Physics-informed Bayesian telemetry for dust-aware urban BIPV fleets: Calibrated forecasting, resource-constrained operations and maintenance, and event-triggered cleaning decisions

Rooftop BIPV in dusty cities incur coupled optical and thermal losses. Common practice still uses static soiling derates and nominal temperature corrections. These choices ignore aerosol dynamics, do not quantify uncertainty, and offer little guidance on when to clean under water constraints. This study develops and validates a telemetry-only, physics-informed, Bayesian framework. It jointly infers time-varying soiling and thermal states from standard AC/DC, irradiance, and temperature telemetry. It then converts posterior distributions into event-triggered cleaning decisions that obey district-level water allocations in Riyadh. The estimator couples a dynamic heat-balance PINN with a single-diode electrical surrogate and a covariate-driven soiling state informed by aerosol optical depth and roadside particulate matter. Randomized cleaning at a subset of rooftops strengthens identifiability. Across thirty rooftops (1-min telemetry, 12 months), hourly AC RMSE decreased by 0.022–0.027 p.u. versus static derates and by 0.015–0.026 p.u. versus SRR + empirical. Cell-temperature RMSE decreased by 1.3–1.8 °C. The nominal 80 % and 90 % prediction-interval coverages were approximately 0.80 and 0.89. The policy executed 242 rather than 360 washes, reduced water use by 94.4 m3, increased energy by 398 MWh, raised revenue by 71,640 SAR, improved PR by 1.9 percentage points, and lowered LCOE by 14 SAR/MWh while respecting district allocations. Physics-guided learning combined with aerosol-aware Bayesian inference thus provides calibrated forecasts and resource-efficient O&M decisions for urban BIPV fleets.