Physics-Informed Data-Driven Models for Electricity Load Forecasting in Complex Energy Systems

The ongoing energy transition, driven by the increasing electrification of end-use sectors, the widespread deployment of renewable energy sources, and the decentralization of power generation, is modifying the operation of modern energy systems. In this context, accurate day-ahead electricity load forecasting has become essential for ensuring system reliability, operational efficiency, and effective energy management. While data-driven machine learning techniques have demonstrated predictive capabilities, their performance depends on the information provided to the models and may deteriorate under atypical demand conditions. This work proposes a general knowledge-informed data-driven framework to investigate how relevant system information can be progressively incorporated into conventional forecasting models. The framework is illustrated through a case study of an Italian power system using 15-min load data for 2022–2025 and two complementary learning approaches, XGBoost and LSTM. Results show that enriching data-driven models with temperature and spatial information improves forecasting accuracy, with XGBoost achieving the best performance and reducing the MAE from 968 to 792 MW. XGBoost outperformed LSTM for the configurations considered.