | +| **STL Forecaster** | Decomposes series into trend, seasonal, residual; forecasts components | Complex seasonality/trends (climate, traffic) | 1. Handles non-fixed seasonality
2. Robust to outliers
3. Components can use other models | +| **Gaussian HMM** | Hidden states generate observations; each state follows Gaussian distribution | State sequence prediction/classification (speech, finance) | 1. Models temporal state transitions
2. Observations independent per state
3. Requires state count | +| **GMM HMM** | Extends Gaussian HMM; each state uses Gaussian Mixture Model | Multi-modal observation scenarios (motion recognition, biosignals) | 1. More flexible than single Gaussian
2. Higher complexity
3. Requires GMM component count | +| **STRAY** (Search for Outliers using Random Projection and Adaptive Thresholding) | Uses SVD to detect anomalies in high-dimensional time series | High-dimensional anomaly detection (sensor networks, IT monitoring) | 1. No distribution assumption
2. Handles high dimensions
3. Sensitive to global anomalies | **Built-in Time Series Large Models:** | Model Name | Core Concept | Applicable Scenarios | Key Features | |-----------------|------------------------------------------------------------------------------|-----------------------------------------------------------|------------------------------------------------------------------------------| -| **Timer-XL** | Long-context time series large model pretrained on massive industrial data | Complex industrial forecasting requiring ultra-long history (energy, aerospace, transport) | 1. Supports input of tens of thousands of time points2. Covers non-stationary, multivariate, and covariate scenarios3. Pretrained on trillion-scale high-quality industrial IoT data | -| **Timer-Sundial** | Generative foundation model with "Transformer + TimeFlow" architecture | Zero-shot forecasting requiring uncertainty quantification (finance, supply chain, renewable energy) | 1. Strong zero-shot generalization; supports point & probabilistic forecasting2. Flexible analysis of any prediction distribution statistic3. Innovative flow-matching architecture for efficient non-deterministic sample generation | -| **Chronos-2** | Universal time series foundation model based on discrete tokenization | Rapid zero-shot univariate forecasting; scenarios enhanced by covariates (promotions, weather) | 1. Powerful zero-shot probabilistic forecasting2. Unified multi-variable & covariate modeling (strict input requirements):  a. Future covariate names ⊆ historical covariate names  b. Each historical covariate length = target length  c. Each future covariate length = prediction length3. Efficient encoder-only structure balancing performance and speed | +| **Timer-XL** | Long-context time series large model pretrained on massive industrial data | Complex industrial forecasting requiring ultra-long history (energy, aerospace, transport) | 1. Supports input of tens of thousands of time points
2. Covers non-stationary, multivariate, and covariate scenarios
3. Pretrained on trillion-scale high-quality industrial IoT data | +| **Timer-Sundial** | Generative foundation model with "Transformer + TimeFlow" architecture | Zero-shot forecasting requiring uncertainty quantification (finance, supply chain, renewable energy) | 1. Strong zero-shot generalization; supports point & probabilistic forecasting
2. Flexible analysis of any prediction distribution statistic
3. Innovative flow-matching architecture for efficient non-deterministic sample generation | +| **Chronos-2** | Universal time series foundation model based on discrete tokenization | Rapid zero-shot univariate forecasting; scenarios enhanced by covariates (promotions, weather) | 1. Powerful zero-shot probabilistic forecasting
2. Unified multi-variable & covariate modeling (strict input requirements):
  a. Future covariate names ⊆ historical covariate names
  b. Each historical covariate length = target length
  c. Each future covariate length = prediction length
3. Efficient encoder-only structure balancing performance and speed | +| **Moirai 2.0** | Lightweight decoder-only Patch Transformer with a single patch size, multi-token prediction, and multi-quantile outputs | Zero-shot univariate forecasting where model size and inference efficiency are important, such as industrial monitoring, energy load, and equipment metrics | 1. Approximately 11.4M parameters
2. Predicts multiple patches per decoding step to reduce autoregressive overhead for long horizons
3. Outputs nine quantiles (0.1–0.9) and uses the p50 median as the point forecast
4. Uses instance normalization to mitigate distribution shift across series
5. Does not support multivariate targets or covariates | +| **Toto 2.0** | Decoder-only Patch Transformer alternating causal temporal attention and variable attention to jointly model temporal and variable dimensions | Zero-shot multivariate forecasting for observability metrics, including joint forecasting of CPU, memory, and network traffic | 1. Supports univariate and multivariate target forecasting
2. Outputs fixed quantiles from 0.1 to 0.9 and uses the p50 median as the point forecast
3. Supports cached block decoding for efficient scaling to longer forecast horizons
4. Covariates are not currently supported | ### 4.4 Deleting Models @@ -656,4 +663,4 @@ To add a new built-in custom model to AINode (using Chronos2 as example): ``` * **Complete Example** - Reference implementation: https://github.com/apache/iotdb/pull/16903 \ No newline at end of file + Reference implementation: https://github.com/apache/iotdb/pull/16903 diff --git a/src/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index bc20138a4..284198ab6 100644 --- a/src/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -46,7 +46,7 @@ The Timer model (non-built-in model) not only demonstrates excellent few-shot ge ## 4. Timer-XL Model -Timer-XL further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: +Timer-XL[2] further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: * **Ultra-Long Context Support**: This model breaks through the limitations of traditional time series forecasting models, supporting the processing of inputs with thousands of Tokens (equivalent to tens of thousands of time points), effectively solving the context length bottleneck problem. * **Coverage of Multi-Variable Forecasting Scenarios**: Supports various forecasting scenarios, including the prediction of non-stationary time series, multi-variable prediction tasks, and predictions involving covariates, meeting diversified business needs. @@ -56,7 +56,7 @@ Timer-XL further extends and upgrades the network structure based on Timer, achi ## 5. Timer-Sundial Model -Timer-Sundial is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: +Timer-Sundial[3] is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: * **Strong Generalization Performance**: Possesses zero-shot forecasting capabilities and can support both point forecasting and probabilistic forecasting simultaneously. * **Flexible Prediction Distribution Analysis**: Not only can it predict means or quantiles, but it can also evaluate any statistical properties of the prediction distribution through the raw samples generated by the model. @@ -66,7 +66,7 @@ Timer-Sundial is a series of generative foundational models focused on time seri ## 6. Chronos-2 Model -Chronos-2 is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: +Chronos-2[4] is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: * **Probabilistic Forecasting Capability**: The model outputs multi-step prediction results in a generative manner, supporting quantile or distribution-level forecasting to characterize future uncertainty. * **Zero-Shot General Forecasting**: Leveraging the contextual learning ability acquired through pre-training, it can directly execute forecasting on unseen datasets without retraining or parameter updates. @@ -78,7 +78,34 @@ Chronos-2 is a universal time series foundational model developed by the Amazon ![](/img/timeseries-large-model-chronos2.png) -## 7. Performance Showcase +## 7. Moirai2 Model + +Moirai2[5] (Moirai 2.0) is a general-purpose time series foundation model developed by Salesforce AI Research (supported in V2.0.10 and later). AINode currently integrates the Moirai 2.0 R-small variant, which has approximately 11.4M parameters. Unlike Moirai 1.0, which uses a masked encoder architecture, Moirai 2.0 uses a causal decoder-only Patch Transformer. With a single patch size, multi-token prediction, and multi-quantile outputs, it provides efficient univariate forecasting with a compact model. Its core features include: + +- **Lightweight Model Architecture**: Uses a decoder-only Patch Transformer with RMSNorm, rotary positional embeddings, and SiLU-GLU feed-forward networks to balance forecasting capability and inference efficiency at a small parameter scale. +- **Multi-Token Prediction**: Predicts multiple patches at each decoding step, reducing the number of autoregressive decoding steps required for long forecast horizons. +- **Probabilistic Forecasting**: Outputs nine quantiles from 0.1 to 0.9. AINode uses the p50 median as the point forecast. +- **Patch Decoding**: Groups the time series into fixed-size patches before the attention module to improve temporal feature extraction and decoding efficiency. +- **Instance Normalization**: Standardizes each time series before model input and applies denormalization after output to mitigate distribution shifts across series. +- **Input Scope**: Focuses on univariate forecasting and does not support multivariate targets or covariates. + +![](/img/LargeModel-moirai2.png) + +> Note: The Moirai 2.0 R-small model weights are licensed under CC BY-NC 4.0 and are restricted to research use. + +## 8. Toto Model + +Toto[6] (Toto 2.0) is a next-generation time series foundation model developed by Datadog (supported in V2.0.10 and later), primarily for forecasting in observability scenarios. AINode currently integrates the 2.5B-parameter variant. It is based on a decoder-only Patch Transformer architecture that alternates causal temporal attention and variable attention to jointly model the temporal and variable dimensions. Its core features include: + +- **Univariate and Multivariate Forecasting**: Supports both individual target variables and joint forecasting of multiple related target variables, making it suitable for observability metrics such as CPU, memory, and network traffic. +- **Probabilistic Forecasting**: Outputs fixed quantiles from 0.1 to 0.9 to represent forecasting uncertainty. AINode uses the p50 median as the point forecast. +- **Efficient Block Decoding**: Uses cached block decoding to generate forecasts in blocks, reducing repeated computation for longer forecast horizons. +- **Large-Scale Model Capability**: Improves forecasting performance by scaling model parameters and pre-training data, providing strong zero-shot generalization. +- **Covariate Limitation**: The currently integrated version does not support historical covariates or known future covariates. + +![](/img/LargeModel-toto.png) + +## 9. Performance Showcase Time Series Large Models can adapt to real time series data from various different domains and scenarios, demonstrating excellent processing capabilities across various tasks. The following shows the actual performance on different datasets: @@ -100,7 +127,7 @@ Using Time Series Large Models to accurately identify outliers that deviate sign ![](/img/LargeModel05.png) -## 8. Deployment and Usage +## 10. Deployment and Usage 1. Open the IoTDB CLI console and check that the ConfigNode, DataNode, and AINode nodes are all Running. @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer: Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() -**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back](#ref4) -**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back]() +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ Back](#ref5) -**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back]() \ No newline at end of file +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ Back](#ref6) diff --git a/src/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md b/src/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md index 259417fbc..0f614b413 100644 --- a/src/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md +++ b/src/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md @@ -416,7 +416,12 @@ IoTDB> show models | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -424,22 +429,24 @@ Built-in traditional time series model introduction: | Model Name | Core Concept | Applicable Scenario | Main Features | |----------------------------------| ----------------------------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------- | -| **ARIMA** (AutoRegressive Integrated Moving Average) | Combines autoregression (AR), differencing (I), and moving average (MA), used for predicting stationary time series or data that can be made stationary through differencing. | Univariate time series prediction, such as stock prices, sales, economic indicators. | 1. Suitable for linear trends and weak seasonality data. 2. Requires selecting parameters (p,d,q). 3. Sensitive to missing values. | -| **Holt-Winters** (Three-parameter exponential smoothing) | Based on exponential smoothing, introduces level, trend, and seasonal components, suitable for data with trend and seasonality. | Time series with clear seasonality and trend, such as monthly sales, power demand. | 1. Can handle additive or multiplicative seasonality. 2. Gives higher weight to recent data. 3. Simple to implement. | -| **Exponential Smoothing** (Exponential smoothing) | Uses weighted average of historical data, with weights decreasing exponentially over time, emphasizing the importance of recent observations. | Data without significant seasonality but with trend, such as short-term demand prediction. | 1. Few parameters, simple calculation. 2. Suitable for stationary or slowly changing sequences. 3. Can be extended to double or triple exponential smoothing. | -| **Naive Forecaster** (Naive predictor) | Uses the observation of the most recent period as the prediction for the next period, the simplest baseline model. | As a benchmark for other models or simple prediction when data has no obvious pattern. | 1. No training required. 2. Sensitive to sudden changes. 3. Seasonal naive variant can use the same period of the previous season to predict. | -| **STL Forecaster** (Seasonal-Trend Decomposition Forecast) | Based on STL decomposition of time series, predicts trend, seasonal, and residual components separately, then combines them. | Data with complex seasonality, trend, and non-linear patterns, such as climate data, traffic flow. | 1. Can handle non-fixed seasonality. 2. Robust to outliers. 3. After decomposition, other models can be combined to predict each component. | -| **Gaussian HMM** (Gaussian Hidden Markov Model) | Assumes observed data is generated by hidden states, with each state's observation probability following a Gaussian distribution. | State sequence prediction or classification, such as speech recognition, financial state identification. | 1. Suitable for time series state modeling. 2. Assumes observations are independent given the state. 3. Requires specifying the number of hidden states. | -| **GMM HMM** (Gaussian Mixture Hidden Markov Model) | An extension of Gaussian HMM, where each state's observation probability is described by a Gaussian Mixture Model, capturing more complex observation distributions. | Scenarios requiring multi-modal observation distributions, such as complex action recognition, biosignal analysis. | 1. More flexible than single Gaussian. 2. More parameters, higher computational complexity. 3. Requires training the number of GMM components. | -| **STRAY** (Anomaly Detection based on Singular Value Decomposition) | Detects anomalies in high-dimensional data through Singular Value Decomposition (SVD), commonly used for time series anomaly detection. | High-dimensional time series anomaly detection, such as sensor networks, IT system monitoring. | 1. No distribution assumption required. 2. Can handle high-dimensional data. 3. Sensitive to global anomalies, may miss local anomalies. | +| **ARIMA** (AutoRegressive Integrated Moving Average) | Combines autoregression (AR), differencing (I), and moving average (MA), used for predicting stationary time series or data that can be made stationary through differencing. | Univariate time series prediction, such as stock prices, sales, economic indicators. | 1. Suitable for linear trends and weak seasonality data.
