Leveraging Machine Learning to Improve Short-Term PQPF

In 2021, the most deadly and damaging weather phenomenon was flash flooding, which caused 135 deaths and more than 2.4 million dollars worth of damage, according to the summary of natural hazard statistics for 2021 in the United States. However, the impacts of extreme rainfall events are ever-changing due to a warming climate. The National Weather Service (NWS) Weather Prediction Center (WPC) uses Quantitative Precipitation Forecasts (QPF) to produce several operational products to protect lives and property during heavy rain events. The WPC publishes two well-known products for short-term to medium-range flash flood forecasts: the Mesoscale Precipitation Discussion for short-term (0 - 6 hours) flash flood forecasts and the Excessive Rainfall Outlook for flash flood impacts out to 5 days. These products are informed using deterministic QPF and probabilistic QPF (PQPF), but uncertainty remains, even at short lead times, in many aspects of the forecast and its characteristics. It is imperative to have access to skillful products to create accurate forecasts for heavy rainfall events. This work leverages the machine learning (ML) architecture known as U-Net to map hourly forecasted High-Resolution Rapid Refresh (HRRR) environmental variables to the Multi-Radar/Multi-Sensor System’s (MRMS) hourly Quantitative Precipitation Estimate (QPE) to generate 1- and 6-hour PQPF products out to 24 hours that span CONUS. We will also evaluate the performance of this product through various verification techniques to ensure that reliable information about short-term rainfall uncertainty can be conveyed to forecasters and thus better inform the WPC operational flash flood forecasts.