Accurate ultra-short-term power forecasting for distributed photovoltaic (PV) systems is critically important for their integration into the power grid. However, current forecasting efforts face challenges such as insufficient accuracy in meteorological data and incomplete power data. To address these issues, this paper proposes a regional PV forecasting method based on cluster partitioning. First, two metrics, i.e., forward/reverse electricity ratio and the ratio of power median to mean, are selected as the basis for distance calculation. The K-means clustering algorithm is then used to divide all PV power plants in the region into clusters. Based on this clustering, PV power forecasting is carried out for each cluster, and the combined results from all clusters are used to produce a regional forecast for distributed PV output. Finally, data from a distributed PV power plant in a specific region is used for verification. The results show that the proposed algorithm has high accuracy and can meet the practical requirements of field applications.