Abstract: A short-term forecast of photovoltaic power (PV) is an essential component of power plant generation planning and scheduling, contributing to the dynamic stability of the power system. To address noise interference and unstable single-model predictions in photovoltaic time series forecasting, this paper proposes a combined prediction model based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN). Firstly, important meteorological features are extracted using correlation coefficients, and the original dataset is divided into categories such as clear sky, clear-to-partly cloudy, and rainy using fuzzy C-means clustering (FCM). Next, for each similar day, the last day is the target prediction day, and the rest is historical training data. ICEEMDAN decomposes the historical training dataset into several more regular subsequences. These subsequences are then reconstructed using permutation entropy (PE). Finally, the CNN-BiGRU-ATTENTION neural network, which combines convolutional neural network (CNN), bidirectional gated recurrent unit (BiGRU), and attention mechanism, is used to predict the high-frequency, low-frequency terms, and trend terms predicted by least squares support vector regression (LSSVR), and the prediction results are superimposed to get the final Predicted value of PV. Through practical verification, this combined model effectively addresses issues such as low accuracy and unstable predictions under different weather conditions; Compared with other modal decompositions, it can more accurately predict the fluctuating local features.