Abstract: The variation of direct normal irradiance(DNI) affects the reliability and efficiency of concentrated solar power generation. A short-term DNI prediction model of a photovoltaic thermal power station in northwest China is proposed based on clustering, ensemble empirical mode decomposition(EEMD), principal component analysis(PCA), long short-term memory(LSTM) neural network and error compensation. The environmental factors that affect DNI are considered fully. Firstly, the relationship between meteorological parameters and DNI is studied. The typical days in the same weather are obtained by using the affinity propagation(AP) clustering algorithm. The each sub-mode is obtained by using EEMD to decompose the original DNI sequence, reducing the non-stationary of the sequence. Then PCA is used to get the key influencing factors, which reduces the correlation and redundancy of the original sequence and reduces the model input dimension. Secondly, the LSTM network is used to model and predict each decomposed sub-model to obtain the initial predicted DNI sequence, and the error sequence between them is obtained by making a difference between the LSTM network and the real sequence. The LSTM network is reestablished to predict the error sequence, that is, error compensation. Finally, the initial predicted DNI and the error sequence are summed to obtain the final prediction model, and the short-term DNI of the optical thermal power station is predicted. The prediction results show that the prediction model is effective and the prediction accuracy is 94%.