Gradient Boosting Algorithms for Day-ahead Residential Load Forecasting of Individual Household

The accomplishment of electrification targets and the successful integration of renewable energy sources into the residential grids, conceives the urgency for implementing platforms for the optimal dispatch plan of the energy. This study investigates the utilization of various Gradient Boosting Decision Trees (GBDT) models for the efficient Day-ahead Load Forecasting (DALF) of individual households. In this work, GBDT algorithms, such as eXtreme Gradient Boosting (XG- Boost), Categorical Boosting (CatBoost), Light Gradient Boosting Model (LightGBM) are compared and evaluated based on the accuracy of the day-ahead prediction with one hour granularity. The aforementioned models are trained on historical household energy consumption data, incorporating temporal and seasonal features to capture complex consumption patterns. Furthermore, we propose a novel hybrid algorithm that combines the Extreme Learning Machine (ELM) as a weak learner. The preliminary forecasts generated by the ELM are incorporated as an additional input feature into the XGBoost model, forming a hybrid architecture designed to enhance predictive performance. Experimental results demonstrate that the proposed hybrid model significantly outperforms the standalone GBDT algorithms as well as a persistence baseline.