In snow-dominated river basins of the southwestern US, water managers are tasked with producing forecasts of warm-season water supply. They have typically used linear models, such as multiple linear regression, with spring peak snow water equivalent (SWE) as the predictor. Machine learning is a promising approach for improving such forecasts; however, there exist many potential model types (e.g., regression versus tree-based) to employ. Prior studies have sought the best forecasting model type, but each type exhibits different performance depending on predictor variables, or features, and study basin. Moreover, comparatively less research focuses on the influence of different features on forecast performance. This study of multiple snow-dominated river basins in the US southwest unbiasedly compares the predictive performance of five machine learning model types by using a semi-exhaustive feature selection framework for optimizing model performance. Furthermore, this study also provides insight into which features, both currently and in the future with climate change, are important for producing accurate and stable water supply forecasts. Results suggest that optimized models of each type outperform models built only with SWE, although when compared amongst each other, all types performed similarly when using their respective optimized feature set, demonstrating the greater relative importance of feature selection over the choice of model type.
Graduate Student听Civil Engineering, CU听Boulder