Uncertainty Estimation of Wind Gust Predictions Using the Deep Evidential Model

Wind gusts, one of the most unpredictable meteorological variables, wield direct influence over power outages and wind energy generation. In a prior study, we found that machine learning algorithms such as Random Forest and Extreme Gradient Boosting outperformed post-processed gust estimations from the Weather Research and Forecasting (WRF) model and were able to remove the bias almost entirely. However, the performance of the ML models was still not satisfactory for high gust values (<15 m/s) due to intermittent occurrence of wind gusts and scarcity of high gust values in the observation dataset. For a successful transition from traditional physics-based models, such as WRF, to machine or deep learning models for gust predictions, identifying the uncertainty in the predictions is pivotal. In this study, we use evidential deep learning to quantify aleatoric and epistemic uncertainty of wind gust predictions. We selected sixty one storms characterized by high wind values that occurred in the northeastern United States from the years 2005 to 2020. While the majority of the selected storms were low pressure systems accompanied by cold fronts, the accuracy of gust forecasts for each storm might vary depending on the distinct atmospheric conditions intrinsic to that storm. Therefore, we assess uncertainty for the individual storms through leave-one-storm-out cross validation instead of using a holistic approach, e.g., training on certain years of data and subsequent testing and validation on different years of gust data. After we obtain uncertainty estimates, we use evaluations graphics such as the attribute diagram, spread-skill plot, discard diagram and PIT histogram to facilitate comprehensive evaluation of the derived uncertainty estimates.