A novel electricity consumption forecasting model based on kernel extreme learning machine-with generalized maximum correntropy criterion

Abstract

Electricity consumption forecasting (ECF) plays a crucial role in the new open electricity market nowadays, and in order to pass the ECF test, the ECF error is required to be less than 5%. However, for historical electricity consumption (EC) data and the error distribution of ECF is usually nonlinear and non-Gaussian distributed, and some random events and data collection errors have a great impact on the forecasting accuracy. Therefore, the traditional forecasting methods designed based on mean squared error (MSE) have the problem of unstable forecasting results and difficult to improve the accuracy. To solve this problem, this paper first uses the Generalized Maximum Correntropy Criterion to optimize the Kernel Extreme Learning Machine, and proposes a new forecasting model of KELM-GMCC. Finally, the historical EC and corresponding temperature data of a commercial customer in China in 2018 and 2019 are used to test and guide the electricity prediction in 2020 and 2021. The model proposed in this paper has a mean absolute percentage error of 1.46% in forecasting daily EC using annual data. The results show that the proposed model is robust and has high forecasting accuracy compared with existing models such as Back Propagation (BP), Support Vector Regression (SVR), and Extreme Learning Machine (ELM).

Keywords

• Electricity consumption forecasting
• Generalized maximum correntropy criterion
• Kernel extreme learning machine