Power generation forecasting technology for concentrated solar power(CSP) plants can help the grid to make scheduling plans in advance and effectively mitigate the impact of intermittency and fluctuation after gridconnected operation of CSP system. Taking the grid-connected power forecasting problem of CSP system in California

USA as an example

this paper investigates the correlation between meteorological parameters and power generation at first

and then uses the affinity propagation(AP) clustering algorithm to compare the differences in the impact of meteorological parameters on power generation under different generation patterns. Based on this

a long short-term memory(LSTM) neural network benchmark model is built to predict the power generation. Finally

to address the problem of low accuracy of the benchmark model in predicting the zero generation case

an incremental feature prediction model is developed based on the corresponding clustering results. The results show that the improved model works well and can achieve 91.41% accuracy in predicting daily power generation.