Probabilistic net load forecasting based on transformer network and Gaussian process-enabled residual modeling learning method

Abstract

Accurate net load forecasting plays an increasingly pivotal role in ensuring the reliable operation and scheduling of power systems. This paper introduces a novel probabilistic net load forecasting approach that combines the strengths of a Transformer network with Gaussian process regression. The state-of-the-art Transformer network is first employed to capture the net load pattern utilizing relatively abundant historical training samples. The remarkable temporal feature extraction ability allows it to discover the complex structure in net loads. Subsequently, the Gaussian Process is applied to capture the behavior of the Transformer network by modeling its forecasting residual utilizing a specific composite kernel. The modeling of the forecasting residual not only provides valuable uncertainty quantification of net load but also improves the forecasting performance based on the Transformer network. Comparative tests utilizing real-world data verify the superiority of the proposed method over other state-of-the-art net load forecasting algorithms.