To solve the problem of low accuracy in photovoltaic power prediction caused by traditional machine learning methods in the face of changing environmental factors and non-stationary sequences, a photovoltaic short-term power prediction model based on complete empirical mode decomposition (CEEMD) and long short-term memory neural network (LSTM) was proposed. Firstly, four environmental factors affecting photovoltaic output, namely solar irradiance, relative humidity, atmospheric pressure, and air temperature, were fully considered. The meteorological factor characteristic curve was decomposed into multimodal feature data through CEEMD, accurately capturing its different time scales and frequency characteristics, and thus fully preserving the non-stationary characteristics of environmental data. Secondly, based on this, the LSTM network was used to model the time series of multimodal feature data, aiming to preserve the seasonal and non-stationary features of the time series and provide more accurate input features for subsequent modeling. Finally, by training the decomposed signal and adaptively adjusting the prediction model parameters based on changes in input data, the prediction model for specific scenarios was iteratively generated to flexibly respond to real-time environmental changes and obtain corresponding power prediction results. The validation was conducted on an 8-month meteorological and power dataset of a 37 kW sub array distributed photovoltaic power station on an isolated island in Hainan. The experimental results show that the proposed method has significant advantages in preserving environmental data details and local characteristics, and has good adaptability under different meteorological conditions, effectively improving the accuracy of short-term photovoltaic power prediction.