Residential Energy Consumption Forecasting in Electric Utilities: An Approach Based on Random Forests and Time Series

Forecasting the monthly electricity consumption of residential users is a critical task for improving energy planning, demand management, and the efficient integration of renewable energy sources into the electrical system. This study predicts the consumption of a single residential user based on historical data, including monthly consumption and average temperature records from November 2018 to December 2024. Five forecasting approaches are compared: moving averages, ARIMA, standard Random Forest, Random Forest with lag variables, and Random Forest with hyperparameter optimization using RandomizedSearchCV. The models performance is evaluated using MAE, MSE, and RMSE metrics over the last 12 months of the analyzed period, with 95% confidence intervals calculated via bootstrapping for both the validation phase and the estimation for January 2025. The results show that Random Forest models with lag variables and hyperparameter optimization outperform traditional methods such as moving averages and ARIMA in terms of accuracy. Additionally, the use of confidence intervals provides a more robust assessment of prediction reliability. It is concluded that the combined use of machine learning techniques, selection of relevant historical variables, and uncertainty quantification methods offers an effective tool for anticipating residential electricity consumption behavior. This approach can be valuable for electric utilities and policymakers seeking datadriven, reliable, and reproducible decisions.