To address the challenges of meteorological-power response inaccuracy, difficulty in capturing abrupt features, and data scarcity in photovoltaic power prediction under extreme weather conditions, a hybrid prediction framework is proposed based on fuzzy C-means (FCM), maximum information coefficient (MIC), time variational auto-encoders (TimeVAE), 1D convolutional neural network (1DCNN), and simple-Mamba (S-Mamba). Firstly, meteorological features are clustered using FCM to categorize weather into four types: sunny, cloudy, snowy, and rainy. Subsequently, MIC is employed to select the optimal subset of meteorological features. To mitigate the scarcity of extreme weather samples, TimeVAE is adopted for data generation, leveraging its decomposed reconstruction mechanism to synthesize realistic time-series data. Finally, a 1DCNN-S-Mamba combined model is utilized, where 1DCNN captures short-term abrupt features through local convolution, while bidirectional state-space modeling in S-Mamba enables long-range dependency analysis for prediction. Experimental results demonstrate that the proposed model enhances both timeliness and accuracy in PV power prediction under complex weather conditions. Compared to S-Mamba, it reduces the mean absolute error (MAE) and root mean square error (RMSE) by 3.65% and 5.10%, respectively, in snowy weather scenarios.