2. Requires selecting parameters (p,d,q).
3. Sensitive to missing values. | +| **Holt-Winters** (Three-parameter exponential smoothing) | Based on exponential smoothing, introduces level, trend, and seasonal components, suitable for data with trend and seasonality. | Time series with clear seasonality and trend, such as monthly sales, power demand. | 1. Can handle additive or multiplicative seasonality.
2. Gives higher weight to recent data.
3. Simple to implement. | +| **Exponential Smoothing** (Exponential smoothing) | Uses weighted average of historical data, with weights decreasing exponentially over time, emphasizing the importance of recent observations. | Data without significant seasonality but with trend, such as short-term demand prediction. | 1. Few parameters, simple calculation.
2. Suitable for stationary or slowly changing sequences.
3. Can be extended to double or triple exponential smoothing. | +| **Naive Forecaster** (Naive predictor) | Uses the observation of the most recent period as the prediction for the next period, the simplest baseline model. | As a benchmark for other models or simple prediction when data has no obvious pattern. | 1. No training required.
2. Sensitive to sudden changes.
3. Seasonal naive variant can use the same period of the previous season to predict. | +| **STL Forecaster** (Seasonal-Trend Decomposition Forecast) | Based on STL decomposition of time series, predicts trend, seasonal, and residual components separately, then combines them. | Data with complex seasonality, trend, and non-linear patterns, such as climate data, traffic flow. | 1. Can handle non-fixed seasonality.
2. Robust to outliers.
3. After decomposition, other models can be combined to predict each component. | +| **Gaussian HMM** (Gaussian Hidden Markov Model) | Assumes observed data is generated by hidden states, with each state's observation probability following a Gaussian distribution. | State sequence prediction or classification, such as speech recognition, financial state identification. | 1. Suitable for time series state modeling.
2. Assumes observations are independent given the state.
3. Requires specifying the number of hidden states. | +| **GMM HMM** (Gaussian Mixture Hidden Markov Model) | An extension of Gaussian HMM, where each state's observation probability is described by a Gaussian Mixture Model, capturing more complex observation distributions. | Scenarios requiring multi-modal observation distributions, such as complex action recognition, biosignal analysis. | 1. More flexible than single Gaussian.
2. More parameters, higher computational complexity.
3. Requires training the number of GMM components. | +| **STRAY** (Anomaly Detection based on Singular Value Decomposition) | Detects anomalies in high-dimensional data through Singular Value Decomposition (SVD), commonly used for time series anomaly detection. | High-dimensional time series anomaly detection, such as sensor networks, IT system monitoring. | 1. No distribution assumption required.
2. Can handle high-dimensional data.
3. Sensitive to global anomalies, may miss local anomalies. | Built-in time series large model introduction: | Model Name | Core Concept | Applicable Scenario | Main Features | |---------------| ---------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| **Timer-XL** | Time series large model supporting ultra-long context, enhancing generalization capability through large-scale industrial data pre-training. | Complex industrial prediction requiring extremely long historical data, such as energy, aerospace, and transportation. | 1. Ultra-long context support, can handle tens of thousands of time points as input. 2. Multi-scenario coverage, supports non-stationary, multi-variable, and covariate prediction. 3. Pre-trained on trillions of high-quality industrial time series data. | -| **Timer-Sundial** | A generative foundational model based on "Transformer + TimeFlow" architecture, focusing on probabilistic prediction. | Zero-shot prediction scenarios requiring quantification of uncertainty, such as finance, supply chain, and new energy power generation. | 1. Strong zero-shot generalization capability, supports point prediction and probabilistic prediction. 2. Can flexibly analyze any statistical properties of the prediction distribution. 3. Innovative generative architecture, achieving efficient non-deterministic sample generation. | -| **Chronos-2** | A general time series foundational model based on discrete tokenization paradigm, converting prediction into language modeling tasks. | Rapid zero-shot univariate prediction, and scenarios that can leverage covariates (e.g., promotions, weather) to improve results. | 1. Strong zero-shot probabilistic prediction capability. 2. Supports unified covariate modeling, but has strict input requirements: a. The set of names of future covariates must be a subset of the set of names of historical covariates; b. The length of each historical covariate must equal the length of the target variable; c. The length of each future covariate must equal the prediction length; 3. Uses an efficient encoder-style structure, balancing performance and inference speed. | +| **Timer-XL** | Time series large model supporting ultra-long context, enhancing generalization capability through large-scale industrial data pre-training. | Complex industrial prediction requiring extremely long historical data, such as energy, aerospace, and transportation. | 1. Ultra-long context support, can handle tens of thousands of time points as input.
2. Multi-scenario coverage, supports non-stationary, multi-variable, and covariate prediction.
3. Pre-trained on trillions of high-quality industrial time series data. | +| **Timer-Sundial** | A generative foundational model based on "Transformer + TimeFlow" architecture, focusing on probabilistic prediction. | Zero-shot prediction scenarios requiring quantification of uncertainty, such as finance, supply chain, and new energy power generation. | 1. Strong zero-shot generalization capability, supports point prediction and probabilistic prediction.
2. Can flexibly analyze any statistical properties of the prediction distribution.
3. Innovative generative architecture, achieving efficient non-deterministic sample generation. | +| **Chronos-2** | A general time series foundational model based on discrete tokenization paradigm, converting prediction into language modeling tasks. | Rapid zero-shot univariate prediction, and scenarios that can leverage covariates (e.g., promotions, weather) to improve results. | 1. Strong zero-shot probabilistic prediction capability.
2. Supports unified covariate modeling, but has strict input requirements:
  a. The set of names of future covariates must be a subset of the set of names of historical covariates;
  b. The length of each historical covariate must equal the length of the target variable;
  c. The length of each future covariate must equal the prediction length;
3. Uses an efficient encoder-style structure, balancing performance and inference speed. | +| **Moirai 2.0** | Uses a lightweight decoder-only Patch Transformer with a single patch size, multi-token prediction, and multi-quantile outputs for efficient univariate forecasting. | Zero-shot univariate forecasting where model size and inference efficiency are important, such as industrial monitoring, energy load, and equipment metric forecasting. | 1. Approximately 11.4M parameters.
2. Predicts multiple patches per decoding step to reduce autoregressive overhead for long horizons.
3. Outputs nine quantiles (0.1–0.9) and uses the p50 median as the point forecast.
4. Uses instance normalization to mitigate distribution shift across series.
5. Does not support multivariate targets or covariates. | +| **Toto 2.0** | Uses a decoder-only Patch Transformer that alternates causal temporal attention and variable attention to jointly model the temporal and variable dimensions. | Zero-shot forecasting for multivariate time series such as observability metrics, including joint forecasting of CPU, memory, and network traffic. | 1. Supports univariate and multivariate target forecasting.
