Deep Learning-Based Energy Consumption Optimization in Smart Home Systems An LSTM Approach with SHAP Interpretability

Abstract

Accurate and easy-to-understand energy consumption forecasting is a prerequisite for effective demand-side management in smart home systems. This study proposes an integrated framework that combines a stacked Long Short-Term Memory (LSTM) network with SHapley Additive exPlanations (SHAP) for hourly residential energy forecasting and the extraction of optimization insights. The model was evaluated on the UCI Individual Household Electric Power Consumption (IHEPC) dataset—34,589 hourly observations resampled from over two million sensor records per minute—and tested against Stacked GRU, Vanilla RNN, and ARIMA(2,1,2). All deep learning models significantly outperformed ARIMA (R² = −0.1258, MAPE = 126.92%). The proposed LSTM model achieved MAE = 0.3208 kW, RMSE = 0.4568 kW, and R² = 0.5749, while Stacked GRU recorded the best aggregate metrics (MAE = 0.3090 kW, R² = 0.5801). SHAP analysis identified Global Intensity as the dominant predictor (mean |SHAP| = 0.00350), based on the electrophysical relationship P = V × I, followed by the 3-hour Rolling Mean and seasonal encoding. Further temporal attribution analysis revealed that predictive value is concentrated in the most recent input time steps, confirming that real-time current measurements are the most valuable input for HEMS implementation.