With the high percentage of PV power access, accurate and stable short-term PV power prediction has become a hot topic in existing power system planning and operation. In this paper, a prediction method (C-PASST) based on signal decomposition, deep learning and optimization strategy is proposed for PV short-term power prediction. First, the original PV sequences are decomposed using the full systematic empirical modal decomposition with adaptive noise (C-DAN), which captures the temporal features in the data using the probabilistic sparse self-attention mechanism. Then, the decomposed PV sequences are assigned to different temporal convolutional networks (TCNs) for prediction, respectively. Finally, a multiple universe optimizer (MVO) strategy based on the no-negative constraint theory (NNCT) is introduced to integrate the weight coefficients of the TCN strategies and reconstruct the final prediction results. A study of real-time PV data from Alice Springs, Australia, shows that the method outperforms other benchmark methods in four conventional performance metrics and two statistical tests, demonstrating the effectiveness of the method in predicting PV power.