Hyperparameter tuning and generalization of LSTM networks for residential energy prediction

Accurately predicting household electricity consumption is critical for optimizing energy management systems, reducing costs, and enhancing grid stability. This study proposes a Long Short-Term Memory (LSTM) based framework to address the challenges of modeling nonlinear and temporal dependencies in energy usage data. The research systematically examines the effects of hyperparameter configurations, including LSTM units, hidden layer neurons, batch sizes, and epochs to identify optimal settings for improved model performance. Additionally, the study investigates the generalization capability of the LSTM model across appliances with differing energy usage characteristics. The results demonstrate that the proposed model achieves high accuracy (87.8%) and robust performance under optimal hyperparameters (F1-score 0.923). Moreover, the generalization experiments reveal the model's limitations in distinguishing between regular electricity use and standby power consumption across diverse appliances, emphasizing further improvements. The study concludes that while LSTM models are adequate for household electricity prediction, their generalization requires careful tuning and targeted refinements. These findings provide valuable insights into the development of scalable and adaptive energy management systems, contributing to optimization strategies for sustainable energy use in residential contexts.