Leveraging synthetic data to empower AI models to predict photovoltaic energy production to aid in the decarbonization of buildings

This research explores the use of synthetic data to enhance the accuracy of machine learning models predicting the energy production of photovoltaic (PV) systems integrated into buildings. We address the challenge of data scarcity in real-world scenarios by generating a large and diverse dataset using BIMSolar, encompassing a wide range of building types, PV panel models, and installation locations. The synthetic data approach allows us to directly estimate energy production for new scenarios without requiring historical data, enabling a shift from forecasting to prediction. We evaluate the performance of various machine learning models, including Random Forest, Gradient Boosting, and XGBoost, using metrics such as MSE, R², MAE, and MAPE. Notably, XGBoost and Decision Tree models emerge as top performers, demonstrating high accuracy and efficiency while consuming minimal energy during training. This research presents a promising approach to support building decarbonization efforts by providing reliable estimates of PV energy production, ultimately facilitating a transition towards a sustainable energy future.