Recursive and Adaptive Electricity Forecasting for Renewable Energy Communities Using LightGBM

As the shift toward low-carbon energy systems accelerates, scalable and accurate forecasting techniques become essential for renewable energy communities to manage demand-side flexibility effectively. This study investigates short-term electricity demand forecasting in a residential renewable energy community in Bucciano, Benevento, Italy, utilizing hourly consumption data from multiple users. The goal was to assess the ability of advanced machine learning models to generate reliable day-ahead forecasts without weather data, with a focus on consumption patterns driven by temporal and historical usage characteristics. The widely used regression method LightGBM was evaluated using a walk-forward validation framework with dynamic retraining triggered by statistical drift. The model used historical consumption patterns, time-based indicators, and short-term trends to capture daily and weekly seasonality. Model performance was evaluated using the mean absolute error, coefficient of determination, and mean absolute percentage error, providing a comprehensive assessment of robustness. LightGBM achieved strong forecasting accuracy, with a mean absolute error of 2.66, a mean absolute percentage error of 6.88%, and a coefficient of determination of 0.9110, despite not using any weather inputs. These results underscore the model's suitability for real-time energy management in residential communities, encouraging further exploration of its practical deployment.