The accurate prediction of PV power is very important for the safe and stable operation and real-time control of the integrated energy system. In order to solve the problems of noise interference in photovoltaic power prediction and poor prediction accuracy of traditional single prediction model, a short-term photovoltaic power prediction model based on ICEEMDAN and TCN-AM-BiGRU is proposed. Firstly, the Pearson correlation coefficient was used to screen the key meteorological factors, and the historical PV power data were divided into three similar days: sunny, cloudy and rainy by fuzzy C-means clustering. Secondly, ICEEMDAN is used to decompose the historical training set into several regular subsequences and reconstruct them according to the permutation entropy. Finally, the sequence features are extracted by TCN, the attention mechanism is introduced to assign different weights, and then the prediction is made by BiGRU to output the final prediction result. Taking the actual data of a photovoltaic power station as an example, the prediction model and other models were verified and analyzed. The results showed that in sunny, cloudy and rainy weather, compared with other comparison models, the accuracy of the proposed model increased by 1.69%, 3.58% and 4.40% on average, the MAE decreased by 57.61%, 36.83% and 40.94% on average, and the RMSE decreased by 56.90%, 34.30% and 36.63% on average, which verified the effectiveness and superiority of the proposed model.