Short-term Load Forecasting Based on Time-Varying Feature Autoencoders

To improve the accuracy of short-term power load forecasting, a method based on time-varying feature autoencoders for power systems is proposed. This method optimizes the extraction and fusion of time-varying features from load time series data. Based on Fourier transform and inverse Fourier transform, the method performs autoencoding and decoding of periodic and trend components of load time series data. The frequency domain feature prediction process allows for the selection of efficient artificial intelligence methods. In this paper, the proposed algorithm was compared with the Autoformer, Informer, Transformer, Graph WaveNet, and MTGN models. The results fully validate the effectiveness and advantages of the time-varying feature autoencoder method in short-term load forecasting for power systems.