Evaluation of the impact of machine learning on the prediction of residential energy consumption

Abstract

The objective of this research was to compare the performance of machine learning models and traditional statistical methods for the prediction of residential energy consumption, using a dataset with relevant variables such as consumption, temperature, time of day, type of housing, and energy usage habits. A quantitative and comparative methodology was applied, involving data preprocessing, variable encoding, and normalization, as well as division into training and testing sets. The random forest, support vector machine (SVM), deep neural network (MLP), and linear regression models were trained and evaluated using standard metrics such as mean absolute error (MAE), root mean squared error (RMSE), and R² on test and cross-validation sets. Results show that SVM and linear regression achieved better accuracy and generalization capability, while random forest and the deep neural network exhibited lower explanatory power, reflected in negative R² values. Using the trained models, a projection of residential energy consumption for the 2026–2030 period was performed, revealing a generally increasing trend across all models, although with differences in the magnitude of the predictions. In conclusion, under the current conditions, traditional models demonstrate greater robustness, highlighting the need to tailor algorithm selection to the data context. These projections provide a valuable tool for future energy planning.