Accurate short-term electric load forecasting via modal-aware preprocessing and hybrid temporal modeling

Abstract Accurate short-term electricity load forecasting (STLF) is essential for improving power system dispatch efficiency and ensuring supply reliability. However, the nonlinear and non-stationary characteristics of load data pose significant challenges to existing forecasting models. To address these limitations, this study proposes a hybrid forecasting framework that integrates mode-aware preprocessing with parallel deep temporal modeling. First, a robust STL decomposition isolates macro-scale trend and seasonality from the raw sequence. The residual component, which contains micro-scale nonlinear load dynamics, is then dissected using an improved variational mode decomposition (VMD). To overcome the subjectivity of manual parameter tuning, a Frequency-Energy Entropy Adaptive (FEA) strategy is developed to objectively determine the optimal number of decomposition modes, while a Lightweight Sparrow Search Algorithm (LSSA) is designed to efficiently optimize VMD parameters. The decomposed components are subsequently fed into a parallel Informer-LSTM hybrid, which synergistically captures global temporal dependencies and local fluctuation patterns. Experiments on real-world datasets from China and Australia show that the proposed method achieves state-of-the-art performance, achieving an R 2 of 0.99926 and reducing the Mean Squared Error (MSE) by 77.29% on the Australian dataset. The results confirm the framework’s superior accuracy, robustness, and cross-regional generalization capability.