Time-Series-Library — Advanced Deep Learning Models for Time Series

Introduction

Time-Series-Library (TSLib) is an open-source Python library from Tsinghua University that provides a unified codebase for advanced deep learning models applied to time series tasks. It enables researchers and practitioners to benchmark and deploy models for forecasting, classification, imputation, and anomaly detection using a consistent interface.

What Time-Series-Library Does

• Implements 15+ state-of-the-art deep time series models (TimesNet, PatchTST, iTransformer, DLinear, and more)
• Supports four core tasks: long-term forecasting, short-term forecasting, imputation, and anomaly detection
• Provides standardized experiment scripts for reproducible benchmarking
• Includes popular time series datasets with unified data loading pipelines
• Offers configurable training with GPU acceleration via PyTorch

Architecture Overview

TSLib follows a modular design with separate layers for data loading, model definition, training, and evaluation. Each model implements a common interface, allowing users to switch between architectures by changing a single configuration parameter. The training loop handles multi-GPU setups and supports early stopping, learning rate scheduling, and checkpoint saving.

Self-Hosting & Configuration

• Clone the repo and install dependencies via pip install -r requirements.txt
• Configure experiments through command-line arguments or shell scripts in the scripts/ directory
• Set --model to choose from available architectures (TimesNet, PatchTST, DLinear, FEDformer, etc.)
• Adjust --seq_len, --pred_len, and --label_len to control input/output horizons
• Place custom datasets in the dataset/ folder following the provided CSV format

Key Features

• Unified codebase covering multiple time series tasks under one framework
• Reproducible baselines with pre-configured experiment scripts
• Supports Transformer-based, MLP-based, and CNN-based model families
• Active maintenance with new models added as they are published
• Clean PyTorch implementation suitable for extension and research

Comparison with Similar Tools

• Darts — Higher-level API with statistical and ML models; TSLib focuses on deep learning research
• sktime — Scikit-learn-compatible framework for classical ML; TSLib targets neural architectures
• GluonTS — AWS probabilistic forecasting library; TSLib provides a broader model zoo
• NeuralForecast — Nixtla's neural forecasting; TSLib emphasizes multi-task support beyond forecasting

FAQ

Q: What Python and PyTorch versions are required? A: Python 3.8+ and PyTorch 1.8+ are recommended. Check requirements.txt for exact dependencies.

Q: Can I add my own model to TSLib? A: Yes. Implement the Model class interface in the models/ directory and register it in the experiment runner.

Q: Does it support multivariate time series? A: Yes. Most models support both univariate and multivariate settings via the --features flag (S, M, or MS).

Q: Is GPU training supported? A: Yes. Use --use_gpu and --gpu flags to select devices, with multi-GPU support available.

Sources

• https://github.com/thuml/Time-Series-Library
• https://arxiv.org/abs/2306.06142