High-resolution day-ahead load forecasting under lunar holiday drift using causal wavelet decomposition and event-aware masked learning

High-resolution multi-step load forecasting becomes particularly difficult when the series contains structured extreme variations caused by recurring events. In such settings, conventional “data cleaning” can remove precisely the patterns that matter most for accurate prediction. This study addresses 48-step (half-hourly) day-ahead load forecasting over multiple years, where the largest deviations repeatedly occur around the Lunar New Year and shift across the Gregorian calendar, creating event-driven regime changes that standard models struggle to capture. We propose an event-aware hybrid forecasting framework that models extreme behaviors rather than discarding them as outliers. The approach first applies causal rolling wavelet decomposition to split the signal into low-frequency trend and high-frequency detail components while preventing information leakage from the future. The trend component is forecast using an ensemble DLinear branch for stable global prediction. In parallel, a Lunar-aware XGBoost branch learns event-conditioned residual corrections from the detail component and calendar-related cues. A masked fusion mechanism selectively activates the correction branch mainly during event-like periods, reducing the risk of over-correction on normal days. Experiments against strong transformer-based and linear baselines show that the proposed two-stage strategy consistently improves forecasting accuracy, achieving substantial error reduction and better robustness during holiday-driven extremes. The results highlight the value of combining causal time–frequency decomposition with explicit event awareness for practical forecasting under non-Gregorian, calendar-induced shifts.