In response to the complexity of photovoltaic power data in remote military systems and the low accuracy of existing models, this paper proposes a photovoltaic power prediction method based on ICEEMDAN and TCN-ECA-BiLSTM. First, the Pearson correlation coefficient is used to identify key variables as input sequences. Meanwhile, the photovoltaic power data is classified into stable, fluctuating, and abrupt types through FCM (Fuzzy C-Means) similarity day clustering. Next, the ICEEMDAN decomposition method is applied to decompose the photovoltaic power data, which is then reconstructed based on permutation entropy. Then, TCN (Temporal Convolutional Network) is used as a spatiotemporal feature extraction layer, with an embedded ECA (Efficient Channel Attention) unit to enhance the feature capture ability of the convolutional network. Finally, BiLSTM (Bidirectional Long Short-Term Memory) is employed for prediction, and the output is obtained. Experimental results demonstrate that the proposed model achieves high prediction accuracy and can effectively predict photovoltaic output under various power change trends.