Improving predictions and generating actionable forecast insights for downslope windstorms with machine learning

Abstract

Downslope windstorms are an extreme weather phenomenon characterized by accelerating winds down the lee slope of a mountain with gusts often exceeding 45 m s-1. These impact society through damage directly related to the high winds, ground transportation concerns in the vicinity of the windstorm, aviation impacts through the accompanying mountain wave turbulence, and fueling the rapid intensification and spread of wildfires such as the 2018 Camp Fire, the 2021 Marshall Fire, and the 2023 Lahaina Fire. Despite improvements in numerical weather prediction and observational datasets, predictability of these windstorms still rarely exceeds 12 hours further exacerbating their impacts. Recent advances have made machine learning (ML) more accessible to researchers and have shown promise in improving forecasts of other extreme weather phenomena. We first present models driven by two different types of ML architectures that classify wind events as moderate or high at three locations along the Rocky Mountain Front Range: Cheyenne, Wyoming; Fort Collins, Colorado; and Boulder, Colorado. The first type of architecture is the random forest (RF), which is comprised of multiple decision trees, and the second type is the convolutional neural network (CNN), which is a deep learning method that excels at image recognition. These models make forecasts at the Day 1 and Day 2 lead times based on predictors derived from a 12-km version of the WRF operated at Colorado State University. The results show improvement over the direct weather model forecasts. CNNs show enhanced event detection capability compared to the RFs but with a higher false alarm rate limiting their utility in some cases. Next, explainable artificial intelligence (XAI) techniques are presented. Feature importances indicate that the ML models rely on predictors at geographic locations that align with known atmospheric variables important to downslope wind forecast along the terrain. Also, a framework for reducing the dimensions of the predictor data and clustering these data with a Gaussian mixture model yields insights to the forecast ML models' performance and the synoptic conditions in which downslope windstorms along the Front Range occur. The ML models perform better in regimes characterized by prominent synoptic features such as cold air advection or the presence of the jet stream aloft. Lastly, we investigate whether increasing the resolution of the traditional weather model creating the ML predictors results in performance improvements. We use NOAA's High Resolution Rapid Refresh (HRRR) model to derive input predictors for newly trained CNNs and observe a decrease in false alarms that results in an overall performance boost over the direct HRRR forecasts. A case study on the Marshall Fire is conducted and indicates that the HRRR-based CNN is able to correctly forecast the subsequent downslope wind event before the wind event is explicitly depicted in the HRRR output itself. This study is an example of how ML fused with current weather models closes the forecast gap in these impactful weather phenomena with incomplete physical understandings.