SolarImputer: Conditional diffusion-model based spatio-temporal solar irradiance imputation model

Missing observations are pervasive in distributed photovoltaic monitoring systems and can substantially compromise the reliability of irradiance-driven downstream applications. Existing solar irradiance imputation approaches often prioritize temporal dependency modeling at individual sites, while underutilizing intrinsic spatio-temporal coupling and multivariate structure across geographically distributed stations. To address this limitation, SolarImputer is introduced as a conditional diffusion-based spatio-temporal imputation framework that performs joint reconstruction of solar irradiance sequences over multiple regions. The framework comprises three components. First, physically grounded solar-trajectory information provides an a priori clear-sky pre-fill for missing entries. Second, readily available meteorological variables are leveraged to extract time–frequency–space conditional features that guide the generative process. Third, a conditional diffusion model refines the initialized sequences to produce spatially coherent and physically plausible solar irradiance reconstructions. Experiments on public datasets comprising eight regions from the National Solar Radiation Database and six regions from the SolarCube dataset demonstrate that the proposed method achieves lower reconstruction errors than representative baseline approaches, with relative improvements in reconstruction accuracy ranging from 6.8% to 60% across diverse missing ratios and missingness patterns. Performance gains are more pronounced under high-missingness and long-gap settings, indicating improved robustness to realistic sensor outages.