Short-Term State Electricity Load Forecasting Based on Transfer-Informer

Abstract

The worldwide COVID-19 pandemic has caused an enormous impact on the operation mode of human society. Such sudden events bring sharp fluctuations and data inadequacy in datasets of several areas, which leads to challenges in solving related problems. Traditional deep learning models like CNN have shown relatively poor performance with small datasets during the COVID-19 pandemic. This is because the data insufficiency and fluctuations lead to serious problems in the training process. In our work, an Informer framework combined with Transfer learning methods (Transfer-Informer) is proposed to solve the data insufficiency in emergency situations, as well as to provide a more efficient self-attention mechanism for deep feature mining, with two distinctive advantages: (1) The ProbSpares self-attention mechanisms, which enables the proposed model to highlight dominant information and extract more typical features from time-series datasets. (2) The Transfer learning framework improves the generalization capability of the model, by transferring basic knowledge from normal situations to emergency cases with fewer data. In our experiments, Transfer-Informer is applied to short-term load forecasting, which achieves better predicting accuracy than traditional models. The empirical results indicate that the proposed model has put forward a baseline for short-term load forecasting in emergency situations and provided a feasible method to tackle sudden fluctuations in real problem-solving.