Probabilistic forecast of thunderstorms using Artificial Neural Network with Google Keras libraries for deep learning

An Artificial neural network (ANN) model for classifying storm vs. no storm is implemented with the Google Keras deep learning libraries to generate one hour forecast at 1 km resolution over the conus. A feed-forward back-propagation ANN with a single hidden layer with 500 neurons is used. Eight predictors used in the operational AutoNowCaster (ANC) running within the Multi-Radar/Multi-Sensor (MRMS) and based on output from Rapid Refresh (RAP) model are used in the ANN model. The ANN model performance was compared to the corresponding performance of the operational ANC thunderstorm forecast. The ANN model was found to have significant skill as measured by standard statistics metrics to classify storm vs. no storm using Rapid RAP derived instability fields. The results of this exploratory study shows that the ANN outperforms the current operational ANC based forecast.