2. Outputs fixed quantiles from 0.1 to 0.9 and uses the p50 median as the point forecast.
3. Supports cached block decoding for efficient scaling to longer forecast horizons.
4. Covariates are not currently supported. | ### 4.4 Delete Models @@ -675,4 +682,4 @@ Support adding new built-in models to AINode. The specific steps are as follows * **Full Example** -Full example can be referenced at https://github.com/apache/iotdb/pull/16903 \ No newline at end of file +Full example can be referenced at https://github.com/apache/iotdb/pull/16903 diff --git a/src/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index bc20138a4..284198ab6 100644 --- a/src/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -46,7 +46,7 @@ The Timer model (non-built-in model) not only demonstrates excellent few-shot ge ## 4. Timer-XL Model -Timer-XL further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: +Timer-XL[2] further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: * **Ultra-Long Context Support**: This model breaks through the limitations of traditional time series forecasting models, supporting the processing of inputs with thousands of Tokens (equivalent to tens of thousands of time points), effectively solving the context length bottleneck problem. * **Coverage of Multi-Variable Forecasting Scenarios**: Supports various forecasting scenarios, including the prediction of non-stationary time series, multi-variable prediction tasks, and predictions involving covariates, meeting diversified business needs. @@ -56,7 +56,7 @@ Timer-XL further extends and upgrades the network structure based on Timer, achi ## 5. Timer-Sundial Model -Timer-Sundial is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: +Timer-Sundial[3] is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: * **Strong Generalization Performance**: Possesses zero-shot forecasting capabilities and can support both point forecasting and probabilistic forecasting simultaneously. * **Flexible Prediction Distribution Analysis**: Not only can it predict means or quantiles, but it can also evaluate any statistical properties of the prediction distribution through the raw samples generated by the model. @@ -66,7 +66,7 @@ Timer-Sundial is a series of generative foundational models focused on time seri ## 6. Chronos-2 Model -Chronos-2 is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: +Chronos-2[4] is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: * **Probabilistic Forecasting Capability**: The model outputs multi-step prediction results in a generative manner, supporting quantile or distribution-level forecasting to characterize future uncertainty. * **Zero-Shot General Forecasting**: Leveraging the contextual learning ability acquired through pre-training, it can directly execute forecasting on unseen datasets without retraining or parameter updates. @@ -78,7 +78,34 @@ Chronos-2 is a universal time series foundational model developed by the Amazon ![](/img/timeseries-large-model-chronos2.png) -## 7. Performance Showcase +## 7. Moirai2 Model + +Moirai2[5] (Moirai 2.0) is a general-purpose time series foundation model developed by Salesforce AI Research (supported in V2.0.10 and later). AINode currently integrates the Moirai 2.0 R-small variant, which has approximately 11.4M parameters. Unlike Moirai 1.0, which uses a masked encoder architecture, Moirai 2.0 uses a causal decoder-only Patch Transformer. With a single patch size, multi-token prediction, and multi-quantile outputs, it provides efficient univariate forecasting with a compact model. Its core features include: + +- **Lightweight Model Architecture**: Uses a decoder-only Patch Transformer with RMSNorm, rotary positional embeddings, and SiLU-GLU feed-forward networks to balance forecasting capability and inference efficiency at a small parameter scale. +- **Multi-Token Prediction**: Predicts multiple patches at each decoding step, reducing the number of autoregressive decoding steps required for long forecast horizons. +- **Probabilistic Forecasting**: Outputs nine quantiles from 0.1 to 0.9. AINode uses the p50 median as the point forecast. +- **Patch Decoding**: Groups the time series into fixed-size patches before the attention module to improve temporal feature extraction and decoding efficiency. +- **Instance Normalization**: Standardizes each time series before model input and applies denormalization after output to mitigate distribution shifts across series. +- **Input Scope**: Focuses on univariate forecasting and does not support multivariate targets or covariates. + +![](/img/LargeModel-moirai2.png) + +> Note: The Moirai 2.0 R-small model weights are licensed under CC BY-NC 4.0 and are restricted to research use. + +## 8. Toto Model + +Toto[6] (Toto 2.0) is a next-generation time series foundation model developed by Datadog (supported in V2.0.10 and later), primarily for forecasting in observability scenarios. AINode currently integrates the 2.5B-parameter variant. It is based on a decoder-only Patch Transformer architecture that alternates causal temporal attention and variable attention to jointly model the temporal and variable dimensions. Its core features include: + +- **Univariate and Multivariate Forecasting**: Supports both individual target variables and joint forecasting of multiple related target variables, making it suitable for observability metrics such as CPU, memory, and network traffic. +- **Probabilistic Forecasting**: Outputs fixed quantiles from 0.1 to 0.9 to represent forecasting uncertainty. AINode uses the p50 median as the point forecast. +- **Efficient Block Decoding**: Uses cached block decoding to generate forecasts in blocks, reducing repeated computation for longer forecast horizons. +- **Large-Scale Model Capability**: Improves forecasting performance by scaling model parameters and pre-training data, providing strong zero-shot generalization. +- **Covariate Limitation**: The currently integrated version does not support historical covariates or known future covariates. + +![](/img/LargeModel-toto.png) + +## 9. Performance Showcase Time Series Large Models can adapt to real time series data from various different domains and scenarios, demonstrating excellent processing capabilities across various tasks. The following shows the actual performance on different datasets: @@ -100,7 +127,7 @@ Using Time Series Large Models to accurately identify outliers that deviate sign ![](/img/LargeModel05.png) -## 8. Deployment and Usage +## 10. Deployment and Usage 1. Open the IoTDB CLI console and check that the ConfigNode, DataNode, and AINode nodes are all Running. @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer: Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() -**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back](#ref4) -**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back]() +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ Back](#ref5) -**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back]() \ No newline at end of file +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ Back](#ref6) diff --git a/src/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md b/src/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md index 907aa13b6..fc7d36253 100644 --- a/src/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md +++ b/src/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md @@ -417,7 +417,7 @@ State Flow Explanation: **Viewing Example:** ```SQL -IoTDB> SHOW MODELS +IoTDB> show models +---------------------+--------------+--------------+-------------+ | ModelId| ModelType| Category| State| +---------------------+--------------+--------------+-------------+ @@ -432,7 +432,12 @@ IoTDB> SHOW MODELS | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -440,22 +445,24 @@ IoTDB> SHOW MODELS | Model Name | Core Concept | Applicable Scenarios | Key Features | |-----------------------------------------|------------------------------------------------------------------------------|-----------------------------------------------------------|------------------------------------------------------------------------------| -| **ARIMA** (Autoregressive Integrated Moving Average) | Combines AR, differencing (I), and MA for stationary or differenced series | Univariate forecasting (stock prices, sales, economics) | 1. For linear trends with weak seasonality2. Requires (p,d,q) tuning3. Sensitive to missing values | -| **Holt-Winters** (Triple Exponential Smoothing) | Exponential smoothing with level, trend, and seasonal components | Data with clear trend & seasonality (monthly sales, power demand) | 1. Handles additive/multiplicative seasonality2. Weights recent data higher3. Simple implementation | -| **Exponential Smoothing** | Weighted average of history with exponentially decaying weights | Trending but non-seasonal data (short-term demand) | 1. Few parameters, simple computation2. Suitable for stable/slow-changing series3. Extensible to double/triple smoothing | -| **Naive Forecaster** | Uses last observation as next prediction (simplest baseline) | Benchmarking or data with no clear pattern | 1. No training needed2. Sensitive to sudden changes3. Seasonal variant uses prior season value | -| **STL Forecaster** | Decomposes series into trend, seasonal, residual; forecasts components | Complex seasonality/trends (climate, traffic) | 1. Handles non-fixed seasonality2. Robust to outliers3. Components can use other models | -| **Gaussian HMM** | Hidden states generate observations; each state follows Gaussian distribution | State sequence prediction/classification (speech, finance) | 1. Models temporal state transitions2. Observations independent per state3. Requires state count | -| **GMM HMM** | Extends Gaussian HMM; each state uses Gaussian Mixture Model | Multi-modal observation scenarios (motion recognition, biosignals) | 1. More flexible than single Gaussian2. Higher complexity3. Requires GMM component count | -| **STRAY** (Search for Outliers using Random Projection and Adaptive Thresholding) | Uses SVD to detect anomalies in high-dimensional time series | High-dimensional anomaly detection (sensor networks, IT monitoring) | 1. No distribution assumption2. Handles high dimensions3. Sensitive to global anomalies | +| **ARIMA** (Autoregressive Integrated Moving Average) | Combines AR, differencing (I), and MA for stationary or differenced series | Univariate forecasting (stock prices, sales, economics) | 1. For linear trends with weak seasonality
2. Requires (p,d,q) tuning
3. Sensitive to missing values | +| **Holt-Winters** (Triple Exponential Smoothing) | Exponential smoothing with level, trend, and seasonal components | Data with clear trend & seasonality (monthly sales, power demand) | 1. Handles additive/multiplicative seasonality
2. Weights recent data higher
3. Simple implementation | +| **Exponential Smoothing** | Weighted average of history with exponentially decaying weights | Trending but non-seasonal data (short-term demand) | 1. Few parameters, simple computation
2. Suitable for stable/slow-changing series
3. Extensible to double/triple smoothing | +| **Naive Forecaster** | Uses last observation as next prediction (simplest baseline) | Benchmarking or data with no clear pattern | 1. No training needed
2. Sensitive to sudden changes
3. Seasonal variant uses prior season value | +| **STL Forecaster** | Decomposes series into trend, seasonal, residual; forecasts components | Complex seasonality/trends (climate, traffic) | 1. Handles non-fixed seasonality
2. Robust to outliers
3. Components can use other models | +| **Gaussian HMM** | Hidden states generate observations; each state follows Gaussian distribution | State sequence prediction/classification (speech, finance) | 1. Models temporal state transitions
2. Observations independent per state
3. Requires state count | +| **GMM HMM** | Extends Gaussian HMM; each state uses Gaussian Mixture Model | Multi-modal observation scenarios (motion recognition, biosignals) | 1. More flexible than single Gaussian
2. Higher complexity
3. Requires GMM component count | +| **STRAY** (Search for Outliers using Random Projection and Adaptive Thresholding) | Uses SVD to detect anomalies in high-dimensional time series | High-dimensional anomaly detection (sensor networks, IT monitoring) | 1. No distribution assumption
2. Handles high dimensions
3. Sensitive to global anomalies | **Built-in Time Series Large Models:** | Model Name | Core Concept | Applicable Scenarios | Key Features | |-----------------|------------------------------------------------------------------------------|-----------------------------------------------------------|------------------------------------------------------------------------------| -| **Timer-XL** | Long-context time series large model pretrained on massive industrial data | Complex industrial forecasting requiring ultra-long history (energy, aerospace, transport) | 1. Supports input of tens of thousands of time points2. Covers non-stationary, multivariate, and covariate scenarios3. Pretrained on trillion-scale high-quality industrial IoT data | -| **Timer-Sundial** | Generative foundation model with "Transformer + TimeFlow" architecture | Zero-shot forecasting requiring uncertainty quantification (finance, supply chain, renewable energy) | 1. Strong zero-shot generalization; supports point & probabilistic forecasting2. Flexible analysis of any prediction distribution statistic3. Innovative flow-matching architecture for efficient non-deterministic sample generation | -| **Chronos-2** | Universal time series foundation model based on discrete tokenization | Rapid zero-shot univariate forecasting; scenarios enhanced by covariates (promotions, weather) | 1. Powerful zero-shot probabilistic forecasting2. Unified multi-variable & covariate modeling (strict input requirements):  a. Future covariate names ⊆ historical covariate names  b. Each historical covariate length = target length  c. Each future covariate length = prediction length3. Efficient encoder-only structure balancing performance and speed | +| **Timer-XL** | Long-context time series large model pretrained on massive industrial data | Complex industrial forecasting requiring ultra-long history (energy, aerospace, transport) | 1. Supports input of tens of thousands of time points
2. Covers non-stationary, multivariate, and covariate scenarios
3. Pretrained on trillion-scale high-quality industrial IoT data | +| **Timer-Sundial** | Generative foundation model with "Transformer + TimeFlow" architecture | Zero-shot forecasting requiring uncertainty quantification (finance, supply chain, renewable energy) | 1. Strong zero-shot generalization; supports point & probabilistic forecasting
2. Flexible analysis of any prediction distribution statistic
3. Innovative flow-matching architecture for efficient non-deterministic sample generation | +| **Chronos-2** | Universal time series foundation model based on discrete tokenization | Rapid zero-shot univariate forecasting; scenarios enhanced by covariates (promotions, weather) | 1. Powerful zero-shot probabilistic forecasting
2. Unified multi-variable & covariate modeling (strict input requirements):
  a. Future covariate names ⊆ historical covariate names
  b. Each historical covariate length = target length
  c. Each future covariate length = prediction length
