Wind speed sequence itself has randomness and non-stationary characteristics, which constrain the accuracy of traditional prediction methods. A hybrid prediction framework was constructed for this purpose, which integrates seasonal trend decomposition, optimized variational mode decomposition, sparrow search algorithm, bidirectional gated recurrent unit, and attention mechanism. This method first uses seasonal and trend decomposition (STL) to adaptively decompose the original wind speed data, obtaining trend terms, seasonal terms, and residual terms. Then, the optimized variational mode decomposition (OVMD) algorithm is used to perform secondary decomposition on the residual terms, obtaining several intrinsic mode functions, effectively reducing signal complexity and mode aliasing phenomena. Secondly, for the trend and seasonal components, a bidirectional gated recurrent unit (BiGRU) is used to capture long-term dependencies and periodic features, and the global hidden state is output through bidirectional information fusion of forward and backward GRUs; Introducing selfattention mechanism to dynamically weight the local detail features of each intrinsic mode function (IMF) component, generate context vectors focused on key features, and suppress redundant information interference. Next, in response to the significant impact of BiGRU key hyper parameters on model performance, the SSA algorithm is introduced for global optimization to solve the problems of traditional grid search being prone to local optima and low computational efficiency. Finally, the final wind speed prediction value is output by integrating the trend component, seasonal component, and the predicted results of each IMF component. Taking the short-term wind speed dataset of actual wind farms as the research object, the results were compared and verified with five benchmark models. The results showed that the prediction accuracy and trend capture ability were significantly better than other models. Compared with the basis benchmark model, the MSE decreased by 33.5%, RMSE decreased by 18.5%, MAE decreased by 16.9%, and R2 increased by 3.6%, training time decreased by 42.0%, and parameter number decreased by 41.0%, providing an effective solution for high-precision prediction of non-stationary wind speed sequences.