A Two-Stage Neural Network for Day-Ahead PV Forecasting Using Environmental Data from the LAMBDA Microgrid in Rome

Accurate day-ahead forecasting of photovoltaic (PV) generation is essential for the reliable operation of smart grids and energy management systems (EMS). This paper proposes a lightweight, two-stage neural network model for 24-hour ahead PV prediction at 15-minute resolution. The framework integrates a daylight classifier, implemented using a bagged ensemble of decision trees, with a Bayesian-regularized feedforward neural network (FFNN) trained exclusively on daylight samples. The model is trained and validated using real-world PV and environmental data from the LAMBDA Microgrid Laboratory in Rome, Italy. Experimental results demonstrate an RMSE of 0.0755 and a daylight MAPE of 19.94%, along with 100% accuracy in detecting zero-generation periods. The architecture ensures physically consistent forecasts and supports real-time deployment in EMS. Compared to standard baselines such as SVR and single-stage FFNN, the proposed model achieves superior daylight accuracy and robustness. It offers a practical foundation for real-time EMS deployment and future extensions involving probabilistic forecasting and hybrid deep learning architectures.