Cleaning Prioritization of Distributed Photovoltaic Systems Using Counterfactual Inference and Calibrated Stacking Ensembles

Fleet-scale cleaning prioritization for distributed rooftop PV systems typically relies on basic rules such as days since last cleaning or performance ratio deviation, neither of which estimates the energy that cleaning would actually recover. This paper presents a counterfactual machine learning framework that estimates site-level recoverable energy gain, using per-cluster stacking ensembles of LSTM, XGBoost, and Random Forest. The framework is validated on 124 rooftop PV sites across industrial estates in West Java, Indonesia, using 15 months of daily SCADA records. The ensemble achieves 2.6 to 3.5% MAPE for daily performance ratio prediction. A key finding is that counterfactual predictions systematically underestimate cleaning gain and must be calibrated against historical cleaning events; 191 events across the fleet correct biases of -1.6% to -4.3%, and all five per-cluster bootstrap confidence intervals exclude zero. After calibration, the framework recovers 87% more energy than days-since-cleaning ranking and 62% more than PR-deviation ranking for a 20-site crew allocation. Calibrated estimates across the full 124-site fleet total 5,744 kWh/day of recoverable energy. The framework uses only routine SCADA and maintenance log data, making it suitable for direct integration into operation and maintenance decision-making workflows.