Data-driven approach of atrium skylights-integrated photovoltaic systems design based on multimodal deep learning: Considering different regions in China

Atriums in public buildings significantly enhance daylighting but can increase heating and cooling loads in extreme climates. Skylight-integrated photovoltaic (SIPV) systems efficiently reduce carbon emissions but may compromise daylight quality and visual comfort due to excessive PV coverage. Therefore, this study develops a data-driven approach that integrates simulation, prediction, and optimization on the Grasshopper-Python platform. Image features are extracted using ResNet-50, ViT-32, and Inception-V4; numerical features are processed via MLP and TabTransformer. Feature fusion to form a multimodal neural network model, which serves as a surrogate model within an optimization framework. Using 7 cities in China as case studies, the approach optimizes atrium form, skylight, PV system, and fenestration factors to minimize carbon emissions and energy return on investment (EROI), while improving daylighting. The results show that: (1) Carbon emission reduction rate (CER) exceeds 89 %, Daylight factor (DF) increases by 1.2 %–3.4 %, Daylight glare probability (DGP) remains below 0.28, and Energy return on investment (EROI) reaches 9.37; (2) atrium width, skylight-to-roof ratio (SRR), PV coverage, and horizontal panel are key factors; (3) the ViT-32+MLP model significantly improves prediction accuracy and efficiency. This research offers a practical method for SIPV design, supporting regional adaptation and scalable application in public building atriums.