Short-term load forecasting based on strategies of daily load classification and feature set reconstruction

Accurate short-term load forecasting (STLF) is an important aspect for the safe and stable operation of the power system. When multiple factors are taken into consideration, the input data preprocessing method and the variable selection strategy are essential for improving the accuracy of STLF. In this article, both are discussed, and two improved algorithms are proposed to deal with massive data. Firstly, a cosine similarity-based cluster analysis is introduced to enhance the efficiency of model learning by extracting the information of time variables and summarizing in classification labels. Secondly, considering the relationship between load and multiple factors at hourly intervals, a feature selection method based on reconstruction of feature set and maximal information coefficient (MIC) is proposed to achieve an optimized performance. Then, temporal convolutional network (TCN) with great learning ability for high-dimensional data is adopted as the forecasting model. Experimental results prove that the improved CL-FR-MIC-TCN model can achieve higher prediction accuracy, and the proposed strategies can also effectively improve the accuracy of multiple classic forecasting models.