Simulation-Based Photovoltaic Production Data for Training Data-Driven PV Power Forecasting Models

Accurate day-ahead photovoltaic (PV) power forecasting is essential for grid operation and energy market participation. Nevertheless, PV plants frequently do not have enough historical production data to develop accurate forecasting models. In situations where real measurements are scarce, this work investigates whether physics-based simulated data can support data-driven forecasting models. We propose a methodology that leverages a subset of numerical weather prediction (NWP) variables, provided by the ICON-EU model that predicts dayahead DC production for a target PV plant. Specifically, a 3D convolutional neural network (3D-CNN) is trained to map spatial-temporal NWP forecasts to day-ahead PV power production. To overcome the issue of data scarcity, we use a physics-based model implemented with pvlib to simulate the DC power production for the target PV plant. We compare two experimental settings to assess the use of simulated data as an alternative to scarce real measurements, one using simulated production data and another using the limited available real measurements. Results obtained on a real-world demo site show that the model trained on simulated data achieves a lower root mean squared error (RMSE) than the model trained solely on measured production, highlighting the potential of combining physical modeling and machine learning to improve PV forecasting in data-scarce scenarios such as newly deployed solar plants.