Abstract: The day-ahead photovoltaic (PV) power generation forecasting was an important evidence in power grid economic scheduling, and the K-means clustering algorithm and neural network were widely used in the power generation forecasting. In allusion to the defects in K-means clustering algorithm, i.e., not easy to determine the initial clustering center and the number of clusters, and the imperfection of neural network, i.e., too many training parameters and easy to falling into local optimum, an algorithm, in which the DPK-means was combined with extreme learning machine (ELM), was constructed to implement the day-ahead forecasting of PV power generation. Firstly, utilizing density peaks clustering (DPC) abovementioned defects in K-means clustering was revised. Secondly, the DPK-means algorithm was used to carry out the clustering analysis on historical meteorological data samples, on this basis an ELM forecasting model was established to implement the forecasting of day-ahead PV power generation. Case study results show that using the proposed combined forecasting algorithm a better forecasting result can be obtained, so the proposed algorithm is feasible.