3. Efficient encoder-only structure balancing performance and speed | +| **Moirai 2.0** | Lightweight decoder-only Patch Transformer with a single patch size, multi-token prediction, and multi-quantile outputs | Zero-shot univariate forecasting where model size and inference efficiency are important, such as industrial monitoring, energy load, and equipment metrics | 1. Approximately 11.4M parameters
2. Predicts multiple patches per decoding step to reduce autoregressive overhead for long horizons
3. Outputs nine quantiles (0.1–0.9) and uses the p50 median as the point forecast
4. Uses instance normalization to mitigate distribution shift across series
5. Does not support multivariate targets or covariates | +| **Toto 2.0** | Decoder-only Patch Transformer alternating causal temporal attention and variable attention to jointly model temporal and variable dimensions | Zero-shot multivariate forecasting for observability metrics, including joint forecasting of CPU, memory, and network traffic | 1. Supports univariate and multivariate target forecasting
2. Outputs fixed quantiles from 0.1 to 0.9 and uses the p50 median as the point forecast
3. Supports cached block decoding for efficient scaling to longer forecast horizons
4. Covariates are not currently supported | ### 4.4 Deleting Models @@ -656,4 +663,4 @@ To add a new built-in custom model to AINode (using Chronos2 as example): ``` * **Complete Example** - Reference implementation: https://github.com/apache/iotdb/pull/16903 \ No newline at end of file + Reference implementation: https://github.com/apache/iotdb/pull/16903 diff --git a/src/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index bc20138a4..284198ab6 100644 --- a/src/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -46,7 +46,7 @@ The Timer model (non-built-in model) not only demonstrates excellent few-shot ge ## 4. Timer-XL Model -Timer-XL further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: +Timer-XL[2] further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: * **Ultra-Long Context Support**: This model breaks through the limitations of traditional time series forecasting models, supporting the processing of inputs with thousands of Tokens (equivalent to tens of thousands of time points), effectively solving the context length bottleneck problem. * **Coverage of Multi-Variable Forecasting Scenarios**: Supports various forecasting scenarios, including the prediction of non-stationary time series, multi-variable prediction tasks, and predictions involving covariates, meeting diversified business needs. @@ -56,7 +56,7 @@ Timer-XL further extends and upgrades the network structure based on Timer, achi ## 5. Timer-Sundial Model -Timer-Sundial is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: +Timer-Sundial[3] is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: * **Strong Generalization Performance**: Possesses zero-shot forecasting capabilities and can support both point forecasting and probabilistic forecasting simultaneously. * **Flexible Prediction Distribution Analysis**: Not only can it predict means or quantiles, but it can also evaluate any statistical properties of the prediction distribution through the raw samples generated by the model. @@ -66,7 +66,7 @@ Timer-Sundial is a series of generative foundational models focused on time seri ## 6. Chronos-2 Model -Chronos-2 is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: +Chronos-2[4] is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: * **Probabilistic Forecasting Capability**: The model outputs multi-step prediction results in a generative manner, supporting quantile or distribution-level forecasting to characterize future uncertainty. * **Zero-Shot General Forecasting**: Leveraging the contextual learning ability acquired through pre-training, it can directly execute forecasting on unseen datasets without retraining or parameter updates. @@ -78,7 +78,34 @@ Chronos-2 is a universal time series foundational model developed by the Amazon ![](/img/timeseries-large-model-chronos2.png) -## 7. Performance Showcase +## 7. Moirai2 Model + +Moirai2[5] (Moirai 2.0) is a general-purpose time series foundation model developed by Salesforce AI Research (supported in V2.0.10 and later). AINode currently integrates the Moirai 2.0 R-small variant, which has approximately 11.4M parameters. Unlike Moirai 1.0, which uses a masked encoder architecture, Moirai 2.0 uses a causal decoder-only Patch Transformer. With a single patch size, multi-token prediction, and multi-quantile outputs, it provides efficient univariate forecasting with a compact model. Its core features include: + +- **Lightweight Model Architecture**: Uses a decoder-only Patch Transformer with RMSNorm, rotary positional embeddings, and SiLU-GLU feed-forward networks to balance forecasting capability and inference efficiency at a small parameter scale. +- **Multi-Token Prediction**: Predicts multiple patches at each decoding step, reducing the number of autoregressive decoding steps required for long forecast horizons. +- **Probabilistic Forecasting**: Outputs nine quantiles from 0.1 to 0.9. AINode uses the p50 median as the point forecast. +- **Patch Decoding**: Groups the time series into fixed-size patches before the attention module to improve temporal feature extraction and decoding efficiency. +- **Instance Normalization**: Standardizes each time series before model input and applies denormalization after output to mitigate distribution shifts across series. +- **Input Scope**: Focuses on univariate forecasting and does not support multivariate targets or covariates. + +![](/img/LargeModel-moirai2.png) + +> Note: The Moirai 2.0 R-small model weights are licensed under CC BY-NC 4.0 and are restricted to research use. + +## 8. Toto Model + +Toto[6] (Toto 2.0) is a next-generation time series foundation model developed by Datadog (supported in V2.0.10 and later), primarily for forecasting in observability scenarios. AINode currently integrates the 2.5B-parameter variant. It is based on a decoder-only Patch Transformer architecture that alternates causal temporal attention and variable attention to jointly model the temporal and variable dimensions. Its core features include: + +- **Univariate and Multivariate Forecasting**: Supports both individual target variables and joint forecasting of multiple related target variables, making it suitable for observability metrics such as CPU, memory, and network traffic. +- **Probabilistic Forecasting**: Outputs fixed quantiles from 0.1 to 0.9 to represent forecasting uncertainty. AINode uses the p50 median as the point forecast. +- **Efficient Block Decoding**: Uses cached block decoding to generate forecasts in blocks, reducing repeated computation for longer forecast horizons. +- **Large-Scale Model Capability**: Improves forecasting performance by scaling model parameters and pre-training data, providing strong zero-shot generalization. +- **Covariate Limitation**: The currently integrated version does not support historical covariates or known future covariates. + +![](/img/LargeModel-toto.png) + +## 9. Performance Showcase Time Series Large Models can adapt to real time series data from various different domains and scenarios, demonstrating excellent processing capabilities across various tasks. The following shows the actual performance on different datasets: @@ -100,7 +127,7 @@ Using Time Series Large Models to accurately identify outliers that deviate sign ![](/img/LargeModel05.png) -## 8. Deployment and Usage +## 10. Deployment and Usage 1. Open the IoTDB CLI console and check that the ConfigNode, DataNode, and AINode nodes are all Running. @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer: Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() -**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back](#ref4) -**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back]() +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ Back](#ref5) -**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back]() \ No newline at end of file +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ Back](#ref6) diff --git a/src/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md b/src/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md index 259417fbc..0f614b413 100644 --- a/src/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md +++ b/src/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md @@ -416,7 +416,12 @@ IoTDB> show models | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -424,22 +429,24 @@ Built-in traditional time series model introduction: | Model Name | Core Concept | Applicable Scenario | Main Features | |----------------------------------| ----------------------------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------- | -| **ARIMA** (AutoRegressive Integrated Moving Average) | Combines autoregression (AR), differencing (I), and moving average (MA), used for predicting stationary time series or data that can be made stationary through differencing. | Univariate time series prediction, such as stock prices, sales, economic indicators. | 1. Suitable for linear trends and weak seasonality data. 2. Requires selecting parameters (p,d,q). 3. Sensitive to missing values. | -| **Holt-Winters** (Three-parameter exponential smoothing) | Based on exponential smoothing, introduces level, trend, and seasonal components, suitable for data with trend and seasonality. | Time series with clear seasonality and trend, such as monthly sales, power demand. | 1. Can handle additive or multiplicative seasonality. 2. Gives higher weight to recent data. 3. Simple to implement. | -| **Exponential Smoothing** (Exponential smoothing) | Uses weighted average of historical data, with weights decreasing exponentially over time, emphasizing the importance of recent observations. | Data without significant seasonality but with trend, such as short-term demand prediction. | 1. Few parameters, simple calculation. 2. Suitable for stationary or slowly changing sequences. 3. Can be extended to double or triple exponential smoothing. | -| **Naive Forecaster** (Naive predictor) | Uses the observation of the most recent period as the prediction for the next period, the simplest baseline model. | As a benchmark for other models or simple prediction when data has no obvious pattern. | 1. No training required. 2. Sensitive to sudden changes. 3. Seasonal naive variant can use the same period of the previous season to predict. | -| **STL Forecaster** (Seasonal-Trend Decomposition Forecast) | Based on STL decomposition of time series, predicts trend, seasonal, and residual components separately, then combines them. | Data with complex seasonality, trend, and non-linear patterns, such as climate data, traffic flow. | 1. Can handle non-fixed seasonality. 2. Robust to outliers. 3. After decomposition, other models can be combined to predict each component. | -| **Gaussian HMM** (Gaussian Hidden Markov Model) | Assumes observed data is generated by hidden states, with each state's observation probability following a Gaussian distribution. | State sequence prediction or classification, such as speech recognition, financial state identification. | 1. Suitable for time series state modeling. 2. Assumes observations are independent given the state. 3. Requires specifying the number of hidden states. | -| **GMM HMM** (Gaussian Mixture Hidden Markov Model) | An extension of Gaussian HMM, where each state's observation probability is described by a Gaussian Mixture Model, capturing more complex observation distributions. | Scenarios requiring multi-modal observation distributions, such as complex action recognition, biosignal analysis. | 1. More flexible than single Gaussian. 2. More parameters, higher computational complexity. 3. Requires training the number of GMM components. | -| **STRAY** (Anomaly Detection based on Singular Value Decomposition) | Detects anomalies in high-dimensional data through Singular Value Decomposition (SVD), commonly used for time series anomaly detection. | High-dimensional time series anomaly detection, such as sensor networks, IT system monitoring. | 1. No distribution assumption required. 2. Can handle high-dimensional data. 3. Sensitive to global anomalies, may miss local anomalies. | +| **ARIMA** (AutoRegressive Integrated Moving Average) | Combines autoregression (AR), differencing (I), and moving average (MA), used for predicting stationary time series or data that can be made stationary through differencing. | Univariate time series prediction, such as stock prices, sales, economic indicators. | 1. Suitable for linear trends and weak seasonality data.
2. Requires selecting parameters (p,d,q).
3. Sensitive to missing values. | +| **Holt-Winters** (Three-parameter exponential smoothing) | Based on exponential smoothing, introduces level, trend, and seasonal components, suitable for data with trend and seasonality. | Time series with clear seasonality and trend, such as monthly sales, power demand. | 1. Can handle additive or multiplicative seasonality.
2. Gives higher weight to recent data.
3. Simple to implement. | +| **Exponential Smoothing** (Exponential smoothing) | Uses weighted average of historical data, with weights decreasing exponentially over time, emphasizing the importance of recent observations. | Data without significant seasonality but with trend, such as short-term demand prediction. | 1. Few parameters, simple calculation.
2. Suitable for stationary or slowly changing sequences.
3. Can be extended to double or triple exponential smoothing. | +| **Naive Forecaster** (Naive predictor) | Uses the observation of the most recent period as the prediction for the next period, the simplest baseline model. | As a benchmark for other models or simple prediction when data has no obvious pattern. | 1. No training required.
2. Sensitive to sudden changes.
3. Seasonal naive variant can use the same period of the previous season to predict. | +| **STL Forecaster** (Seasonal-Trend Decomposition Forecast) | Based on STL decomposition of time series, predicts trend, seasonal, and residual components separately, then combines them. | Data with complex seasonality, trend, and non-linear patterns, such as climate data, traffic flow. | 1. Can handle non-fixed seasonality.
2. Robust to outliers.
3. After decomposition, other models can be combined to predict each component. | +| **Gaussian HMM** (Gaussian Hidden Markov Model) | Assumes observed data is generated by hidden states, with each state's observation probability following a Gaussian distribution. | State sequence prediction or classification, such as speech recognition, financial state identification. | 1. Suitable for time series state modeling.
2. Assumes observations are independent given the state.
3. Requires specifying the number of hidden states. | +| **GMM HMM** (Gaussian Mixture Hidden Markov Model) | An extension of Gaussian HMM, where each state's observation probability is described by a Gaussian Mixture Model, capturing more complex observation distributions. | Scenarios requiring multi-modal observation distributions, such as complex action recognition, biosignal analysis. | 1. More flexible than single Gaussian.
2. More parameters, higher computational complexity.
3. Requires training the number of GMM components. | +| **STRAY** (Anomaly Detection based on Singular Value Decomposition) | Detects anomalies in high-dimensional data through Singular Value Decomposition (SVD), commonly used for time series anomaly detection. | High-dimensional time series anomaly detection, such as sensor networks, IT system monitoring. | 1. No distribution assumption required.
