Hybrid Metaheuristic-Machine Learning Framework For Optimizing Solar Pv Layouts On Irregular Urban Rooftops

Designing solar photovoltaic (PV) layouts on irregular urban rooftops is a challenging combinatorial problem, complicated by shading, structural obstacles, and irradiance variability. Traditional metaheuristic techniques such as Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Simulated Annealing (SA) are often computationally intensive for high-dimensional layouts, while purely machine learning (ML)-based methods struggle to explore the vast solution space effectively. This work proposes a hybrid optimization framework that integrates GA, PSO, and SA with a Random Forest surrogate model, which approximates the irradiance-adjusted power generation landscape and guides efficient global exploration, with final solutions validated against the actual fitness function. Applied to rooftop datasets from Singapore, Rio de Janeiro, Nairobi, and Surakarta, the framework achieved more than 90% of the realistic maximum power output while significantly reducing computational demand. The surrogate model maintained R² values above 0.8, ensuring dependable estimations, and outperformed standalone algorithms and pure ML approaches, confirming the advantages of the hybrid strategy.