Performance analysis of vertical patterned glass BIPV in zero energy buildings using explainable AI

Building integrated photovoltaic (BIPV) systems offer a key alternative for urban spatial constraints, yet detailed analyses of power generation across various module concepts are still required. This study conducts an in depth analysis combining data driven approaches and explainable artificial intelligence (XAI) on a one year field dataset (20,700 observations) from 110 full sized, vertically mounted patterned glass BIPV modules. First, data analysis reveals that while voltage (Vmp) decreased with rising temperatures, current (Imp) correlated strongly with plane of array (POA) irradiance. High summer solar altitudes increased incidence angles, reducing effective irradiance and Imp. Consequently, daily energy yield peaked in March (1.68 kWh/kWp) and minimized in July (0.89 kWh/kWp). Second, an XGBoost predictive model achieved high forecasting accuracy (R 2 = 0.9534, nRMSE = 5.661%). SHAP analyses confirmed that POA irradiance at time t was the most dominant factor for predictions, whereas global horizontal irradiance had the greatest overall influence when evaluating all input time steps from a long-term perspective. Temperature and humidity showed relatively low importance. Furthermore, temporal feature analysis revealed that while the most recent data was naturally the most critical, the most distant data emerged as the second most important factor. Finally, cross verifying field data with XAI refutes hypotheses attributing power degradation primarily to heat accumulation. Generation fluctuations are governed by reduced effective irradiance rather than temperature elevation, suggesting new perspectives worth exploring for BIPV design. Ultimately, this research overcomes confirmation bias, establishing a novel paradigm for objective system interpretation.