Summary
The proposed work will explorean innovative ML scheme based on the advanced recurrent neural networks – Long Short-TermMemory (LSTM) that will exploit the time-dependent skills rendered by additional predictorssuch as convective inhibition, convective available potential energy, and precipitable water. Thenetwork will incorporate embedding (a regularization approach) to sharpen the1predictor-predictand relationship. These innovations have the potential of enhancing the skills ofPQPF for thresholds beyond 3-5in/day for lead times out to day 8.The precipitation type forecast scheme will be based on a multiclass ANN scheme demonstratedin Scheuerer et al., (2020) for precipitation amount forecast. Adoption of this method forprecipitation type forecasts will mark a major advance from an existing regression-based methodemployed in the Localized Aviation MOS Program in terms of robustness, and will be morecomputational efficient than the Bayesian Classifier by Schuerer et al. (2020). The scheme willbe developed and trained using ASOS, NSSL MPing data set and NEXRAD Hydrometeor Classproducts to produce high-resolution (3-km) probabilistic precipitation type forecasts.