时间序列库（TSLib）

TSLib 是一个面向深度学习研究者的开源库，特别适用于深度时间序列分析。

English README：README.md

我们提供了一个整洁的代码库，用于评测先进的深度时间序列模型或开发自定义模型，覆盖 长短期预测、插补、异常检测和分类 等五大主流任务。

:triangular_flag_on_post:最新动态（2025.12）非常感谢 ailuntz 的杰出贡献，提供了更新的依赖要求和 Docker 部署，以及完善的文档。这对本项目和初学者都很有意义。

:triangular_flag_on_post:最新动态（2025.11）鉴于大型时间序列模型（LTSM）的快速发展，我们在 TSLib 中新增了[零样本预测]功能，可参考 此脚本 评测 LTSM。

:triangular_flag_on_post:最新动态（2025.10）针对近期研究者在标准基准上追求微小提升而产生的困惑，我们提出了[精度定律]，以刻画深度时间序列预测任务的目标，并可据此识别已饱和的数据集。

:triangular_flag_on_post:最新动态（2024.10）我们已纳入 [TimeXer]，其定义了一个实用的预测范式：带外生变量的预测。考虑到实用性与计算效率，我们认为 TimeXer 所定义的新范式将成为未来研究的“正确”任务。

:triangular_flag_on_post:最新动态（2024.10）实验室已开源 [OpenLTM]，提供了有别于 TSLib 的预训练 - 微调范式。如果您对大型时间序列模型感兴趣，该仓库值得参考。

:triangular_flag_on_post:最新动态（2024.07）我们撰写了关于[深度时间序列模型]的综述，并基于 TSLib 构建了严谨的基准。论文总结了当前时间序列模型的设计原则，并通过深入实验验证，期望对未来研究有所帮助。

:triangular_flag_on_post:最新动态（2024.04）感谢 frecklebars 的贡献，著名的序列模型 Mamba 已加入本库。参见该文件，需要先用 pip 安装 mamba_ssm。

:triangular_flag_on_post:最新动态（2024.03）鉴于各论文使用的回溯窗口长度不一致，我们将排行榜中的长期预测拆分为 Look-Back-96 与 Look-Back-Searching 两类。建议阅读 TimeMixer，其实验同时包含两种窗口设置，更具科学性。

:triangular_flag_on_post:最新动态（2023.10）我们添加了 iTransformer 的实现，这是长期预测领域的最新 SOTA。官方代码与完整脚本参见 此处。

:triangular_flag_on_post:最新动态（2023.09）我们为 TimesNet 及本库添加了详细教程，对时间序列初学者十分友好。

:triangular_flag_on_post:最新动态（2023.02）我们发布了 TSlib，作为一个面向时间序列模型的综合基准与代码库，扩展自此前的 Autoformer 仓库。

时间序列分析排行榜

截至 2024 年 3 月，各任务排行榜前三名如下：

| 模型

排名 | 长期预测

Look-Back-96 | 长期预测

Look-Back-Searching | 短期预测 | 插补 | 分类 | 异常检测 | | ------------ | ------------------------ | -------------------------------- | -------- | ---- | ---- | -------- | | 🥇 第一名 | TimeXer | TimeMixer | TimesNet | TimesNet | TimesNet | TimesNet | | 🥈 第二名 | iTransformer | PatchTST | [Non-stationary

Transformer](https://github.com/thuml/Nonstationary_Transformers) | [Non-stationary

Transformer](https://github.com/thuml/Nonstationary_Transformers) | [Non-stationary

Transformer](https://github.com/thuml/Nonstationary_Transformers) | FEDformer | | 🥉 第三名 | TimeMixer | DLinear | FEDformer | Autoformer | Informer | Autoformer |

