Assessing the Predictability of Downslope Windstorms in the Rocky Mountain Front Range Using Explainable Convolutional Neural Networks

Convolutional neural networks (CNNs) and other machine learning techniques have been applied to forecast various types of extreme weather events in the past. While value certainly exists in being able to make skillful predictions of impactful weather, additional insights are gained by understanding why and how the model architecture makes a particular prediction. Specifically, understanding the strengths and weaknesses of the model allows the forecaster to assess the predictability of a certain weather event. This study focuses on applying explainable artificial intelligence (XAI) analysis to CNNs that forecast downslope windstorms at three locations along the Rocky Mountain Front Range (Boulder, CO, Fort Collins, CO, and Cheyenne, WY) to assess the predictability of such high wind events. Saliency maps and layer-wise relevance propagation are utilized to understand how model predictors (in this case, numerical weather model output) activate neurons throughout the network contributing to the final output. Additionally, backwards optimization is applied to create the “perfect windstorm” that maximizes the CNN’s positive prediction output. By comparing real-world windstorms to the optimal model windstorm, assessments are made of the model’s ability to forecast the real-world windstorms. Finally, unsupervised learning is applied to the predictors to create clusters of similar antecedent numerical weather prediction output. By analyzing the CNNs’ performance in each cluster, an expectation can be made of how the model will perform in advance of a windstorm given the preceding forecast from traditional numerical weather models.