Improving Residential Load Forecasting: A Probabilistic Tree-based Approach for Multi-Step Ahead Prediction

Accurate short-term residential load forecasting is crucial for efficient energy planning and management. This study introduces a tree-based approach for point forecasting and probabilistic machine-learning approaches to predict short-term electrical loads, leveraging feature selection based on trends and periodic patterns in historical data. The method can predict electrical load patterns of any specific region based on the historical load profile and its temporal features and weather forecasts, such as wind speed, air pressure, temperature, humidity, and precipitation. The algorithm is flexible and can be used for different prediction horizons, from a few hours up to several days ahead using a rolling window approach. Four variants of the tree-based approach are tested and validated here, two for point forecasting and two for probabilistic forecasting based on selected evaluation metrics. Comparative simulation results demonstrate that the proposed tree-based approach delivers superior predictive accuracy, achieving notable improvements in electrical load forecasts for residential applications.