Machine learning-driven optimization of robotic cleaning schedules for photovoltaic panels: a cost-benefit analysis

Solar power plants in arid and semi-arid regions suffer substantial performance losses due to dust accumulation on photovoltaic modules, making the design of an effective and economically sound cleaning schedule a key operational challenge. This study proposes an integrated ML-assisted decision-support framework that combines five years of meteorological and dust observations from a utility-scale photovoltaic plant in Semnan, Iran, with two machine learning models, Extreme Gradient Boosting and Random Forest, to generate a season-adaptive 365-day robotic cleaning schedule. Model outputs are translated into actionable decisions by enforcing minimum inter-cleaning intervals and benchmarking the learned policy against alternative scenarios. The constrained ML-assisted schedules recommend 32 and 33 robotic cleaning events for the Extreme Gradient Boosting and Random Forest models, respectively, leading to an estimated annual gain of about 411.6 MWh in electricity generation, corresponding to an efficiency improvement of approximately 2.1% in the XGBoost-based scenario, and a reduction of total cleaning costs by roughly 51%. These base-case results indicate that the proposed framework can provide a practical decision-support strategy for mitigating soiling-induced losses and supporting intelligent maintenance in dusty utility-scale PV plants, while the economic outcomes remain dependent on site-specific tariff, cleaning-cost, and robotic-system assumptions.