To enhance the precision of ultra short-term photovoltaic (PV) power prediction, this paper proposes a PV power prediction model driven by multi-scale feature fusion. Firstly, the ICEEMDAN decomposes the original PV power sequence into several modal components and permutation entropy (PE) is used to reconstruct data to decrease the computational complexity. Secondly, the K-means clustering algorithm is used to classify weather patterns based on reconstructed dataset. Finally, a PV power ensemble prediction model is constructed by combining Bidirectional Long Short-Term Memory (Bi-LSTM), Temporal Convolutional Network (TCN), and Multilayer Perceptron (MLP), with a self-attention mechanism integrated to achieve multi-scale feature fusion by effectively capturing both long-term dependencies and local features, and significantly improves the accuracy of PV power prediction. Experimental results show that the proposed model outperforms individual models in PV power prediction, with more precise fitting performance across different time periods, confirming its effectiveness and reliability.