Addressing the issue of inaccurate photovoltaic (PV) power prediction due to changing environmental factors and non-stationary sequences, this study introduces a novel PV short-term power prediction model. Unlike traditional machine learning algorithms that merely fit historical datasets, this model combines Complete Ensemble Empirical Mode Decomposition (CEEMD) and Long Short-Term Memory (LSTM). Firstly, using CEEMD, meteorological data were decomposed into multi-modal feature data, capturing temporal and frequency characteristics. Then, the LSTM network then modeled these as time series, preserving their inherent properties and non-stationary features. After training on decomposed signals, power forecasts were obtained by superimposed reconstruction. Validated with an 8-month dataset from a southern China PV station, the method enhanced short-term PV power forecasting accuracy.