Multi-temporal-spatial-scale temporal convolution network for short-term load forecasting of power systems

Abstract

With the advancement of power market reform, accurate load forecasting can ensure the stable operation of power systems increasingly. The randomness of feature change such as climate and day type increases the complexity of short-term load forecasting. To simplify the data processing process to facilitate the practical application and predict short-term loads more accurately, this paper takes the past load data as a feature and considers the time series characteristics of load data simultaneously. The multi-temporal-spatial-scale method is applied to process the load data by reducing the load data noise error and enhancing the time series characteristics. Then, a novel short-term load forecasting model, which is named a multi-temporal-spatial-scale temporal convolutional network, is applied to load forecasting tasks in this paper. The proposed approach can learn the nonlinear feature and time series characteristics of load data simultaneously. To predict the power load of a city in Guangxi Zhuang Autonomous Region (China) in the next day and the next week, the forecasting model is trained by the historical feature load of 7 days, 21 days, 99 days, and 199 days. Compared with 22 artificial intelligent short-term load forecasting models, such as backpropagation neural network and bagging regression, the simulation results show that the proposed multi-temporal-spatial-scale temporal convolutional network can obtain higher accuracy for the short-term load forecasting of power systems than other compared methods.