Context-Aware Smart-Home Energy Forecasting Using an Energy-Derived Presence Proxy: A Hybrid Deep Learning and Machine Learning Approach

Abstract

The demand for household electricity continues to rise, increasing both energy costs and greenhouse-gas emissions. An accurate forecast of short-horizon consumption will contribute significantly to sustain such efforts through facilitating demand response, load shifting, and general smart home energy management. This paper presents a hybrid model that integrates (i) appliance-level usage and weather features with (ii) an energy-derived occupancy proxy (presence rate), aimed at improved next-step household electricity consumption prediction. The proposed framework first estimates a presence-proxy signal from the appliance activity patterns and then utilizes this signal as an added contextual feature of consumption forecasting. We evaluate tree-based ensemble learners (Random Forest, XGBoost, LightGBM) and deep learning models (CNN and LSTM) using a time-aware train/test split. Results show that incorporating the presence-proxy enhances forecasting accuracy for the best-performing tree-based model, while LSTM maintains robustness due to its temporal modeling capacity. Beyond predictive accuracy, we discuss how the improved forecasts can be operationalized for peak reduction and energy-efficiency interventions, and we highlight validity considerations when inferring occupancy-related signals from smart meter data.