A hybrid Autoformer framework for electricity demand forecasting

Abstract

Electricity demand forecasting is of great significance to the electricity system and residents’ life, but it is difficult to forecast the electricity demand series because of the influence of cyclical factors. Electricity demand forecasting also faces the problem of small data amounts. Therefore, we need to design a model that is less affected by data volume and can cope with complex electricity demand series. Based on the Autoformer model, this paper establishes a novel forecasting framework with excellent performance. In the part of data preprocessing, multiple linear regression with 10 variables and Bootstrap processing are added. In the part of the model, the Auto-Correlation mechanism is modified to better extract the historical and nonlinear characteristics of electricity demand series from different time spans. Using this framework, we further analyze the impact of working days and seasonal changes on the electricity demand in Taixing City and New South Wales. In addition, we propose a new electricity demand forecasting method, which can adjust the original sequence according to the actual situation. The experimental results show that this method can achieve good precision in demand forecasting. Taking Taixing of China and New South Wales of Australia as examples, the forecasting performance with the proposed framework is better than that of Autoformer, Reformer, Informer, and other mainstream models. The forecasting indexes with our proposed framework of the test set are MAE: 35.05, RMSE: 47.28, MAPE: 1.63 in Taixing and MAE: 193.17, RMSE: 239.96, MAPE: 2.43 in NSW.

Keywords

• Autoformer
• Bootstrap
• Auto-correlation
• Deep learning
• Prediction