Operational data of photovoltaic (PV) systems are often affected by equipment failure, communication interruption, and other objective factors, leading to widespread data missing issues and degraded data quality. Effective data restoration is an important foundation for achieving high-precision PV power forecasting and optimal power plant operation. Traditional data imputation methods struggle to effectively capture the complex spatiotemporal correlations in distributed PV data and exhibit limited adaptability to special scenarios such as missing meteorological data, resulting in unsatisfactory imputation accuracy. Therefore, a spatial-attention temporal-convolution network (SATCN) is proposed. A one-dimensional convolutional neural network is used to learn temporal dependencies, while a spatial attention network inspired by the self-attention mechanism of the Transformer is constructed as a spatial aggregator. By leveraging limited observable data, the proposed model fully exploits spatiotemporal dependencies to achieve high-quality restoration of data from multiple distributed PV clusters. In addition, a joint optimization training strategy is adopted, incorporating both observed value reconstruction and artificially masked data imputation tasks, thereby preventing the model from focusing solely on observed data while neglecting missing data imputation. Experimental results show that the proposed method has good imputation performance under various missing rates and achieves high-precision restoration of missing data without supervision.