Improving the accuracy of PV energy production estimates is essential to ensure the safe command and continued reliable operation of the power network. A novel photovoltaic power prediction method is proposed, which combines the variational mode decomposition (VMD). This paper analyzes how various ecological factors affect the efficiency of solar photovoltaic power generation system, and uses variational modal decomposition method to analyze the sequence data of related environmental factors. This kind of analysis is very helpful to reveal the signal fluctuation of different scales over time, which can significantly improve the stability of the environmental variable series. Then, the KPCA algorithm is used to identify the key elements that determine the feature arrangement, which not only removes all the correlations and redundancies in the initial sequence, but also reduces the dimensions of the input to the model. The LSTM network is finally used to dynamically extract the temporal features of the multi-dimensional feature train to generate the prediction of the solar projection. Monitoring data from the plant in Australia for verification purposes. The research data reveal that the innovative VMD-KPCA-LSTM algorithm achieves more excellent prediction performance than SVR, RNN, LSTM, VMD-RNN and VMD-LSTM in the equivalent environment. The VMD-KPCA-LSTM algorithm can significantly reduce the instability of the environmental factor sequence and remove the redundant information in the photovoltaic power prediction, thus achieving a more accurate prediction. The use of this prediction technique has a crucial impact on optimizing the accuracy of solar photovoltaic energy production in the short term.