Decision-support forecasting and predictive maintenance for transparent photovoltaic façades in hot-arid, dust-prone climates
Transparent photovoltaic (PV) façades operated under combined heat and dust stress in hot-arid regions, which increased generation variability and complicated maintenance planning. This study designed and assessed an operational decision-support workflow that combined short-term energy forecasting with predictive maintenance classification for transparent PV façades deployed across four Saudi sites (NEOM, Riyadh, Red Sea, and AlUla). A gated recurrent unit (GRU) model forecasted normalized AC power at 15-, 30-, and 60-minute horizons using engineered inputs that represented irradiance drivers, heat-stress indicators, and surface-condition proxies. A Random Forest classifier estimated anomaly probability from operational features and maintenance-aligned labels, and probability calibration, threshold governance, and persistence logic controlled false alarms. Forecast and anomaly outputs were then translated into explicit maintenance triggers (Watch, Inspect, Clean, Urgent) and a dispatch-priority score based on expected energy loss, anomaly likelihood, and time since the last cleaning. The multi-site evaluation characterized how coastal adhesion regimes and inland event-driven dust exposure influenced forecast reliability, alert burden, and lead time to actionable maintenance. The workflow produced auditable decision rules that supported planning and dispatch across heterogeneous hot-arid operating conditions and reduced reliance on ad hoc interventions.