Evaluation of a Machine-Learning Model for the Prediction of Heavy and Extreme Rainfall

This presentation will describe a new model designed to provide guidance to forecasters at the NOAA Weather Prediction Center (WPC), with a goal of improving WPC's excessive rainfall outlooks. The model uses 10+ years of forecast output from the Global Ensemble Forecast System Reforecast-2 (GEFS-R), past observations of rainfall exceeding average recurrence intervals (ARIs), and machine learning algorithms, to produce daily forecasts of the probability of exceeding various ARIs across the continental United States. The model considers quantitative precipitation forecasts from the 11-member GEFS-R, as well as numerous atmospheric quantities relevant to rainfall prediction (including precipitable water, convective available potential energy, low-level temperatures, moisture, and winds). The machine learning method of random forests (RF) then processes that day's model output along with the historical data about ARI exceedances in 8 sub-regions of the United States, and generates a gridded probability forecast.

During the Flash Flood and Intense Rainfall (FFaIR) experiment at the Hydrometeorology Testbed and WPC in June-July 2017, this model was used as a "first guess'' for forecasters in their creation of excessive rainfall outlooks (indicating the likelihood of flooding rains) for days 2 and 3 (i.e., approximately 18-42 and 42-66 hour forecasts). The presentation will include both subjective feedback regarding the usefulness and skill of the forecasts during the experiment, along with objective evaluation of the forecasts in comparison with WPC's operational excessive rainfall outlooks. Preliminary results suggest that the model has great promise for providing useful, bias-adjusted information about the potential for extreme precipitation, in a format that is familiar to forecasters, and at forecast lead times when other types of guidance (e.g., convection-allowing models) are not generally available. Some areas requiring further development and improvement of the model were also identified in the FFaIR evaluation, which will be discussed along with the steps being taken to address them.