A Patch TST-Bi Transformer Hybrid Framework for Short-Term Power Load Forecasting

Hanyang Wei

Abstract: This paper proposes a hybrid Patch TST-Bi Transformer framework for short-term power load forecasting to simultaneously capture local temporal patterns and long-range dependencies. Historical sequences are partitioned into temporal patches and encoded using a channel-independent Patch TST backbone, while a Bidirectional Transformer models global contextual relationships across patches. Reversible Instance Normalization and moving average decomposition are incorporated to improve robustness against non-stationarity and enhance long-horizon forecasting performance. Experiments conducted on benchmark and real-world power load datasets demonstrate consistent improvements over Autoformer, Informer, FEDformer, and other baseline models. The proposed framework achieves higher forecasting accuracy while maintaining computational efficiency and robust generalization.

Keywords: PatchTST, BiTransformer, Power Load Forecasting, Time Series Forecasting, Transformer, Deep Learning, RevIN, Smart Grid