An Effective N-BEATS Network Model for Short Term Load Forecasting

Abstract

This paper addresses the issue of short-term load forecasting in the power system, a domain where risks are liable to take place during the peak period of power consumption. To prevent such risks, it is crucial to have a precise load forecasting that can be carried out beforehand, allowing for the arrangement of the power peak period in advance, and greatly minimize the occurrence of power accidents. To achieve this goal, a power load forecasting method of high reliability and precision is indispensable. In this paper, an effective approach is proposed that is capable of effectively resolving short- and medium-term power forecasting issues. Prior to the forecasting, the Seasonal-Trend decomposition procedure based on Loess (The full term for Loess is locally weighted scatterplot smoothing, LOWESS or LOESS) (STL) time series decomposition is conducted on the data set, enabling the observation of the trend, periodicity, and corresponding residual items. Based on the STL decomposition results, the prediction lengths of the previous and subsequent items in the N-BEATS (Neural basis expansion analysis for interpretable time series forecasting) network model are adjusted, and the corresponding network structure is altered to identify the optimal model through continuous experiments. The method’s efficacy was evaluated on datasets obtained from two distinct regions, and the results indicated a better performance compare with other similar forecasting methods in terms of accuracy and forecasting bias.