Advanced and Optimized Forecasting Techniques for Wind Power Generation: A Comparative Analysis

Wind power forecasting and subsequently ramp event detection are vital for maintaining the stability and efficiency of modern power grids that increasingly rely on renewable energy sources. This study presents a new hybrid forecasting framework that combines Long Short-Term Memory (LSTM) networks with Extreme Gradient Boosting (XGBoost) to address the limitations of each model individually. While LSTMs are effective at capturing temporal dependencies, they tend to perform poorly in detecting rapid changes such as ramp events. Conversely, XGBoost provides accurate point predictions but lacks inherent temporal awareness. By merging these models, our framework leverages LSTM’s ability to model time sequences and XGBoost’s strength in predicting sharp transitions. This integration markedly improves short-term forecasting accuracy, especially during dynamic ramp conditions. The framework focuses on short-term forecasting with a 12 -hour horizon using 24 hours of historical and forecasted exogenous data. Tested on a four-year meteorological dataset from Atlantic Canada, our results show that this integrated approach, leveraging the complementary strengths of hybrid deep learning (Dual Transformer-LSTM) and tree-based models (XGBoost), achieve a $30 \%$ relative improvement in deterministic and a $40 \%$ improvement in deterministic accuracy when capturing both general and ramp-specific features of wind power signals compared to the other two methods.