Abstract: Short-term photovoltaic (PV) power forecasting often suffers from significant prediction errors due to nonlinear variations in meteorological conditions. To address this issue, a novel forecasting method integrating feature clustering with the Superb Fairy-wren Optimization Algorithm-based TimeMixer (SFOA-TimeMixer) is proposed. First, key meteorological features were selected using mutual information (MI) to reduce irrelevant inputs. To mitigate data heterogeneity, a feature clustering algorithm, FMVHC, was developed. This algorithm reconstructs features using Variational Mode Decomposition (VMD) and constructs a composite distance matrix by integrating Euclidean distance, cosine similarity, and Pearson correlation coefficient. An improved hierarchical clustering approach then categorizes daily data into three typical weather types: sunny, cloudy, and rainy. Next, a dual-stage optimization strategy is employed to enhance the adaptive modeling capability of the forecasting framework. The SFOA algorithm is applied separately to optimize both the VMD parameters for feature reconstruction and the hyperparameters of the TimeMixer model. Finally, the SFOA-TimeMixer model performs separate predictions for the datasets of the three weather categories, enabling more accurate capture of complex temporal patterns. The method was validated using 2022 data from a PV power station in Jiangsu Province. Results show that the proposed approach outperforms other comparative models across all three weather conditions, significantly improving prediction accuracy. The integration of feature clustering and dual-stage optimization effectively addresses the challenges of nonlinear meteorological variations and heterogeneous data in short-term PV power forecasting, providing a robust framework for practical deployment in PV generation management.