ПІДВИЩЕННЯ ЕНЕРГОЕФЕКТИВНОСТІ ФАСАДНОЇ СОНЯЧНОЇ ЕЛЕКТРОСТАНЦІЇ ЗІ ЗАСТОСУВАННЯМ ШТУЧНОГО ІНТЕЛЕКТУ

Optimizing solar photovoltaic systems in complex urban environments requires intelligent and adaptive control strategies to account for the dynamic operational factors. This study presents a data-driven simulation and optimization framework for the “River Park, Facade” solar Photovoltaic installation using PyPSA for system modeling and multi- agent Deep Q-Learning (DQN) for control optimization. The system under study features a 30. 2 kW DC-rated array and a 30 kW AC inverter, with a measured annual output of 23. 38 MWh. The performance ratio (PR) of 84.3% based on meteorological inputs from Meteonorm are used for the analysis. To enhance operational efficiency, we deployed a multi-agent reinforcement learning (MARL) framework that exploits three cooperative agents: a cleaning agent to mitigate soiling losses, a tilt agent to maximize irradiance capture, and an inverter agent to regulate conversion efficiency. Each agent was trained within a PyPSA– OpenAI Gym environment using shaped rewards aligned with specific loss categories. The MARL system was validated through comparative overlay plots of the original system versus the RL-optimized version. Results demonstrate an increased monthly grid output in nearly all months and a slight but consistent improvement in PR. Additionally, higher cumulative energy production throughout the year, and a notable gain in total annual energy yield was observed in case of RL-enhanced regime. These findings confirm the potential of modular, reward-shaped MARL for real-time optimization of facade-integrated solar Photovoltaics systems for smart building energy systems in urban settings.