Domain-informed multi-step wind speed forecasting: evaluating extreme wind conditions and seasonal variations

Most of the literature concentrates on single-step wind speed forecasting for low-frequency data samples, often neglecting analysing the seasonal and extreme wind speed factors in their predictions. This study conducts a comprehensive evaluation of high-temporal-resolution wind speed data from mast measurements, while concurrently analyzing local seasonal variations in wind speed, shear, wind direction, and extreme wind conditions. Furthermore, a novel multi-step deep learning-based wind speed forecasting framework, enhanced by domain knowledge, is introduced. Further, seasonal models were developed to properly characterize the model capabilities while trained on extreme wind speed observations. Interestingly, applying different feature variation techniques did not improve the model forecasting accuracies for 24 hours ahead predictions (144 steps per day for 10 minutes temporal resolution samples). Instead, a simple Long Short Term Memory (LSTM) model can effectively capture short-term variations in wind speed. The extreme value analysis identified the extreme wind speed of 40.45 m/s for the site, and the wind speed was modeled using a Weibull distribution with parameters \(A= 9.501\) $A=9.501$ and \(k= 2.682\) $k=2.682$ . The study also examined high wind events, such as the Madi cyclone in 2013, emphasizing the importance of structural stability planning and wind energy generation potential at the site. Additionally, it highlights the need to enhance the capability of deep learning models to handle extreme wind speed events and improve forecast-related energy security risk assessments.