Why theoretical photovoltaic suitability does not fully predict deployment: interpretable deep learning of urban-scale siting across Chinese cities

Macro-scale photovoltaic (PV) siting studies usually estimate where PV should be built, but rarely test whether theoretical suitability surfaces correspond to where PV is actually deployed. This study addresses that gap by first comparing reconstructed PV construction-potential surfaces with observed empirical PV installation footprints across nine Chinese cities spanning three solar-resource and deployment contexts, and then using interpretable deep learning to diagnose the spatial factors and interactions that organize realized deployment. Twenty-one aligned geographic, environmental, infrastructural, and socio-economic factors were used to train four models under a leave-one-city-out (LOCO) design. The best-performing architecture was then retrained on the pooled nine-city dataset for SHAP and Average Causal Effect (ACE)-style interpretation. The LOCO results show that cross-city transfer is difficult: all models produced low absolute segmentation scores, while ResNet achieved the strongest relative performance, with the highest mean IoU, F1-score, and precision. The interpretive results show that theoretical potential and observed deployment are only partially aligned, and that the mismatch varies substantially across cities. Power System (F15) is a recurrent driver across city classes, but its role is conditional. Class I and II cities show stronger grid-resource-climate dependence, especially through power-system connectivity, GHI, terrain, and temperature-related factors. Class III cities retain grid sensitivity but distribute importance more broadly across POI-based land-use proxies, population, GDP, water bodies, road access, and local demand-related signals. ACE further reveals non-additive interactions among grid, resource, terrain, population, and economic variables. Because the footprint data likely underrepresent small residential rooftop PV, the findings are interpreted primarily for large-area and land-intensive PV deployment. Overall, the framework supports differentiated PV planning and policy evaluation.