Short-term wind power prediction is crucial for power system scheduling and operational security. However, the accuracy of such predictions is severely compromised by the inherent strong randomness and non-stationarity of wind power data, as well as limitations in existing methods, including insufficient shape-preserving capability in data preprocessing, modal aliasing, and inefficient parameter optimization in prediction models. To address these issues, this paper proposes a novel hybrid framework combining a piecewise cubic hermite interpolating polynomial (PCHIP) with variational mode decomposition (VMD) for data preprocessing and a sparrow search algorithm (SSA)-optimized long short-term memory (LSTM) network for prediction. First, abnormal values in raw wind power data are identified and repaired using PCHIP, which preserves the local monotonicity and curvature of the original sequence through Hermite interpolation. Second, the preprocessed data are decomposed into four intrinsic mode components (IMFs) via VMD to capture multi-scale temporal features. Finally, the stabilized IMF sequences are input into the SSA-LSTM wind power forecasting model to yield prediction outcomes. Experimental validation using 21-day measured power data from a wind farm demonstrates that the proposed model achieves a fitting degree of 0.989 1 with actual values, improving prediction accuracy by 5.558% compared to conventional LSTM, thereby verifying the effectiveness and superiority of the method.