Time series power load forecasting based on PCC-MIC hybrid feature selection and differential attention improved PatchTST

Accurate power load forecasting is essential for the secure and stable operation of power systems. Nevertheless, the inherent complexity and stochasticity of load time series pose great challenges to precise prediction. In this paper, we propose a novel forecasting framework integrated with hybrid feature selection and Differential Attention mechanism, namely Seasonal and Trend decomposition using Loess (STL) enhanced Patch Time Series Transformer (PatchTST), abbreviated as STL-PatchTDT. First, key features are selected using Pearson Correlation Coefficient (PCC) and Maximum Information Coefficient (MIC). Next, the STL technique decomposes load series into trend, seasonal, and residual components. Taking the decomposition and feature selection results as input to the model, the forecasting model enhances the foundational PatchTST model by incorporating a “Differential Attention” mechanism, thereby boosting its ability to model complex patterns. Using multi-load data from Arizona State University’s Tempe Campus, experiments show the proposed method improves R 2 by 2.7%, and reduces MAE and RMSE by 21.7% and 15.5%, respectively, outperforming baseline models and achieving high-precision forecasting in complex scenarios.