2. Can handle high-dimensional data.
3. Sensitive to global anomalies, may miss local anomalies. | Built-in time series large model introduction: | Model Name | Core Concept | Applicable Scenario | Main Features | |---------------| ---------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| **Timer-XL** | Time series large model supporting ultra-long context, enhancing generalization capability through large-scale industrial data pre-training. | Complex industrial prediction requiring extremely long historical data, such as energy, aerospace, and transportation. | 1. Ultra-long context support, can handle tens of thousands of time points as input. 2. Multi-scenario coverage, supports non-stationary, multi-variable, and covariate prediction. 3. Pre-trained on trillions of high-quality industrial time series data. | -| **Timer-Sundial** | A generative foundational model based on "Transformer + TimeFlow" architecture, focusing on probabilistic prediction. | Zero-shot prediction scenarios requiring quantification of uncertainty, such as finance, supply chain, and new energy power generation. | 1. Strong zero-shot generalization capability, supports point prediction and probabilistic prediction. 2. Can flexibly analyze any statistical properties of the prediction distribution. 3. Innovative generative architecture, achieving efficient non-deterministic sample generation. | -| **Chronos-2** | A general time series foundational model based on discrete tokenization paradigm, converting prediction into language modeling tasks. | Rapid zero-shot univariate prediction, and scenarios that can leverage covariates (e.g., promotions, weather) to improve results. | 1. Strong zero-shot probabilistic prediction capability. 2. Supports unified covariate modeling, but has strict input requirements: a. The set of names of future covariates must be a subset of the set of names of historical covariates; b. The length of each historical covariate must equal the length of the target variable; c. The length of each future covariate must equal the prediction length; 3. Uses an efficient encoder-style structure, balancing performance and inference speed. | +| **Timer-XL** | Time series large model supporting ultra-long context, enhancing generalization capability through large-scale industrial data pre-training. | Complex industrial prediction requiring extremely long historical data, such as energy, aerospace, and transportation. | 1. Ultra-long context support, can handle tens of thousands of time points as input.
2. Multi-scenario coverage, supports non-stationary, multi-variable, and covariate prediction.
3. Pre-trained on trillions of high-quality industrial time series data. | +| **Timer-Sundial** | A generative foundational model based on "Transformer + TimeFlow" architecture, focusing on probabilistic prediction. | Zero-shot prediction scenarios requiring quantification of uncertainty, such as finance, supply chain, and new energy power generation. | 1. Strong zero-shot generalization capability, supports point prediction and probabilistic prediction.
2. Can flexibly analyze any statistical properties of the prediction distribution.
3. Innovative generative architecture, achieving efficient non-deterministic sample generation. | +| **Chronos-2** | A general time series foundational model based on discrete tokenization paradigm, converting prediction into language modeling tasks. | Rapid zero-shot univariate prediction, and scenarios that can leverage covariates (e.g., promotions, weather) to improve results. | 1. Strong zero-shot probabilistic prediction capability.
2. Supports unified covariate modeling, but has strict input requirements:
  a. The set of names of future covariates must be a subset of the set of names of historical covariates;
  b. The length of each historical covariate must equal the length of the target variable;
  c. The length of each future covariate must equal the prediction length;
3. Uses an efficient encoder-style structure, balancing performance and inference speed. | +| **Moirai 2.0** | Uses a lightweight decoder-only Patch Transformer with a single patch size, multi-token prediction, and multi-quantile outputs for efficient univariate forecasting. | Zero-shot univariate forecasting where model size and inference efficiency are important, such as industrial monitoring, energy load, and equipment metric forecasting. | 1. Approximately 11.4M parameters.
2. Predicts multiple patches per decoding step to reduce autoregressive overhead for long horizons.
3. Outputs nine quantiles (0.1–0.9) and uses the p50 median as the point forecast.
4. Uses instance normalization to mitigate distribution shift across series.
5. Does not support multivariate targets or covariates. | +| **Toto 2.0** | Uses a decoder-only Patch Transformer that alternates causal temporal attention and variable attention to jointly model the temporal and variable dimensions. | Zero-shot forecasting for multivariate time series such as observability metrics, including joint forecasting of CPU, memory, and network traffic. | 1. Supports univariate and multivariate target forecasting.
2. Outputs fixed quantiles from 0.1 to 0.9 and uses the p50 median as the point forecast.
3. Supports cached block decoding for efficient scaling to longer forecast horizons.
4. Covariates are not currently supported. | ### 4.4 Delete Models @@ -675,4 +682,4 @@ Support adding new built-in models to AINode. The specific steps are as follows * **Full Example** -Full example can be referenced at https://github.com/apache/iotdb/pull/16903 \ No newline at end of file +Full example can be referenced at https://github.com/apache/iotdb/pull/16903 diff --git a/src/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index bc20138a4..284198ab6 100644 --- a/src/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -46,7 +46,7 @@ The Timer model (non-built-in model) not only demonstrates excellent few-shot ge ## 4. Timer-XL Model -Timer-XL further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: +Timer-XL[2] further extends and upgrades the network structure based on Timer, achieving comprehensive breakthroughs in multiple dimensions: * **Ultra-Long Context Support**: This model breaks through the limitations of traditional time series forecasting models, supporting the processing of inputs with thousands of Tokens (equivalent to tens of thousands of time points), effectively solving the context length bottleneck problem. * **Coverage of Multi-Variable Forecasting Scenarios**: Supports various forecasting scenarios, including the prediction of non-stationary time series, multi-variable prediction tasks, and predictions involving covariates, meeting diversified business needs. @@ -56,7 +56,7 @@ Timer-XL further extends and upgrades the network structure based on Timer, achi ## 5. Timer-Sundial Model -Timer-Sundial is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: +Timer-Sundial[3] is a series of generative foundational models focused on time series forecasting. The base version has 128 million parameters and has been pre-trained on 1 trillion time points, with the following core characteristics: * **Strong Generalization Performance**: Possesses zero-shot forecasting capabilities and can support both point forecasting and probabilistic forecasting simultaneously. * **Flexible Prediction Distribution Analysis**: Not only can it predict means or quantiles, but it can also evaluate any statistical properties of the prediction distribution through the raw samples generated by the model. @@ -66,7 +66,7 @@ Timer-Sundial is a series of generative foundational models focused on time seri ## 6. Chronos-2 Model -Chronos-2 is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: +Chronos-2[4] is a universal time series foundational model developed by the Amazon Web Services (AWS) research team, evolved from the Chronos discrete token modeling paradigm. This model is suitable for both zero-shot univariate forecasting and covariate forecasting. Its main characteristics include: * **Probabilistic Forecasting Capability**: The model outputs multi-step prediction results in a generative manner, supporting quantile or distribution-level forecasting to characterize future uncertainty. * **Zero-Shot General Forecasting**: Leveraging the contextual learning ability acquired through pre-training, it can directly execute forecasting on unseen datasets without retraining or parameter updates. @@ -78,7 +78,34 @@ Chronos-2 is a universal time series foundational model developed by the Amazon ![](/img/timeseries-large-model-chronos2.png) -## 7. Performance Showcase +## 7. Moirai2 Model + +Moirai2[5] (Moirai 2.0) is a general-purpose time series foundation model developed by Salesforce AI Research (supported in V2.0.10 and later). AINode currently integrates the Moirai 2.0 R-small variant, which has approximately 11.4M parameters. Unlike Moirai 1.0, which uses a masked encoder architecture, Moirai 2.0 uses a causal decoder-only Patch Transformer. With a single patch size, multi-token prediction, and multi-quantile outputs, it provides efficient univariate forecasting with a compact model. Its core features include: + +- **Lightweight Model Architecture**: Uses a decoder-only Patch Transformer with RMSNorm, rotary positional embeddings, and SiLU-GLU feed-forward networks to balance forecasting capability and inference efficiency at a small parameter scale. +- **Multi-Token Prediction**: Predicts multiple patches at each decoding step, reducing the number of autoregressive decoding steps required for long forecast horizons. +- **Probabilistic Forecasting**: Outputs nine quantiles from 0.1 to 0.9. AINode