Short-Term Load Forecasting with Temporal Fusion Transformers for Power Distribution Networks

Abstract

Short-Term Load Forecasting (STLF) is essential to the operation and management of modern power distribution networks. Accurate STLF can significantly improve the demand-side management of the power system. In this paper, a new method with high-performance forecasting performance is presented to forecast short-term loads with deep learning. The temporal fusion transformers (TFT) approach is an attention-based deep learning model with interpretable insights into temporal dynamics. The sequence-to-sequence model processes the historical and future covariates to enhance the forecasting performance. Gated Residual Network (GRN) is applied to drop out unnecessary information and improve efficiency. The proposed method is tested on anonymized data from a university campus. The anomalies and missing data are imputed with the k-nearest neighbor (KNN) method. The testing results demonstrate the effectiveness of the proposed method.