MFformer: An improved transformer-based multi-frequency feature aggregation model for electricity load forecasting

Abstract

Accurate electric load forecasting is crucial for the effective operation and planning of power systems. To address issues such as the weak comprehensive prediction capability of power load forecasting models across the entire cycle and insufficient feature extraction of load volatility, an improved Transformer framework integrated with a multi-frequency feature model, MFformer (Multi-Frequency Transformer), is proposed. This model utilizes a multi-level decomposition hybrid framework with seasonal and trend feature encoders to decompose periodic and trend components. Key frequency points are selected based on power spectral density, with main frequency band information captured using a self-attention mechanism. An MLP layer hierarchically extracts low-frequency trend information, while wavelet decomposition combined with self-attention explores high-frequency random components. The final load prediction curve is reconstructed from these frequency bands. An experiment on the load dataset of a thermal power unit in Henan demonstrated that the model outperforms others in both mid- to long-term and short-term prediction accuracy, across various comparative models and forecast horizons, this model demonstrates average reductions in mean MAE and MSE errors of 27.70 % and 14.68 %, respectively.