As the energy evolves, the types of power sources linked to the distribution grid are growing increasingly varied, with new energy sources like distributed photovoltaic (PV) developing rapidly. Nevertheless, the fluctuating and unpredictable nature of distributed PV systems presents a challenge to the characteristics of the power grid. Thus, analyzing and predicting the capacity of distributed PV access is of great significance. Initially, the impact of weather-related elements on the PV output power is examined. Secondly, based on the research on meteorological factors, a distributed PV scene clustering method using the K-medoids algorithm is proposed. Subsequently, taking into account the spatiotemporal attributes of distributed PV data, a prediction model for PV capacity is introduced, which is founded on the application of deep learning techniques involving Convolutional Neural Networks and Long Short Term Memory (CNN-LSTM). Finally, the efficacy of the suggested approach is verified through the utilization of empirical data, providing theoretical support and data references for predicting the PV access capacity of distribution networks.