Day–Ahead Load Forecasting Using MSTL–VMD–Based Multi–Seasonal Decomposition and LSTM–Based Component–Wise Prediction

This paper proposes a day-ahead load forecasting method that improves interpretability through multi-seasonal decomposition and component-wise prediction. System load exhibits superimposed trend, seasonalities, and irregular fluctuations, which limits direct forecasting with a single model. MSTL (Multi-Seasonal Trend decomposition using Loess) decomposes the load into trend, weekly, daily, and residual components, and VMD (Variational Mode Decomposition) further separates the residual into multiple frequency modes. The trend, seasonal, and VMD-based residual components are predicted using LSTM (Long Short-Term Memory) models. Case studies on the hourly nationwide gross system load during normal days in 2024 show that the proposed method achieves a MAPE of 1.89% and an RMSE of 1572.07 MW, outperforming a single LSTM model with a MAPE of 2.44% and an RMSE of 2268.86 MW. In addition, the proposed framework enables identification of dominant error sources, demonstrating improved accuracy and interpretability.