Weather-guided conditional diffusion model with adaptive group normalization for robust imputation of missing distributed photovoltaic power data

Data integrity is crucial for the reliable monitoring and energy management of distributed photovoltaic (PV) systems. However, the occurrence of missing data is a frequent and inevitable issue due to device failures and communication errors in real-world applications, posing severe challenges to downstream analysis. Although numerous imputation approaches have been developed, most existing methods suffer from insufficient robustness, particularly in extreme scenarios characterized by high missing rates. To this end, this paper proposes a novel weather-guided conditional diffusion model (WGDiff) to achieve high-fidelity and robust data imputation. The proposed WGDiff explicitly integrates meteorological priors as external conditional information into the missing data generation process. This strategic integration effectively bridges the information gap caused by sparse data, enabling the model to maintain superior imputation accuracy and strong robustness even when historical self-dependencies are severely compromised. Furthermore, a specialized adaptive group normalization (AdaGN) mechanism is designed to dynamically modulate the feature distributions of the diffusion network based on environmental contexts. This allows the proposed model to extract complex features from conditional information, specifically capturing the strong stochastic dynamics inherent in meteorological priors, thereby further enhancing the reconstruction fidelity. Extensive experiments are conducted on a real-world dataset from the DKASC database to validate the proposed method. The results demonstrate that WGDiff consistently outperforms all baselines across a wide spectrum of missing rates.