Abstract
Physics-based simulation tools are reliable for photovoltaic system design and annual-yield assessment, but they do not resolve meteorological changes that govern operational forecasting in high-altitude regions. This study develops a day-ahead multi-step forecasting method for residential building-integrated photovoltaic (BIPV) systems on the Qinghai–Tibet Plateau. A physics-enhanced plateau-meteorological-regime-gated mixture of experts (PGMR-MoE) is evaluated for separate 1, 6, and 24 h predictions of hourly grid-connected energy (EGrid) and a normalized horizontal-irradiance index (NHEI) in residential BIPV simulations. The model combines a 72 h lookback span, engineered weather interactions, five experts, validation-based fusion, and a low-irradiance output rule. All accuracy tests use target-time PVsyst weather as a perfect-weather-input proxy. Under a month-stratified split, 24 h EGrid MAE, RMSE, and R2 are 4.3467 kWh, 10.0039 kWh, and 0.9943; NHEI values are 0.0345, 0.0851, and 0.9806. ExtraTrees gives a lower EGrid RMSE of 7.7108 kWh. In a matched full-configuration comparison at the 24 h horizon, EGrid RMSE increases from 9.8021 kWh under the month-stratified split to 40.7179 kWh under the blocked split with a 96 h purge, while R2 decreases from 0.9945 to 0.8928. These results indicate that prediction performance is sensitive to temporal data partitioning, particularly under temporally blocked evaluation. Direct NHEI prediction modestly improves on derivation from predicted EGrid. Cross-city evaluation within the PVsyst simulation framework yields EGrid R2 = 0.9700–0.9831 and NHEI R2 = 0.9484–0.9653 across fifteen cities excluded from fitting, using local historical outputs and target-time simulated weather. The method shows potential to support day-ahead scheduling, energy management, and operation of BIPV systems across diverse geographical regions.