Abstract: Accurate forecasting of photovoltaic (PV) power is essential for stable operation of new electricity systems. This study proposed a novel approach for short-term PV power forecasting combining the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), kernel principal component analysis (KPCA) and an improved carnivorous plant algorithm-long short-term memory (ICPA-LSTM) network. First, ICEEMDAN was employed to extract the implicit features of nonlinear signals from the meteorological data. Next, KPCA was performed to reduce the redundancy of the decomposed data and select model input parameters based on the contribution of principal components. Finally, the ICPA-LSTM model was constructed by improving the carnivorous plant algorithm (CPA). The approach was validated for PV power prediction under four typical weather conditions: sunny, rainy, cloudy, and variable weather. Results show that the proposed model reaches a determination coefficient (R²) of over 99% across all four weather scenarios, and achieves better performance compared to benchmark models.