Comparison of Retraining Schedules in Ensemble Models for Enhanced Day-Ahead Electricity Net-Demand Predictions

This paper presents a comprehensive study comparing the performance of tree ensemble models for day-ahead electricity net-demand predictions, focusing on the impact of different retraining schedules. The models are trained with Least-Squares Boosting (LSBoost) and evaluated based on their predictive accuracy. The study explores retraining schedules at various intervals, namely daily, weekly, fortnightly, and four-weekly schedules. The results reveal that retraining significantly influences predictive performance, with the Daily model demonstrating the highest overall accuracy with an R-squared of 0.833 and an RMSE of 0.664 MW. Surprisingly, the Four-Weekly model performs nearly as well with an R-squared of 0.832 and an RMSE of 0.668 MW, suggesting potential efficiency gains. The Weekly and Fortnightly models exhibit lower performance, indicating that certain demand patterns may require longer or shorter adaptation intervals.