sktime/pytorch-forecasting

What it solves

PyTorch Forecasting simplifies the implementation of state-of-the-art deep learning architectures for time series forecasting. It bridges the gap between complex research models and real-world deployment by providing a high-level API that handles the tedious aspects of data preparation, training, and evaluation for time series data.

How it works

Built on PyTorch Lightning, the library provides a structured workflow for forecasting:

• Data Handling: A specialized TimeSeriesDataSet class manages variable transformations, missing values, randomized subsampling, and multiple history lengths.
• Model Architectures: It implements several advanced neural networks (like Temporal Fusion Transformers, N-BEATS, and PatchTST) that are optimized for real-world use and include built-in interpretation capabilities.
• Training & Scaling: It leverages PyTorch Lightning to automatically scale training across CPUs and single or multiple GPUs with integrated logging via TensorBoard.
• Optimization: It integrates with Optuna for hyperparameter tuning and includes multi-horizon time series metrics to evaluate performance.

Who it’s for

The project is designed for both professionals who need maximum flexibility for deployment and beginners who require reasonable defaults to get started with neural network-based forecasting.

Highlights

• Diverse Model Zoo: Includes high-performance models such as Temporal Fusion Transformers (TFT), N-BEATS, N-HiTS, DeepAR, and PatchTST.
• Integrated Data Pipeline: Simplifies the conversion of pandas DataFrames into model-ready time series datasets.
• Hardware Agnostic: Out-of-the-box support for CPU and GPU training via PyTorch Lightning.
• Interpretability: Provides models with built-in capabilities to interpret predictions.