Exploring Additional Predictors to Improve Probabilistic Forecasting of Thunderstorms Using a 2-D Convolutional Neural Network (CNN)

Machine learning is a powerful tool with the potential to advance the National Weather Service’s objective to provide probabilistic Impact-based Decision Support Services (IDSS) by making skillful probabilistic predictions. A type of machine learning model called a 2-D convolutional neural network (CNN) was developed by NWS’s Meteorological Development Lab (MDL) to produce short-range probabilistic thunderstorm forecast guidance for aviation and other applications. In previous work, the model was trained using thirteen predictors derived from the High-Resolution Rapid Refresh (HRRR) model and produced storm forecasts with skill matching or exceeding HRRR composite reflectivity forecasts. Our work investigates whether the addition of a new training predictor further improves forecast skill. Two new input predictors are individually tested: lightning rate from the HRRR and the probability of convection over an hour from MDL’s Localized Aviation MOS Program (LAMP). In both cases, training using the new predictor improves model performance, with this improvement being significant in some instances depending on the performance metric and forecast lead time. The CNN model using the predictor from LAMP demonstrates improved bias due to fewer false alarms. This presentation will discuss the results of training and validation of the CNN model using the two new predictors. Model performance is evaluated subjectively using case studies and objectively through statistical analysis. Finally, future work, such as tuning of additional model parameters, is suggested to further improve probabilistic forecast capabilities.