Multi‐Scale Transformer Framework for Explainable Medium‐Term Household Energy Forecasting

Abstract

ABSTRACT Medium‐term electricity demand forecasts (days–weeks ahead) are essential for scheduling, maintenance, and tariff or hedging decisions, yet remain challenging due to multi‐scale seasonality and weather‐driven non‐stationarity. We propose a Multi‐Scale Transformer (MSTr) that fuses diurnal, weekly, and seasonal context via scale‐specific pooling and learned softmax gating. The model is trained in a leak‐safe, direct‐ protocol for horizons using only information available at decision time. On hourly household data from Northeast Mexico, MSTr consistently outperforms strong baselines (LightGBM, LSTM, and Seasonal/Naive) across RMSE, MAE, WAPE, sMAPE, and , with the largest gains at 168–336 h where weekly and seasonal signals dominate. Explainability analyses (SHAP, partial dependence, temperature‐load sensitivity, and multi‐scale gating diagnostics) indicate that MSTr captures meteorological and calendar effects more faithfully than baselines. An ablation study confirms the utility of positional encodings, multi‐scale pooling, and learned gates, while a block‐bootstrap evaluation and Diebold–Mariano tests show that MSTr's improvements are statistically significant at key horizons. The approach remains computationally practical, is straightforward to integrate into existing forecasting pipelines, and supports transparent reporting through gate weights, seasonal slices, and feature‐level explanations.