An Improved Informer Model for Short-Term Load Forecasting by Considering Periodic Property of Load Profiles

Abstract

Short-term load forecasting (STLF) is an important but a difficult task due to the uncertainty and complexity of electric power systems. In recent times, an attention-based model, Informer, has been proposed for efficient feature learning of lone sequences. To solve the quadratic complexity of traditional method, this model designs what is called ProbSparse self-attention mechanism. However, this mechanism may neglect daily-cycle property of load profiles, affecting its performance of STLF. To solve this problem, this study proposes an improved Informer model for STLF by considering the periodic property of load profiles. The improved model concatenates the output of Informer, the periodic load values of input sequences, and outputs forecasting results through a fully connected layer. This makes the improved model could not only inherit the superior ability of the traditional model for the feature learning of long sequences, but also extract periodic features of load profiles. The experimental results on three public data sets showed its superior performance than the traditional Informer model and others for STLF.