Short-term extreme wind speed forecasting using dual-output LSTM-based regression and classification model

This study introduces a methodology for forecasting extreme wind speeds (EWS) using a dual-output long short-term memory and transformer (LSTM-Transformer) model that combines regression and classification techniques. The process involves three stages: establishing extreme event thresholds using extreme value analysis (EVA), training the model on historical weather data for precise point forecasting and classification, and calibrating the output for accurate extreme event identification. The model is trained using a combination of the losses corresponding to each output with tuned weights. Evaluated using data from Los Angeles, Chicago, and Houston, for a 60 and 90 min forecast interval, the model demonstrates reasonable performance in specific climatic conditions, outperforming its single-output regression and classification counterparts in terms of both accuracy and generalisation. This indicates strong potential for real-world applications in specific regions. Crucially, the study reveals that the forecast performances of the model are closely related to the imbalance ratios, highlighting a significant link between the model's performance and the distribution of wind speed within the dataset. This highlights the importance of considering the imbalance ratio in the predictive model, especially when integrating EVA according to typical engineering practices. This innovative approach offers a reliable and flexible framework for enhancing EWS predictions, contributing significantly to the safety and decision-making processes in managing infrastructures.