PREDICTING HOUSEHOLD ENERGY USE FROM SMART METER TIME SERIES: A STACKED XGBOOST–LSTM MODEL

Abstract

Smart meters record household electricity use at a fine temporal resolution, providing the data needed to forecast consumption at the level of the individual dwelling. Accurate short-term forecasts support demand-side management, peak shaving and the integration of distributed renewable generation, yet residential load is noisy, non-stationary and shaped by weather, calendar effects and occupant behaviour, which limits the accuracy of any single learner. This study presents a smart meter-based intelligent model that predicts household appliance energy consumption with a stacked ensemble of heterogeneous learners. The model is organised as a layered architecture covering data publishing, middleware preprocessing, structured storage, model training and prediction. Using 19,735 ten-minute records of appliance energy use, indoor temperature and humidity, and weather from a low-energy house, the readings were enriched with twelve lagged consumption values, rolling statistics, calendar features and outdoor conditions. An eXtreme Gradient Boosting (XGBoost) regressor tuned by grid search, a Long Short-Term Memory (LSTM) network and a Prophet model were trained on the first 80% of the series, and their predictions were combined by a linear meta-learner. On the final 20% of the series, the stacked model achieved a mean absolute error of 15.45 Wh, a root mean squared error (RMSE) of 27.37 Wh, a mean absolute percentage error of 16.11% and a coefficient of determination (R²) of 0.909, compared with R² values of 0.892 for the tuned LSTM, 0.870 for XGBoost and 0.557 for a baseline LSTM; its RMSE was 8.2% lower than that of the tuned LSTM and 16.2% lower than that of XGBoost. The results indicate that combining boosted trees with recurrent networks improves household consumption prediction and can support smart home and smart grid energy management.