Aiming at the complexity of PV power data of military systems in remote areas and the low accuracy of existing PV prediction models, a PV power prediction method based on time-domain convolution fusion attention mechanism is proposed. Firstly, the main variables identified by Pearson correlation coefficient are used as the input sequence, and the photovoltaic power data are divided into three types of stationary, fluctuating and abrupt by fuzzy C-means (FCM) similar day clustering to improve the accuracy of the prediction model; Then, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) decomposition method is used to decompose the photovoltaic power. And reconstruct according to that permutation entropy. A temporal convolutional network (TCN) is used as a spatio-temporal feature extraction layer, and an efficient channel attention (ECA) mechanism unit is embedded to enhance the feature capture capability of the convolutional network; Finally, a bidirectional long short term memory (BiLSTM) network is used to predict the output power. The experimental results show that the proposed model has high prediction accuracy and can effectively predict the PV output under different power variation trends.