Solar simulation accuracy is challenged by shading and solar reflections within built environments. Thus, identifying these events in observational datasets is essential for targeted validations and model refinements. This study proposes a clustering-based method that couples feature engineering with k-means clustering to group pyranometer measurements based on physical conditions related to shading and cloudiness. Three engineered features are used: (i) binned global tilted irradiance (GTI) that distinguishes permanent shadows from transient shading from clouds, (ii) the difference between measured GTI and clear-sky GTI to capture the impact of cloudiness and shading from surroundings, and (iii) the difference between measured and modelled GTI to reveal solar reflections. Two case studies have been used to demonstrate the robustness of the methodology: a dense urban canyon in Geneva (Switzerland) and a stand-alone building in Trondheim (Norway). These differentiate for the urban context, climates, and solar geometry. Findings demonstrate that the model can effectively separate cloudy from clear-sky conditions and identify shading by surroundings, self-shading, and noshading. Nonetheless, performance is worse in Trondheim than in Geneva due to lower irradiance variability. Additionally, shading conditions with low solar elevations and solar reflections in open urban settings are poorly detected.
