AI-empowered forecasting of electricity demand for optimal energy management
TL;DR
The results show that an AI-based forecasting framework can provide reliable, decision-ready demand estimates for digital energy management in buildings and public institutions.
Abstract
The digitalization of energy systems is increasing the need for accurate and operationally useful short-term electricity demand forecasts, especially in buildings, where flexible loads can support peak reduction, demand response, and more efficient operation. Improved forecasting can significantly reduce energy costs by helping institutions avoid peak-demand penalties and make better use of flexible resources. This paper presents a lightweight and reproducible machine-learning pipeline for day-ahead building electricity forecasting with calibrated uncertainty. The proposed approach uses gradient-boosted tree models with calendar, weather, and recent-load features to predict hourly electricity demand over the next 24 hours. To make the forecasts more useful in practice, point predictions are complemented with prediction intervals obtained through quantile regression and conformal calibration. We evaluate the method on an open dataset, BDG2. The results show that an AI-based forecasting framework can provide reliable, decision-ready demand estimates for digital energy management in buildings and public institutions.