Dust-resilient optimization of building-integrated PV systems using enhanced deep reinforcement learning: A case study in Saudi Arabia

Building-integrated PV (BIPV) deployed in desert climates is subjected to persistent derating from heat, soiling, and sand-storm events that is seldom reflected in optimization. An enhanced deep reinforcement learning with climatic uncertainty integration (EDRL-CUI) framework is presented for the joint sizing and operation of a hybrid BIPV–wind–battery system serving a commercial building in Riyadh, Saudi Arabia. At the upper level, capacities are selected by NSGA-III; at the lower level, operational policies are learned by DRL based on high-resolution weather and dust inputs. In contrast to prior DRL-based EMS studies, dust/soiling and sand-storm dynamics are explicitly encoded as first-class uncertainties in the state and reward, resilience metrics (CLSR, SRT, MDA) are jointly optimized alongside cost, self-sufficiency, CO2 reduction, and grid-friendliness, and sizing and operation are co-optimized through a hierarchical NSGA-III + DRL scheme tailored to arid BIPV. The approach is benchmarked against model predictive control, rule-based EMS, and classical DRL, and is evaluated on a 15-min, 2021–2023 climate/dust dataset under a leave-one-year-out protocol. Relative to MPC on a fixed sizing, ∼18% lower operating cost, ∼17% improvement in grid-interaction index, +∼9 percentage-point storm CLSR, and ∼9 h faster recovery are observed; with joint sizing–operation, self-sufficiency reaches ∼39–47% and CO2 reductions ∼42–49%. During sand-storm events, resilient performance is maintained via adaptive cleaning and battery dispatch. The case-study scope and simulation-based validation are acknowledged as limitations, and implications for BIPV deployment in arid regions are discussed.