Patch Time Series Transformer in Hugging Face — Forecasting on Electricity Dataset

AI Impact Summary

The Patch Time Series Transformer (PatchTST) model, developed by IBM, offers a novel approach to time series forecasting by vectorizing univariate time series into patches. This model leverages a Transformer architecture to capture long-term dependencies within the data, and demonstrates transfer learning capabilities by adapting a pre-trained model to new domains without retraining. The provided code demonstrates a basic workflow for forecasting on the Electricity dataset, showcasing the model's ability to perform zero-shot forecasting and further refine its performance through linear probing and fine-tuning.