In recent years, the country has vigorously promoted distributed photovoltaic power generation, and accurate and reliable photovoltaic power prediction is essential to ensure large-scale distributed photovoltaic integration into the power grid. The current distributed photovoltaic power prediction methods have not fully considered the impact of meteorological factors, making it difficult to improve prediction accuracy. To address the above issues, a distributed photovoltaic short-term power prediction method based on adaptive classification and matching of weather change processes is proposed. First, scenario partitioning of weather processes is achieved through K-Medoids-Grey, and then the convolutional neural network is optimized using an improved multiverse algorithm to achieve short-term prediction of distributed photovoltaics. Taking a distributed photovoltaic user in Gansu province, China as an example for verification. The results show that in the test set, the prediction accuracy of the IMVO-CNN method under clustering is 9.83 percentage points higher than that under non clustering, verifying the effectiveness of the method.