Accurate forecasting of wind power ramp events—sudden and unpredictable changes in wind generation—is crucial for maintaining grid stability and ensuring efficient energy management, especially with the increasing reliance on renewable energy. This study addresses the challenge of capturing complex wind fluctuations by proposing a hybrid model that integrates Variational Mode Decomposition (VMD), Long Short-Term Memory (LSTM) networks, and eXtreme Gradient Boosting (XGBoost). The model utilizes historical wind speed, direction, and external weather variables such as temperature and gusts, with ramp detection performed using the Definition-Based Sign Indicator (DSI) method. Validated across four geographically diverse wind farms with an 8-h ahead forecasting horizon, the model demonstrates high ramp detection precision (0.814–0.903) and strong overall performance in terms of MSE, RMSE, MAE, and sMAPE. Results confirm that combining signal decomposition with deep and ensemble learning significantly improves forecasting accuracy and generalizability, making the model highly suitable for real-world wind energy integration scenarios.