FV3-LAM CAM Ensemble Predictions and Consensus Products for Predicting Heavy Rain for the Hydrometeorology Testbed

Flash flooding and intense rains have contributed to several billion-dollar high-impact weather events in the United States in recent years. High-resolution ensemble forecasts show promise toward improving predictions of such high-impact events. The US National Weather Service is developing a Unified Forecasting System (UFS) utilizing the FV3 model core for the atmospheric prediction component. In the next few years this will include a high-resolution Convection-Allowing Model (CAM) ensemble, the Rapid Refresh Forecasting System (RRFS). To maximize the utility of RRFS for predicting intense rainfall events, it is thus necessary to explore high resolution ensemble consensus forecasts and ensemble consensus products for improving forecasts of heavy rain that might lead to flash flooding and other impacts.

In recent years, the Center for Analysis and Prediction of Storms (CAPS) has designed ensemble forecast products and run real-time forecasts using the FV3-Limited Area Model (FV3-LAM) for the Hydrometeorology Testbed (HMT) Flash Flood and Intense Rainfall Experiments (FFaIR). In recent years, CAPS has produced real-time CAM ensembles consisting of 12-16 members covering a CONUS-wide domain on a 3 km grid. The ensembles were created through variation of the physics suites employed and by applying perturbations to the initial-conditions and lateral boundary conditions. Ensemble consensus products have been developed that include an ensemble mean, a probability matched (PM) mean, a localized probability matched (LPM) mean, and a spatially aligned mean (SAM). The latter of these aims to preserve the structure and local maxima in individual CAM members by locally adjusting precipitation fields to a common location.

Beginning in 2022, the CAM output products included the use of artificial intelligence machine learning (ML) to predict rainfall, using a U-Net deep learning approach to forecast the probability of rainfall exceeding 0.5 and 1.0 inches during 6-hour periods. The algorithm used a combination of selected CAPS CAM ensemble members and operational High Resolution Ensemble Forecast (HREF) members. The CAPS U-Net for rainfall prediction in the 2022 FFaIR was trained on CAPS FV3 and HREF data from the summers of 2020 and 2021.

The consensus methods described in this section have been evaluated subjectively and objectively scored using the MET-Plus verification tools. One result from 2022 FFaIR is included in the abstract image, Spatial Bias and Equitable Threat Score at the 1-inch threshold for 24 h precipitation. Results for individual ensemble members are shown in the colored bars, plus three different ensemble means in the orange shades, at 3 verification times, 12-36h, 36-60h, 60-84h over 17 cases from 2022 FFaIR. The results show some variation among members, but also illustrate the value of the ensemble means as they score well above any individual member. Details of the ensemble consensus and ML methods and more ensemble results from recent FFaIR experiments will be presented at the conference.