uses the p50 median as the point forecast. +- **Patch Decoding**: Groups the time series into fixed-size patches before the attention module to improve temporal feature extraction and decoding efficiency. +- **Instance Normalization**: Standardizes each time series before model input and applies denormalization after output to mitigate distribution shifts across series. +- **Input Scope**: Focuses on univariate forecasting and does not support multivariate targets or covariates. + +![](/img/LargeModel-moirai2.png) + +> Note: The Moirai 2.0 R-small model weights are licensed under CC BY-NC 4.0 and are restricted to research use. + +## 8. Toto Model + +Toto[6] (Toto 2.0) is a next-generation time series foundation model developed by Datadog (supported in V2.0.10 and later), primarily for forecasting in observability scenarios. AINode currently integrates the 2.5B-parameter variant. It is based on a decoder-only Patch Transformer architecture that alternates causal temporal attention and variable attention to jointly model the temporal and variable dimensions. Its core features include: + +- **Univariate and Multivariate Forecasting**: Supports both individual target variables and joint forecasting of multiple related target variables, making it suitable for observability metrics such as CPU, memory, and network traffic. +- **Probabilistic Forecasting**: Outputs fixed quantiles from 0.1 to 0.9 to represent forecasting uncertainty. AINode uses the p50 median as the point forecast. +- **Efficient Block Decoding**: Uses cached block decoding to generate forecasts in blocks, reducing repeated computation for longer forecast horizons. +- **Large-Scale Model Capability**: Improves forecasting performance by scaling model parameters and pre-training data, providing strong zero-shot generalization. +- **Covariate Limitation**: The currently integrated version does not support historical covariates or known future covariates. + +![](/img/LargeModel-toto.png) + +## 9. Performance Showcase Time Series Large Models can adapt to real time series data from various different domains and scenarios, demonstrating excellent processing capabilities across various tasks. The following shows the actual performance on different datasets: @@ -100,7 +127,7 @@ Using Time Series Large Models to accurately identify outliers that deviate sign ![](/img/LargeModel05.png) -## 8. Deployment and Usage +## 10. Deployment and Usage 1. Open the IoTDB CLI console and check that the ConfigNode, DataNode, and AINode nodes are all Running. @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer: Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() -**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back]() +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ Back](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back](#ref4) -**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ Back]() +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ Back](#ref5) -**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821**. [↩ Back]() \ No newline at end of file +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ Back](#ref6) diff --git a/src/zh/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md b/src/zh/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md index 6356a0489..915ef8691 100644 --- a/src/zh/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md +++ b/src/zh/UserGuide/Master/Table/AI-capability/AINode_Upgrade_apache.md @@ -443,7 +443,12 @@ IoTDB> show models | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -451,22 +456,24 @@ IoTDB> show models | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |----------------------------------| ----------------------------------------------------------------------------------------- | ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | -| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。2. 需要选择参数 (p,d,q)。3. 对缺失值敏感。 | -| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。2. 对近期数据赋予更高权重。3. 简单易实现。 | -| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。2. 适合平稳或缓慢变化序列。3. 可扩展为双指数或三指数平滑。 | -| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。2. 对突发变化敏感。3. 季节性朴素变体可用前一季节同期值预测。 | -| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。2. 对异常值稳健。3. 分解后可结合其他模型预测各分量。 | -| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。2. 假设观测值在给定状态下独立。3. 需指定隐藏状态数量。 | -| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。2. 参数更多，计算复杂度高。3. 需训练GMM成分数。 | -| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。2. 可处理高维数据。3. 对全局异常敏感，局部异常可能漏检。 | +| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。
2. 需要选择参数 (p,d,q)。
3. 对缺失值敏感。 | +| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。
2. 对近期数据赋予更高权重。
3. 简单易实现。 | +| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。
2. 适合平稳或缓慢变化序列。
3. 可扩展为双指数或三指数平滑。 | +| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。
2. 对突发变化敏感。
3. 季节性朴素变体可用前一季节同期值预测。 | +| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。
2. 对异常值稳健。
3. 分解后可结合其他模型预测各分量。 | +| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。
2. 假设观测值在给定状态下独立。
3. 需指定隐藏状态数量。 | +| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。
2. 参数更多，计算复杂度高。
3. 需训练GMM成分数。 | +| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。
2. 可处理高维数据。
3. 对全局异常敏感，局部异常可能漏检。 | 内置时序大模型介绍如下： | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |---------------| ---------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。2. 多场景覆盖，支持非平稳、多变量及协变量预测。3. 基于万亿级高质量工业时序数据预训练。 | -| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测 2. 可灵活分析预测分布的任意统计特性。3. 创新生成架构，实现高效的非确定性样本生成。 | -| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。2. 支持协变量统一建模，但对输入有严格要求：a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；b. 每个历史协变量的长度必须等于目标变量的长度； c. 每个未来协变量的长度必须等于预测长度；3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。
2. 多场景覆盖，支持非平稳、多变量及协变量预测。
3. 基于万亿级高质量工业时序数据预训练。 | +| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测
2. 可灵活分析预测分布的任意统计特性。
3. 创新生成架构，实现高效的非确定性样本生成。 | +| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。
2. 支持协变量统一建模，但对输入有严格要求：
  a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；
  b. 每个历史协变量的长度必须等于目标变量的长度；
  c. 每个未来协变量的长度必须等于预测长度；
3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Moirai 2.0** | 采用轻量级 Decoder-only Patch Transformer，通过单一 Patch 尺寸、多 Token 预测和多分位数输出实现高效的单变量预测。 | 适用于对模型体量和推理效率要求较高的零样本单变量预测，如工业监测、能源负荷及设备指标预测。 | 1. 模型参数量约 11.4M。
2. 每个解码步可预测多个 Patch，降低长预测范围的自回归开销。
3. 输出 9 个分位数（0.1～0.9），使用 p50 中位数作为点预测。
4. 使用实例归一化缓解序列分布漂移。
5. 不支持多变量和协变量。 | +| **Toto 2.0** | 采用 Decoder-only Patch Transformer，交替使用因果时间注意力与变量注意力，对时间维和变量维进行联合建模。 | 面向可观测性指标等多变量时间序列的零样本预测，如 CPU、内存、网络流量等指标的联合预测。 | 1. 支持单变量和多变量目标预测。
2. 输出 0.1～0.9 的固定分位数，使用 p50 中位数作为点预测。
3. 支持缓存式块解码，可高效扩展较长预测范围。
4. 当前不支持协变量。 | ### 4.4 删除模型 diff --git a/src/zh/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/zh/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index e1af1e1d0..5fe5c3771 100644 --- a/src/zh/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/zh/UserGuide/Master/Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -78,7 +78,34 @@ Chronos-2 [4] ![](/img/timeseries-large-model-chronos2.png) -## 7. 效果展示 +## 7. Moirai2 模型 + +Moirai2[5]（Moirai 2.0）是由 Salesforce AI Research 推出的通用时间序列基础模型（V2.0.10 及以后版本支持）。当前集成的是 Moirai 2.0 R-small 版本，模型参数量约为 11.4M。与采用掩码编码器架构的 Moirai 1.0 不同，Moirai 2.0 使用因果 Decoder-only Patch Transformer，并通过单一 Patch 尺寸、多 Token 预测和多分位数输出，在较小模型规模下实现高效的单变量预测。其核心特性包括： + +- **轻量化模型架构**：采用 Decoder-only Patch Transformer，并结合 RMSNorm、旋转位置编码和 SiLU-GLU 前馈网络，在较小参数规模下兼顾预测能力和推理效率。 +- **多 Token 预测**：每个解码步骤可同时预测多个 Patch，减少长预测范围下所需的自回归解码次数。 +- **概率性预测能力**：模型输出 0.1～0.9 共 9 个分位数，AINode 使用 p50 中位数作为点预测结果。 +- **Patch 解码**：在进入注意力模块前，将时间序列按照固定大小的 Patch 进行分组，以提高时序特征提取和解码效率。 +- **实例归一化**：在模型输入前对每条时间序列进行标准化，并在输出后执行反归一化，以缓解不同序列间的分布漂移。 +- **输入范围**：模型聚焦于单变量预测，不支持多变量目标及协变量输入。 + +![](/img/LargeModel-moirai2.png) + +> 注意：Moirai 2.0 R-small 模型权重采用 CC BY-NC 4.0 许可证，仅限研究用途。 + +## 8. Toto 模型 + +Toto[6]（Toto 2.0）是由 Datadog 推出的新一代时间序列基础模型（V2.0.10 及以后版本支持），主要面向可观测性场景中的时间序列预测。当前集成的模型采用 2.5B 参数版本，基于 Decoder-only Patch Transformer 架构，交替使用因果时间注意力与变量注意力，对时间维度和变量维度进行联合建模。其核心特性包括： + +- **单变量与多变量预测**：支持单个目标变量以及多个相关目标变量的联合预测，适用于 CPU、内存、网络流量等可观测性指标的分析场景。 +- **概率性预测能力**：模型输出 0.1～0.9 的固定分位数，可刻画预测结果的不确定性；AINode 使用 p50 中位数作为点预测结果。 +- **高效块解码**：采用缓存式块解码机制，可分块生成预测结果，降低较长预测范围下的重复计算开销。 +- **大规模模型能力**：通过扩大模型参数量和预训练数据规模提升预测性能，具备良好的零样本泛化能力。 +- **协变量限制**：当前集成版本不支持历史协变量或已知未来协变量。 + +![](/img/LargeModel-toto.png) + +## 9. 效果展示 时序大模型能够适应多种不同领域和场景的真实时序数据，在各种任务上拥有优异的处理效果，以下是在不同数据上的真实表现： @@ -100,7 +127,7 @@ Chronos-2 [4] ![](/img/LargeModel05.png) -## 8. 部署使用 +## 10. 部署使用 1. 打开 IoTDB cli 控制台，检查 ConfigNode、DataNode、AINode 节点确保均为 Running。 @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer- Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref1) -**[2]** TIMER-XL- LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING ,Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.** [↩ 返回](#ref4) -**[3]** Sundial- A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ 返回](#ref5) -**[4] **Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.**[↩ 返回](#ref4) +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ 返回](#ref6) diff --git a/src/zh/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md b/src/zh/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md index a86a0ea7c..255eaff80 100644 --- a/src/zh/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md +++ b/src/zh/UserGuide/Master/Tree/AI-capability/AINode_Upgrade_apache.md @@ -419,7 +419,12 @@ IoTDB> show models | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -427,22 +432,24 @@ IoTDB> show models | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |----------------------------------| ----------------------------------------------------------------------------------------- | ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | -| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。2. 需要选择参数 (p,d,q)。3. 对缺失值敏感。 | -| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。2. 对近期数据赋予更高权重。3. 简单易实现。 | -| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。2. 适合平稳或缓慢变化序列。3. 可扩展为双指数或三指数平滑。 | -| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。2. 对突发变化敏感。3. 季节性朴素变体可用前一季节同期值预测。 | -| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。2. 对异常值稳健。3. 分解后可结合其他模型预测各分量。 | -| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。2. 假设观测值在给定状态下独立。3. 需指定隐藏状态数量。 | -| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。2. 参数更多，计算复杂度高。3. 需训练GMM成分数。 | -| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。2. 可处理高维数据。3. 对全局异常敏感，局部异常可能漏检。 | +| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。
