In order to improve the accuracy of short-term PV power prediction and reduce the influence of climate factors on forecasting results, a short-term photovoltaic power generation forecasting scheme based on Improved Kernel Fuzzy C-Means (IKFCM) and multi-mode social spider optimization SVR is proposed. Firstly, the improved KFCM (IKFCM) clustering method is used to process the training sample set. By introducing intra class scatter clustering validity index, the automatic training sample set separation is realized and the clustering accuracy of IKFCM is improved, thereby the effect of sample data difference on prediction performance is effectively reduced. Then, the SVR prediction models corresponding to training samples set classifications one to one are built, and the Multi-modal SSO (MSSO) optimization algorithm is used to optimize the parameters of SVR model, which helps to obtain the optimal SVR parameters combination for each SVR model. Finally, the MSSO optimization SVR model is used to predict the test data. Simulation results show that, the scheme can realize accurate short-term PV power prediction for different weather conditions, and compared with other prediction algorithms, the prediction accuracy is improved by 25.2%~37.8%. This work is supported by China Spark Program (No. 2015GA780024).