To address the non-stationarity, strong noise interference, and feature redundancy in wind power sequences, a hybrid forecasting model integrating feature selection with the multi-strategy improved slime mould algorithm (MISMA)-optimized multivariate variational mode decomposition (MVMD) and support vector machine (SVM) has been proposed to improve the short-term wind power prediction accuracy. First, the density-based spatial clustering of applications with noise (DBSCAN) algorithm was used to detect wind power outliers and correct them. Random forest (RF) was then employed to screen for feature variables strongly correlated with wind power output, thereby reducing input redundancy. Secondly, a multi-modal variational decomposition was introduced to perform time-frequency synchronous decomposition on power sequences and feature sequences, yielding an equal number of mode components with aligned frequencies to reduce data volatility. Furthermore, a multi-strategy improved slime mould algorithm was obtained by integrating the initialization of the Logistic-Tent chaotic map, generalized normal distribution search, leader mechanism position update, and adaptive proportional mutation strategy, thereby enhancing the search capability of the traditional slime mould algorithm. Based on this, MISMA was used to optimize the MVMD parameters and SVM hyperparameters in stages, and a MISMA-MVMD-SVM prediction model was constructed. Finally, simulation experiments were conducted using actual wind power generation data from the Hami region of Xinjiang, and comparisons were made with different models. Melting experiments were also designed to verify the effectiveness of each module. The experimental results showed that the proposed model demonstrated significant advantages in terms of prediction accuracy and stability, providing an efficient and reliable solution for wind power forecasting.