2. 需要选择参数 (p,d,q)。
3. 对缺失值敏感。 | +| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。
2. 对近期数据赋予更高权重。
3. 简单易实现。 | +| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。
2. 适合平稳或缓慢变化序列。
3. 可扩展为双指数或三指数平滑。 | +| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。
2. 对突发变化敏感。
3. 季节性朴素变体可用前一季节同期值预测。 | +| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。
2. 对异常值稳健。
3. 分解后可结合其他模型预测各分量。 | +| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。
2. 假设观测值在给定状态下独立。
3. 需指定隐藏状态数量。 | +| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。
2. 参数更多，计算复杂度高。
3. 需训练GMM成分数。 | +| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。
2. 可处理高维数据。
3. 对全局异常敏感，局部异常可能漏检。 | 内置时序大模型介绍如下： | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |---------------| ---------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。2. 多场景覆盖，支持非平稳、多变量及协变量预测。3. 基于万亿级高质量工业时序数据预训练。 | -| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测 2. 可灵活分析预测分布的任意统计特性。3. 创新生成架构，实现高效的非确定性样本生成。 | -| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。2. 支持协变量统一建模，但对输入有严格要求：a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；b. 每个历史协变量的长度必须等于目标变量的长度； c. 每个未来协变量的长度必须等于预测长度；3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。
2. 多场景覆盖，支持非平稳、多变量及协变量预测。
3. 基于万亿级高质量工业时序数据预训练。 | +| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测
2. 可灵活分析预测分布的任意统计特性。
3. 创新生成架构，实现高效的非确定性样本生成。 | +| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。
2. 支持协变量统一建模，但对输入有严格要求：
  a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；
  b. 每个历史协变量的长度必须等于目标变量的长度；
  c. 每个未来协变量的长度必须等于预测长度；
3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Moirai 2.0** | 采用轻量级 Decoder-only Patch Transformer，通过单一 Patch 尺寸、多 Token 预测和多分位数输出实现高效的单变量预测。 | 适用于对模型体量和推理效率要求较高的零样本单变量预测，如工业监测、能源负荷及设备指标预测。 | 1. 模型参数量约 11.4M。
2. 每个解码步可预测多个 Patch，降低长预测范围的自回归开销。
3. 输出 9 个分位数（0.1～0.9），使用 p50 中位数作为点预测。
4. 使用实例归一化缓解序列分布漂移。
5. 不支持多变量和协变量。 | +| **Toto 2.0** | 采用 Decoder-only Patch Transformer，交替使用因果时间注意力与变量注意力，对时间维和变量维进行联合建模。 | 面向可观测性指标等多变量时间序列的零样本预测，如 CPU、内存、网络流量等指标的联合预测。 | 1. 支持单变量和多变量目标预测。
2. 输出 0.1～0.9 的固定分位数，使用 p50 中位数作为点预测。
3. 支持缓存式块解码，可高效扩展较长预测范围。
4. 当前不支持协变量。 | ### 4.4 删除模型 diff --git a/src/zh/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/zh/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index e1af1e1d0..5fe5c3771 100644 --- a/src/zh/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/zh/UserGuide/Master/Tree/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -78,7 +78,34 @@ Chronos-2 [4] ![](/img/timeseries-large-model-chronos2.png) -## 7. 效果展示 +## 7. Moirai2 模型 + +Moirai2[5]（Moirai 2.0）是由 Salesforce AI Research 推出的通用时间序列基础模型（V2.0.10 及以后版本支持）。当前集成的是 Moirai 2.0 R-small 版本，模型参数量约为 11.4M。与采用掩码编码器架构的 Moirai 1.0 不同，Moirai 2.0 使用因果 Decoder-only Patch Transformer，并通过单一 Patch 尺寸、多 Token 预测和多分位数输出，在较小模型规模下实现高效的单变量预测。其核心特性包括： + +- **轻量化模型架构**：采用 Decoder-only Patch Transformer，并结合 RMSNorm、旋转位置编码和 SiLU-GLU 前馈网络，在较小参数规模下兼顾预测能力和推理效率。 +- **多 Token 预测**：每个解码步骤可同时预测多个 Patch，减少长预测范围下所需的自回归解码次数。 +- **概率性预测能力**：模型输出 0.1～0.9 共 9 个分位数，AINode 使用 p50 中位数作为点预测结果。 +- **Patch 解码**：在进入注意力模块前，将时间序列按照固定大小的 Patch 进行分组，以提高时序特征提取和解码效率。 +- **实例归一化**：在模型输入前对每条时间序列进行标准化，并在输出后执行反归一化，以缓解不同序列间的分布漂移。 +- **输入范围**：模型聚焦于单变量预测，不支持多变量目标及协变量输入。 + +![](/img/LargeModel-moirai2.png) + +> 注意：Moirai 2.0 R-small 模型权重采用 CC BY-NC 4.0 许可证，仅限研究用途。 + +## 8. Toto 模型 + +Toto[6]（Toto 2.0）是由 Datadog 推出的新一代时间序列基础模型（V2.0.10 及以后版本支持），主要面向可观测性场景中的时间序列预测。当前集成的模型采用 2.5B 参数版本，基于 Decoder-only Patch Transformer 架构，交替使用因果时间注意力与变量注意力，对时间维度和变量维度进行联合建模。其核心特性包括： + +- **单变量与多变量预测**：支持单个目标变量以及多个相关目标变量的联合预测，适用于 CPU、内存、网络流量等可观测性指标的分析场景。 +- **概率性预测能力**：模型输出 0.1～0.9 的固定分位数，可刻画预测结果的不确定性；AINode 使用 p50 中位数作为点预测结果。 +- **高效块解码**：采用缓存式块解码机制，可分块生成预测结果，降低较长预测范围下的重复计算开销。 +- **大规模模型能力**：通过扩大模型参数量和预训练数据规模提升预测性能，具备良好的零样本泛化能力。 +- **协变量限制**：当前集成版本不支持历史协变量或已知未来协变量。 + +![](/img/LargeModel-toto.png) + +## 9. 效果展示 时序大模型能够适应多种不同领域和场景的真实时序数据，在各种任务上拥有优异的处理效果，以下是在不同数据上的真实表现： @@ -100,7 +127,7 @@ Chronos-2 [4] ![](/img/LargeModel05.png) -## 8. 部署使用 +## 10. 部署使用 1. 打开 IoTDB cli 控制台，检查 ConfigNode、DataNode、AINode 节点确保均为 Running。 @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer- Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref1) -**[2]** TIMER-XL- LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING ,Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.** [↩ 返回](#ref4) -**[3]** Sundial- A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ 返回](#ref5) -**[4] **Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.**[↩ 返回](#ref4) +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ 返回](#ref6) diff --git a/src/zh/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md b/src/zh/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md index 6356a0489..915ef8691 100644 --- a/src/zh/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md +++ b/src/zh/UserGuide/latest-Table/AI-capability/AINode_Upgrade_apache.md @@ -443,7 +443,12 @@ IoTDB> show models | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -451,22 +456,24 @@ IoTDB> show models | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |----------------------------------| ----------------------------------------------------------------------------------------- | ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | -| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。2. 需要选择参数 (p,d,q)。3. 对缺失值敏感。 | -| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。2. 对近期数据赋予更高权重。3. 简单易实现。 | -| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。2. 适合平稳或缓慢变化序列。3. 可扩展为双指数或三指数平滑。 | -| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。2. 对突发变化敏感。3. 季节性朴素变体可用前一季节同期值预测。 | -| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。2. 对异常值稳健。3. 分解后可结合其他模型预测各分量。 | -| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。2. 假设观测值在给定状态下独立。3. 需指定隐藏状态数量。 | -| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。2. 参数更多，计算复杂度高。3. 需训练GMM成分数。 | -| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。2. 可处理高维数据。3. 对全局异常敏感，局部异常可能漏检。 | +| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。
2. 需要选择参数 (p,d,q)。
3. 对缺失值敏感。 | +| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。
2. 对近期数据赋予更高权重。
3. 简单易实现。 | +| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。
2. 适合平稳或缓慢变化序列。
3. 可扩展为双指数或三指数平滑。 | +| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。
2. 对突发变化敏感。
3. 季节性朴素变体可用前一季节同期值预测。 | +| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。
2. 对异常值稳健。
3. 分解后可结合其他模型预测各分量。 | +| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。
2. 假设观测值在给定状态下独立。
3. 需指定隐藏状态数量。 | +| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。
2. 参数更多，计算复杂度高。
3. 需训练GMM成分数。 | +| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。
2. 可处理高维数据。
3. 对全局异常敏感，局部异常可能漏检。 | 内置时序大模型介绍如下： | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |---------------| ---------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。2. 多场景覆盖，支持非平稳、多变量及协变量预测。3. 基于万亿级高质量工业时序数据预训练。 | -| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测 2. 可灵活分析预测分布的任意统计特性。3. 创新生成架构，实现高效的非确定性样本生成。 | -| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。2. 支持协变量统一建模，但对输入有严格要求：a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；b. 每个历史协变量的长度必须等于目标变量的长度； c. 每个未来协变量的长度必须等于预测长度；3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。
2. 多场景覆盖，支持非平稳、多变量及协变量预测。
3. 基于万亿级高质量工业时序数据预训练。 | +| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测
2. 可灵活分析预测分布的任意统计特性。
3. 创新生成架构，实现高效的非确定性样本生成。 | +| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。
2. 支持协变量统一建模，但对输入有严格要求：
  a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；
  b. 每个历史协变量的长度必须等于目标变量的长度；
  c. 每个未来协变量的长度必须等于预测长度；
3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Moirai 2.0** | 采用轻量级 Decoder-only Patch Transformer，通过单一 Patch 尺寸、多 Token 预测和多分位数输出实现高效的单变量预测。 | 适用于对模型体量和推理效率要求较高的零样本单变量预测，如工业监测、能源负荷及设备指标预测。 | 1. 模型参数量约 11.4M。
2. 每个解码步可预测多个 Patch，降低长预测范围的自回归开销。
3. 输出 9 个分位数（0.1～0.9），使用 p50 中位数作为点预测。
4. 使用实例归一化缓解序列分布漂移。
5. 不支持多变量和协变量。 | +| **Toto 2.0** | 采用 Decoder-only Patch Transformer，交替使用因果时间注意力与变量注意力，对时间维和变量维进行联合建模。 | 面向可观测性指标等多变量时间序列的零样本预测，如 CPU、内存、网络流量等指标的联合预测。 | 1. 支持单变量和多变量目标预测。
2. 输出 0.1～0.9 的固定分位数，使用 p50 中位数作为点预测。
3. 支持缓存式块解码，可高效扩展较长预测范围。
