PatchTST is a Transformer-based architecture for long-term time series forecasting. It segments time series into subseries-level patches that serve as input tokens to the Transformer, along with channel-independence where each univariate series shares embedding weights. The model achieves significant MSE and MAE reductions over prior Transformer-based and non-Transformer baselines. It has been adopted into GluonTS, NeuralForecast, and tsai libraries.