Evaluation of Ensemble and Hybrid Models for Predicting Household Energy Consumption: A Comparative Study of Machine Learning Approaches

Abstract

Accurately predicting household energy consumption is critical for efficient energy management, particularly as global energy demands rise. This study explores the predictive performance of various machine learning models, including linear regression, Ridge regression, Lasso regression, Random Forest, Gradient Boosting, XGBoost, CatBoost, and a hybrid model combining Long Short-Term Memory (LSTM) networks with Random Forest regression. The models were evaluated on a dataset consisting of minute-level energy readings over a 350-day period. Key performance metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and the coefficient of determination (𝑅²) were used to assess model accuracy. The results demonstrate that ensemble models, particularly Random Forest and CatBoost, outperformed traditional regression models in terms of error minimization. CatBoost achieved the lowest MSE among all models, highlighting its effectiveness in handling non-linearities and categorical data. However, none of the models achieved a positive (𝑅²) score, indicating their limitations in fully explaining the variance within the dataset. The hybrid LSTM + Random Forest model, despite its expected strength in capturing temporal dependencies, performed worse than simpler models, suggesting issues with feature extraction and model integration.These findings suggest that while ensemble methods are well-suited for energy consumption prediction, more advanced modeling techniques or enhanced feature engineering are needed to improve performance. Future research could explore deeper neural networks or time-series models such as ARIMA to better capture the temporal patterns in household energy consumption.