说明：排行榜会持续更新。 如果您提出了先进的模型，可通过发送论文或代码链接、或提交 PR 与我们联系，我们会尽快将其加入仓库并更新排行榜。

排行榜中的对比模型（☑ 表示代码已收录）。 - [x] TimeXer - TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables [NeurIPS 2024] [代码] - [x] TimeMixer - TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting [ICLR 2024] [代码] - [x] TSMixer - TSMixer: An All-MLP Architecture for Time Series Forecasting [arXiv 2023] [代码] - [x] iTransformer - iTransformer: Inverted Transformers Are Effective for Time Series Forecasting [ICLR 2024] [代码] - [x] PatchTST - A Time Series is Worth 64 Words: Long-term Forecasting with Transformers [ICLR 2023] [代码] - [x] TimesNet - TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis [ICLR 2023] [代码] - [x] DLinear - Are Transformers Effective for Time Series Forecasting? [AAAI 2023] [代码] - [x] LightTS - Less Is More: Fast Multivariate Time Series Forecasting with Light Sampling-oriented MLP Structures [arXiv 2022] [代码] - [x] ETSformer - ETSformer: Exponential Smoothing Transformers for Time-series Forecasting [arXiv 2022] [代码] - [x] Non-stationary Transformer - Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting [NeurIPS 2022] [代码] - [x] FEDformer - FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting [ICML 2022] [代码] - [x] Pyraformer - Pyraformer: Low-complexity Pyramidal Attention for Long-range Time Series Modeling and Forecasting [ICLR 2022] [代码] - [x] Autoformer - Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting [NeurIPS 2021] [代码] - [x] Informer - Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting [AAAI 2021] [代码] - [x] Reformer - Reformer: The Efficient Transformer [ICLR 2020] [代码] - [x] Transformer - Attention is All You Need [NeurIPS 2017] [代码]

更多详情可参考我们关于 [TimesNet] 的最新论文，实时在线版本即将发布。

新增基线模型（综合评测后将加入排行榜）。 - [x] MambaSL - MambaSL: Exploring Single-Layer Mamba for Time Series Classification [ICLR 2026] [Code] - [x] TimeFilter - TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting [ICML 2025] [代码] - [x] KAN-AD - KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks [ICML 2025] [代码] - [x] MultiPatchFormer - A multiscale model for multivariate time series forecasting [Scientific Reports 2025] [代码] - [x] WPMixer - WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting [AAAI 2025] [代码] - [x] MSGNet - MSGNet: Learning Multi-Scale Inter-Series Correlations for Multivariate Time Series Forecasting [AAAI 2024] [代码] - [x] PAttn - Are Language Models Actually Useful for Time Series Forecasting? [NeurIPS 2024] [代码] - [x] Mamba - Mamba: Linear-Time Sequence Modeling with Selective State Spaces [arXiv 2023] [代码] - [x] SegRNN - SegRNN: Segment Recurrent Neural Network for Long-Term Time Series Forecasting [arXiv 2023] [代码] - [x] Koopa - Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors [NeurIPS 2023] [代码] - [x] FreTS - Frequency-domain MLPs are More Effective Learners in Time Series Forecasting [NeurIPS 2023] [代码] - [x] MICN - MICN: Multi-scale Local and Global Context Modeling for Long-term Series Forecasting [ICLR 2023] [代码] - [x] Crossformer - Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting [ICLR 2023] [代码] - [x] TiDE - Long-term Forecasting with TiDE: Time-series Dense Encoder [arXiv 2023] [代码] - [x] SCINet - SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction [NeurIPS 2022] [代码] - [x] FiLM - FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting [NeurIPS 2022] [代码] - [x] TFT - Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting [arXiv 2019] [代码]

新增大型时间序列模型。本库同样支持以下 LTSM 的零样本评测：

• [x] Chronos2 - Chronos-2: From Univariate to Universal Forecasting [arXiv 2025] [代码]
• [x] TiRex - TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning [NeurIPS 2025] [代码]
• [x] Sundial - Sundial: A Family of Highly Capable Time Series Foundation Models [ICML 2025] [代码]
• [x] Time-MoE - Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts [ICLR 2025] [代码]
• [x] Toto - Toto: Time Series Optimized Transformer for Observability [arXiv 2024]
• [x] Chronos - Chronos: Learning the Language of Time Series [TMLR 2024] [代码]
• [x] Moirai - Unified Training of Universal Time Series Forecasting Transformers [ICML 2024]
• [x] TimesFM - TimesFM: A decoder-only foundation model for time-series forecasting [ICML 2024] [代码]

快速开始

准备数据

可从 [Google Drive]、[Baidu Drive] 或 [Hugging Face] 下载预处理数据，并置于 ./dataset 目录。

安装

1. 克隆本仓库 bash git clone https://github.com/thuml/Time-Series-Library.git cd Time-Series-Library

2. 创建新的 Conda 环境 bash conda create -n tslib python=3.11 conda activate tslib

3. 安装核心依赖

   ⚠️ CUDA 兼容性提示 torch 预编译包与 CUDA 版本强相关。（查看 https://pytorch.org/get-started/previous-versions/ ） 请确保torch安装与本地 CUDA 版本匹配的包（如 cu118 或 cu121）。 推荐torch==2