Cities need photovoltaic (PV) systems to meet climate-neutral goals, yet dense urban forms and variable weather limit their output. This review synthesizes how machine learning (ML) models capture both static factors (orientation, roof, and façade geometry) and dynamic drivers (irradiance, transient shading, and meteorology) to predict and optimize urban PV performance. Following PRISMA 2020, we screened 111 records and analyzed 61 peer-reviewed studies (2020–2025), eight Horizon-Europe projects, as well as market reports. Deep learning models—mainly artificial and convolutional neural networks—typically reduce the mean absolute error by 10–30% (median ≈ 15%) compared with physical or empirical baselines, while random forests support transparent feature ranking. Short-term irradiance variability and local shading are the dominant dynamic drivers; roof shape and façade tilt lead the static set. Industry evidence aligns with these findings: ML-enabled inverters and module-level power electronics increase the measured annual yields by about 3–15%. A compact meta-analysis shows a pooled correlation of r ≈ 0.966 (R2 ≈ 0.933; 95% CI 0.961–0.970) and a pooled log error ratio of −0.16 (≈15% relative error reduction), with moderate heterogeneity. Key gaps remain, such as limited data from equatorial megacities, sparse techno-economic or life-cycle metrics, and few validations under heavy soiling. We call for open datasets from multiple cities and climates, and for on-device ML (Tiny Machine Learning) with uncertainty reporting to support bankable, city-scale PV deployment.”
Tabatabaei et al. (Tue,) studied this question.