Explainable Machine Learning for Household Electricity Demand Forecasting Using Hybrid Residual Modeling

Abstract

—This study proposes an interpretable hybrid residual learning framework for household electricity demand prediction under limited data conditions. The framework integrates statistical forecasting models with machine learning residual learners and SHapley Additive exPlanations (SHAP)-based explainability to improve predictive accuracy while preserving transparency. Using 1,000 hourly household electricity consumption observations collected over approximately 41.7 consecutive days, the proposed model was compared with Linear Regression, Autoregressive Integrated Moving Average (ARIMA), Random Forest, Support Vector Machine, and Deep Neural Network baselines under identical chronological train-validation-test conditions. Results show that the hybrid framework achieved the best predictive performance, with Mean Absolute Error (MAE) = 0.141, Root Mean Squared Error (RMSE) = 0.198, Mean Absolute Percentage Error (MAPE) = 7.5%, and Coefficient of Determination (R²) = 0.90, outperforming both statistical and standalone machine learning models. The results are interpreted within a short-term limited-data forecasting scenario. Although the proposed framework showed the best performance on the independent chronological test set, robustness under progressive training-data reduction is considered a direction for future validation because the available dataset contains only 1,000 hourly observations, while statistical comparison was performed using the Diebold-Mariano test. SHAP explanations identified temperature, hour-of-day patterns, and household occupancy as the most influential demand drivers, providing physically meaningful insights for residential energy planning. The findings suggest that interpretable hybrid residual models represent a reliable alternative for short-term household electricity forecasting in data-scarce urban environments