Enhancing wind power forecasting and ramp detection using long short-term memory networks and the swinging door algorithm

Abstract

Accurate prediction of short-term wind power ramps is essential for effective smart grid management. This study introduces the swinging door algorithm for ramp detection, which outperforms traditional methods by precisely identifying ramp events. Additionally, a long short-term memory (LSTM) network is evaluated against established models such as support vector machines, artificial neural networks, convex multi-task feature learning, and random forest for wind power ramp forecasting. The LSTM model demonstrates superior performance, achieving the lowest weighted mean absolute percentage error of 8.36% and normalized root mean squared error of 0.60, alongside the highest R-squared (R²) value of 0.73, indicating strong predictive accuracy and correlation with observed data. Furthermore, the combined swinging door algorithm-LSTM framework improved ramp event detection by 15% compared to traditional methods, showcasing its robustness in capturing both mild and extreme ramp events. This research underlines LSTM's effectiveness in wind power forecasting, marking a notable advancement in prediction methodologies. By illustrating the strengths of LSTM and swinging door algorithm, the study contributes to the refinement of prediction models for smart grid applications, highlighting their potential to transform wind power ramp prediction and detection.