4. 当前不支持协变量。 | ### 4.4 删除模型 diff --git a/src/zh/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/zh/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index 99712b134..0a75fce5b 100644 --- a/src/zh/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/zh/UserGuide/latest-Table/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -78,7 +78,34 @@ Chronos-2 [4] ![](/img/timeseries-large-model-chronos2.png) -## 7. 效果展示 +## 7. Moirai2 模型 + +Moirai2[5]（Moirai 2.0）是由 Salesforce AI Research 推出的通用时间序列基础模型（V2.0.10 及以后版本支持）。当前集成的是 Moirai 2.0 R-small 版本，模型参数量约为 11.4M。与采用掩码编码器架构的 Moirai 1.0 不同，Moirai 2.0 使用因果 Decoder-only Patch Transformer，并通过单一 Patch 尺寸、多 Token 预测和多分位数输出，在较小模型规模下实现高效的单变量预测。其核心特性包括： + +- **轻量化模型架构**：采用 Decoder-only Patch Transformer，并结合 RMSNorm、旋转位置编码和 SiLU-GLU 前馈网络，在较小参数规模下兼顾预测能力和推理效率。 +- **多 Token 预测**：每个解码步骤可同时预测多个 Patch，减少长预测范围下所需的自回归解码次数。 +- **概率性预测能力**：模型输出 0.1～0.9 共 9 个分位数，AINode 使用 p50 中位数作为点预测结果。 +- **Patch 解码**：在进入注意力模块前，将时间序列按照固定大小的 Patch 进行分组，以提高时序特征提取和解码效率。 +- **实例归一化**：在模型输入前对每条时间序列进行标准化，并在输出后执行反归一化，以缓解不同序列间的分布漂移。 +- **输入范围**：模型聚焦于单变量预测，不支持多变量目标及协变量输入。 + +![](/img/LargeModel-moirai2.png) + +> 注意：Moirai 2.0 R-small 模型权重采用 CC BY-NC 4.0 许可证，仅限研究用途。 + +## 8. Toto 模型 + +Toto[6]（Toto 2.0）是由 Datadog 推出的新一代时间序列基础模型（V2.0.10 及以后版本支持），主要面向可观测性场景中的时间序列预测。当前集成的模型采用 2.5B 参数版本，基于 Decoder-only Patch Transformer 架构，交替使用因果时间注意力与变量注意力，对时间维度和变量维度进行联合建模。其核心特性包括： + +- **单变量与多变量预测**：支持单个目标变量以及多个相关目标变量的联合预测，适用于 CPU、内存、网络流量等可观测性指标的分析场景。 +- **概率性预测能力**：模型输出 0.1～0.9 的固定分位数，可刻画预测结果的不确定性；AINode 使用 p50 中位数作为点预测结果。 +- **高效块解码**：采用缓存式块解码机制，可分块生成预测结果，降低较长预测范围下的重复计算开销。 +- **大规模模型能力**：通过扩大模型参数量和预训练数据规模提升预测性能，具备良好的零样本泛化能力。 +- **协变量限制**：当前集成版本不支持历史协变量或已知未来协变量。 + +![](/img/LargeModel-toto.png) + +## 9. 效果展示 时序大模型能够适应多种不同领域和场景的真实时序数据，在各种任务上拥有优异的处理效果，以下是在不同数据上的真实表现： @@ -100,7 +127,7 @@ Chronos-2 [4] ![](/img/LargeModel05.png) -## 8. 部署使用 +## 10. 部署使用 1. 打开 IoTDB cli 控制台，检查 ConfigNode、DataNode、AINode 节点确保均为 Running。 @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer- Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref1) -**[2]** TIMER-XL- LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING ,Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.** [↩ 返回](#ref4) -**[3]** Sundial- A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ 返回](#ref5) -**[4] **Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.**[↩ 返回](#ref4) +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ 返回](#ref6) diff --git a/src/zh/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md b/src/zh/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md index a86a0ea7c..255eaff80 100644 --- a/src/zh/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md +++ b/src/zh/UserGuide/latest/AI-capability/AINode_Upgrade_apache.md @@ -419,7 +419,12 @@ IoTDB> show models | custom| | user_defined| active| | timer_xl| timer| builtin| activating| | sundial| sundial| builtin| active| +| sundialx_1| sundial| fine_tuned| active| +| sundialx_4| sundial| fine_tuned| training| +| sundialx_5| sundial| fine_tuned| failed| | chronos2| t5| builtin| inactive| +| moirai2| moirai| builtin| inactive| +| toto| toto| builtin| inactive| +---------------------+--------------+--------------+-------------+ ``` @@ -427,22 +432,24 @@ IoTDB> show models | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |----------------------------------| ----------------------------------------------------------------------------------------- | ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | -| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。2. 需要选择参数 (p,d,q)。3. 对缺失值敏感。 | -| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。2. 对近期数据赋予更高权重。3. 简单易实现。 | -| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。2. 适合平稳或缓慢变化序列。3. 可扩展为双指数或三指数平滑。 | -| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。2. 对突发变化敏感。3. 季节性朴素变体可用前一季节同期值预测。 | -| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。2. 对异常值稳健。3. 分解后可结合其他模型预测各分量。 | -| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。2. 假设观测值在给定状态下独立。3. 需指定隐藏状态数量。 | -| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。2. 参数更多，计算复杂度高。3. 需训练GMM成分数。 | -| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。2. 可处理高维数据。3. 对全局异常敏感，局部异常可能漏检。 | +| **ARIMA**(自回归整合移动平均模型) | 结合自回归(AR)、差分(I)和移动平均(MA)，用于预测平稳时间序列或可通过差分变为平稳的数据。 | 单变量时间序列预测，如股票价格、销量、经济指标等。| 1. 适用于线性趋势和季节性较弱的数据。
2. 需要选择参数 (p,d,q)。
3. 对缺失值敏感。 | +| **Holt-Winters**(三参数指数平滑) | 基于指数平滑，引入水平、趋势和季节性三个分量，适用于具有趋势和季节性的数据。 | 有明显季节性和趋势的时间序列，如月度销售额、电力需求等。 | 1. 可处理加性或乘性季节性。
2. 对近期数据赋予更高权重。
3. 简单易实现。 | +| **Exponential Smoothing**(指数平滑) | 通过加权平均历史数据，权重随时间指数递减，强调近期观测值的重要性。 | 无显著季节性但存在趋势的数据，如短期需求预测。 | 1. 参数少，计算简单。
2. 适合平稳或缓慢变化序列。
3. 可扩展为双指数或三指数平滑。 | +| **Naive Forecaster**(朴素预测器) | 使用最近一期的观测值作为下一期的预测值，是最简单的基准模型。 | 作为其他模型的比较基准，或数据无明显模式时的简单预测。 | 1. 无需训练。
2. 对突发变化敏感。
3. 季节性朴素变体可用前一季节同期值预测。 | +| **STL Forecaster**(季节趋势分解预测) | 基于STL分解时间序列，分别预测趋势、季节性和残差分量后组合。 | 具有复杂季节性、趋势和非线性模式的数据，如气候数据、交通流量。 | 1. 能处理非固定季节性。
2. 对异常值稳健。
3. 分解后可结合其他模型预测各分量。 | +| **Gaussian HMM**(高斯隐马尔可夫模型) | 假设观测数据由隐藏状态生成，每个状态的观测概率服从高斯分布。 | 状态序列预测或分类，如语音识别、金融状态识别。 | 1. 适用于时序数据的状态建模。
2. 假设观测值在给定状态下独立。
3. 需指定隐藏状态数量。 | +| **GMM HMM** (高斯混合隐马尔可夫模型) | 扩展Gaussian HMM，每个状态的观测概率由高斯混合模型描述，可捕捉更复杂的观测分布。 | 需要多模态观测分布的场景，如复杂动作识别、生物信号分析。 | 1. 比单一高斯更灵活。
2. 参数更多，计算复杂度高。
3. 需训练GMM成分数。 | +| **STRAY**(基于奇异值的异常检测) | 通过奇异值分解(SVD)检测高维数据中的异常点，常用于时间序列异常检测。 | 高维时间序列的异常检测，如传感器网络、IT系统监控。 | 1. 无需分布假设。
2. 可处理高维数据。
3. 对全局异常敏感，局部异常可能漏检。 | 内置时序大模型介绍如下： | 模型名称 | 核心概念 | 适用场景 | 主要特点 | |---------------| ---------------------------------------------------------------------- | ------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。2. 多场景覆盖，支持非平稳、多变量及协变量预测。3. 基于万亿级高质量工业时序数据预训练。 | -| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测 2. 可灵活分析预测分布的任意统计特性。3. 创新生成架构，实现高效的非确定性样本生成。 | -| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。2. 支持协变量统一建模，但对输入有严格要求：a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；b. 每个历史协变量的长度必须等于目标变量的长度； c. 每个未来协变量的长度必须等于预测长度；3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Timer-XL** | 支持超长上下文的时序大模型，通过大规模工业数据预训练增强泛化能力。 | 需利用极长历史数据的复杂工业预测，如能源、航空航天、交通等领域。 | 1. 超长上下文支持，可处理数万时间点输入。
2. 多场景覆盖，支持非平稳、多变量及协变量预测。
3. 基于万亿级高质量工业时序数据预训练。 | +| **Timer-Sundial** | 采用“Transformer + TimeFlow”架构的生成式基础模型，专注于概率预测。 | 需要量化不确定性的零样本预测场景，如金融、供应链、新能源发电预测。 | 1. 强大的零样本泛化能力，支持点预测与概率预测
2. 可灵活分析预测分布的任意统计特性。
3. 创新生成架构，实现高效的非确定性样本生成。 | +| **Chronos-2** | 基于离散词元化范式的通用时序基础模型，将预测转化为语言建模任务。 | 快速零样本单变量预测，以及可借助协变量（如促销、天气）提升效果的场景。 | 1. 强大的零样本概率预测能力。
2. 支持协变量统一建模，但对输入有严格要求：
  a. 未来协变量的名称组成的集合必须是历史协变量的名称组成的集合的子集；
  b. 每个历史协变量的长度必须等于目标变量的长度；
  c. 每个未来协变量的长度必须等于预测长度；
3. 采用高效的编码器式结构，兼顾性能与推理速度。 | +| **Moirai 2.0** | 采用轻量级 Decoder-only Patch Transformer，通过单一 Patch 尺寸、多 Token 预测和多分位数输出实现高效的单变量预测。 | 适用于对模型体量和推理效率要求较高的零样本单变量预测，如工业监测、能源负荷及设备指标预测。 | 1. 模型参数量约 11.4M。
2. 每个解码步可预测多个 Patch，降低长预测范围的自回归开销。
3. 输出 9 个分位数（0.1～0.9），使用 p50 中位数作为点预测。
4. 使用实例归一化缓解序列分布漂移。
5. 不支持多变量和协变量。 | +| **Toto 2.0** | 采用 Decoder-only Patch Transformer，交替使用因果时间注意力与变量注意力，对时间维和变量维进行联合建模。 | 面向可观测性指标等多变量时间序列的零样本预测，如 CPU、内存、网络流量等指标的联合预测。 | 1. 支持单变量和多变量目标预测。
2. 输出 0.1～0.9 的固定分位数，使用 p50 中位数作为点预测。
3. 支持缓存式块解码，可高效扩展较长预测范围。
4. 当前不支持协变量。 | ### 4.4 删除模型 diff --git a/src/zh/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md b/src/zh/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md index 99712b134..0a75fce5b 100644 --- a/src/zh/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md +++ b/src/zh/UserGuide/latest/AI-capability/TimeSeries-Large-Model_Upgrade_apache.md @@ -78,7 +78,34 @@ Chronos-2 [4] ![](/img/timeseries-large-model-chronos2.png) -## 7. 效果展示 +## 7. Moirai2 模型 + +Moirai2[5]（Moirai 2.0）是由 Salesforce AI Research 推出的通用时间序列基础模型（V2.0.10 及以后版本支持）。当前集成的是 Moirai 2.0 R-small 版本，模型参数量约为 11.4M。与采用掩码编码器架构的 Moirai 1.0 不同，Moirai 2.0 使用因果 Decoder-only Patch Transformer，并通过单一 Patch 尺寸、多 Token 预测和多分位数输出，在较小模型规模下实现高效的单变量预测。其核心特性包括： + +- **轻量化模型架构**：采用 Decoder-only Patch Transformer，并结合 RMSNorm、旋转位置编码和 SiLU-GLU 前馈网络，在较小参数规模下兼顾预测能力和推理效率。 +- **多 Token 预测**：每个解码步骤可同时预测多个 Patch，减少长预测范围下所需的自回归解码次数。 +- **概率性预测能力**：模型输出 0.1～0.9 共 9 个分位数，AINode 使用 p50 中位数作为点预测结果。 +- **Patch 解码**：在进入注意力模块前，将时间序列按照固定大小的 Patch 进行分组，以提高时序特征提取和解码效率。 +- **实例归一化**：在模型输入前对每条时间序列进行标准化，并在输出后执行反归一化，以缓解不同序列间的分布漂移。 +- **输入范围**：模型聚焦于单变量预测，不支持多变量目标及协变量输入。 + +![](/img/LargeModel-moirai2.png) + +> 注意：Moirai 2.0 R-small 模型权重采用 CC BY-NC 4.0 许可证，仅限研究用途。 + +## 8. Toto 模型 + +Toto[6]（Toto 2.0）是由 Datadog 推出的新一代时间序列基础模型（V2.0.10 及以后版本支持），主要面向可观测性场景中的时间序列预测。当前集成的模型采用 2.5B 参数版本，基于 Decoder-only Patch Transformer 架构，交替使用因果时间注意力与变量注意力，对时间维度和变量维度进行联合建模。其核心特性包括： + +- **单变量与多变量预测**：支持单个目标变量以及多个相关目标变量的联合预测，适用于 CPU、内存、网络流量等可观测性指标的分析场景。 +- **概率性预测能力**：模型输出 0.1～0.9 的固定分位数，可刻画预测结果的不确定性；AINode 使用 p50 中位数作为点预测结果。 +- **高效块解码**：采用缓存式块解码机制，可分块生成预测结果，降低较长预测范围下的重复计算开销。 +- **大规模模型能力**：通过扩大模型参数量和预训练数据规模提升预测性能，具备良好的零样本泛化能力。 +- **协变量限制**：当前集成版本不支持历史协变量或已知未来协变量。 + +![](/img/LargeModel-toto.png) + +## 9. 效果展示 时序大模型能够适应多种不同领域和场景的真实时序数据，在各种任务上拥有优异的处理效果，以下是在不同数据上的真实表现： @@ -100,7 +127,7 @@ Chronos-2 [4] ![](/img/LargeModel05.png) -## 8. 部署使用 +## 10. 部署使用 1. 打开 IoTDB cli 控制台，检查 ConfigNode、DataNode、AINode 节点确保均为 Running。 @@ -143,6 +170,8 @@ IoTDB> show models | timer_xl| timer| builtin| active| | sundial| sundial| builtin| active| | chronos2| t5| builtin| active| +| moirai2| moirai| builtin| active| +| toto| toto| builtin| active| +---------------------+---------+--------+--------+ ``` @@ -150,8 +179,12 @@ IoTDB> show models **[1]** Timer- Generative Pre-trained Transformers Are Large Time Series Models, Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref1) -**[2]** TIMER-XL- LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING ,Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) +**[2]** TIMER-XL: LONG-CONTEXT TRANSFORMERS FOR UNIFIED TIME SERIES FORECASTING, Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, Mingsheng Long. [↩ 返回](#ref2) + +**[3]** Sundial: A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) + +**[4]** Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.** [↩ 返回](#ref4) -**[3]** Sundial- A Family of Highly Capable Time Series Foundation Models, Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, Mingsheng Long, **ICML 2025 spotlight**. [↩ 返回](#ref3) +[5] Moirai 2.0: When Less Is More for Time Series Forecasting, Salesforce AI Research, arXiv:2511.11698. [↩ 返回](#ref5) -**[4] **Chronos-2: From Univariate to Universal Forecasting, Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider, **arXiv:2510.15821.**[↩ 返回](#ref4) +[6] Toto 2.0: Time Series Forecasting Enters the Scaling Era, Datadog, arXiv:2605.20119. [↩ 返回](#ref6)