Considering the characteristics of volatility and high complexity in wind power sequences, a combined short-term wind power prediction model based on Singular Spectrum Analysis (SSA), Bidirectional Long Short-Term Memory Network (BiLSTM), and ensemble learning is proposed. First, SSA is used to effectively extract features from historical power data, reducing noise interference. Second, the decomposed data is input into a prediction model that combines BiLSTM with ensemble learning to form a short-term wind power prediction model. Finally, wind power data from the Shanghai Lingang area is used to verify the feasibility and effectiveness of the proposed model. Experimental results show that compared with the benchmark model, the improved model reduces the mean absolute squared error by 55.45%, the root mean squared error by 52.62%, and increases the coefficient of determination by 9.83%. These results verify that the proposed model can effectively improve the accuracy of short-term wind power prediction.