Deep Learning for Residential Electrical Energy Consumption Forecasting: A Hybrid Framework with Multiscale Temporal Analysis and Weather Integration

This paper presents an evaluation of the use of deep learning architectures for forecasting electrical energy consumption in residential environments. The main contribution of this study lies in the development and assessment of a hybrid forecasting framework that integrates multiscale temporal analysis and weather data, enabling evaluation of predictive performance across different temporal granularities, forecast horizons, and aggregation levels. Single and hybrid models were compared, trained with high-resolution data from a single residence, both considering only endogenous variables and including exogenous variables (weather data). The results showed that, among all models tested in this study, the hybrid LSTM + GRU model achieved the highest predictive performance, with R2 values of 94.62% using energy data and 95.25% when weather variables were included. Intermediary granularities, particularly the 6 steps, offered the best balance between temporal detail and predictive robustness for the tests performed. Furthermore, short-time windows aggregation (1 to 5 min) showed better accuracy, while the inclusion of weather data in scenarios with larger aggregation windows and longer horizons provided additional gains. The results reinforce the potential of hybrid deep learning models as effective tools for forecasting residential electricity consumption, with possible practical applications in energy management, automation, and integration of distributed energy resources.