There is a problem of a strong randomness, volatility and low prediction accuracy for photovoltaic power generation. Thus a prediction model based on variational mode decomposition (VMD) and an improved squirrel search algorithm optimization kernel extreme learning machine (ISSA-KELM) is proposed. First, photovoltaic power data is clustered using a Gaussian mixture model to obtain similar samples under different weather types. Second, the original photovoltaic power generation power sequence is stabilized using VMD to obtain a number of regular subsequences. Then, the KELM prediction model is constructed for different subsequences and ISSA is used to optimize nuclear and regularization parameters of the KELM. Finally, the predicted value of different subsequences is reconstructed to obtain the final prediction result. Combined with an actual example, the results show that the proposed VMD-ISSA-KELM model can obtain satisfactory prediction accuracy in different weather conditions, and is significantly better than other models, verifying its effectiveness and superiority. This work is supported by the Natural Science Basic Research Program of Shaanxi Province (No. 2021JM-393).