Improving the accuracy of photovoltaic power generation prediction is of great importance to ensure safe dispatch and stable operation of the power system. In this research, an improved hybrid method is adopted which called SA-EEMD-BiLSTM. It is proposed as a solution to the challenge of short-term photovoltaic power prediction. Which employs cutting-edge techniques from machine learning, Embracing self attention(SA) mechanism, ensemble empirical mode decomposition (EEMD) algorithm, and bidirectional long-short term memory (BiLSTM) network. In this approach, the initial photovoltaic power sequence is initially broken down into multiple subsequences using the EEMD algorithm, The goal is to break down the intricate issue into more manageable sub-problems. Subsequently, the sub-sequences are reconstructed using an enhanced SA mechanism designed to uncover the connections between the sub-sequences. Ultimately, the reconstituted sequence set serves as the input for the BiLSTM model, from which prediction results for the entire issue are derived. Compared with the traditional method, SA-EEMD-BiLSTM shows better prediction effect.