Implementation of an AI-driven dynamic control system to optimize excess photovoltaic energy management in grid-connected sustainable BIPV

This study proposes an integrated approach for optimizing grid-connected
photovoltaic (PV) systems through AI-based forecasting and a Dynamic Automatic Control
System (DACS). Using Convolutional Neural Networks (CNN), Long Short-Term Memory
networks (LSTM), and a hybrid CNN-LSTM model, we predict PV production and building
energy demand. The CNN-LSTM model achieved the best performance for PV forecasting
(MAE = 0.0423), while LSTM excelled in energy demand prediction (MAE = 0.0130). A
disaggregated analysis of key building surfaces (South Facade, South Roof, East Facade,
and North Roof) revealed significant variations in energy contributions. MATLAB/Simulink
simulations based on one day of demand indicate that the DACS attains a 70% PV utilization
rate, with the South Roof and South Facade reaching 69% and 42% utilization, respectively.
These results demonstrate the potential of AI-driven energy management systems to optimize
renewable energy use and support decarbonization by reducing grid dependency. This study
integrates, for the first time, ultra-short-term (5-minute) forecasting and real-time dynamic
control, offering continuous PV–battery–grid operation and establishing a new standard in
energy efficiency.