Multi-performance prediction and optimization for building-integrated photovoltaics facades with passive design via explainable machine learning

Building-integrated photovoltaic (BIPV) facades with passive design represent a low-carbon, sustainable architectural strategy for addressing climate change and energy challenges. Given that early-stage design decisions significantly impact project outcomes, this study focused on developing rapid assessment methods for three key performance aspects: daylight availability, solar energy generation, and building energy efficiency. To achieve this, we established Shanghai-specific dataset through building performance simulations and label classification. Using this dataset, we developed predictive models for four critical metrics: spatial daylight autonomy (sDA), solar radiation, EUI_heat, EUI_cool. By comparing Random Forest and XGBoost algorithms, we found that both achieved strong performance (F1 scores: 0.856 for sDA, 0.808 for solar radiation, 0.878 for EUI_cool, and 0.924 for EUI_heat). Notably, SHAP-based explainability analysis not only validated the models’ reliability by aligning with correlation results but also revealed the relative importance of different design parameters. Furthermore, when tested in other cities with similar climates, the models maintained high accuracy, demonstrating their practical value for regional applications. The proposed method reduces the computational time from 106–239 h to 60.6 h. After optimization, the optimal solution can achieve 25 % to 48 % of cooling and heating energy supplied by photovoltaic power generation. This research provides architects with a predictive tool to assess multiple performance metrics of BIPV façade with passive design, supporting sustainable design decisions.