Short-Term Net Load Forecasting Based on MSTL Decomposition and PGA-TimeXer

Accurate short-term net load forecasting is challenging due to the superimposition of meteorological influences, daily and weekly seasonal patterns, and irregular disturbances in the observed series. On the one hand, redundant meteorological input data risks hindering the learning process. On the other hand, methods that directly model the raw time series may struggle to distinguish these heterogeneous patterns. To address the above issues, this paper proposes a component-wise hybrid framework combining multiple seasonal-trend decomposition using loess (MSTL) and phototropic growth algorithm (PGA)-optimized TimeXer with the selected features. First, both the Pearson correlation coefficient and Shapley additive explanation are employed for meteorological feature selection. Then, MSTL is utilized to decompose the net load into trend, daily seasonal, weekly seasonal, and residual components. Following this, independent PGA-TimeXer models are trained to forecast the trend and seasonal components, while a light gradient boosting machine (LightGBM) is designed to forecast the residual component from lagged residual features. By recombining the forecasting results from each component, the final net load prediction is achievable. Finally, field measurement data from an actual 220 kV substation verify the effectiveness of the method proposed in this paper.