Improving Medium-Range Predictions of the Locally Extreme Precipitation with Machine Learning

Improving Medium-Range Predictions of the Locally Extreme Precipitation with Machine Learning

The random forest (RF) algorithm and logistic regression (LR) are implemented to develop skillful, calibrated contiguous United States (CONUS)-wide probabilistic forecasts of locally extreme precipitation, as quantified by 1- and 10-year average recurrence interval (ARI) exceedances. Models are created for two different 24-hour periods representing lead times of 36-60 hours and 60-84 hours. CONUS is partitioned into eight regions which exhibit similar hydrometeorological properties. Within each of these regions, a model is trained to produce probabilistic exceedances forecasts on a ~0.5°grid, based on historical forecasts spanning an 11-year 2003-2013 period. Predictor data used to generate forecast probabilities come from simulated atmospheric fields taken from a record of NOAA’s 11-member Second Generation Global Ensemble Forecast System Reforecast (GEFS/R), and includes not only the quantitative precipitation forecast (QPF) output from the model, but also variables that characterize the meteorological regime, including winds, moisture, and instability; spatiotemporal variability of fields is also considered. Results from a variety of sensitivity experiments are presented, and the use of these models to explore the physics of the forecast problem and objectively quantify the statistical biases of the GEFS/R is explored. These models are being developed for operational implementation at the Weather Prediction Center to assist forecasters with Excessive Rainfall Outlook generation. In association with that effort, forecasts from this model were evaluated during their 2017 Flash Flood and Intense Rainfall Experiment. Subjective forecaster evaluations are presented alongside objective verification during this period as well as an extended 4-year period beginning in September 2013. Overall, it is found that the machine learning-based forecasts add significant (>1 day lead time) skill over forecasts produced from the raw QPFs of both the GEFS/R and ECMWF ensemble across almost all regions of CONUS. The seminar will conclude with discussion of how well the methodology extrapolates to other datasets and predictands.

Speaker Bio
Greg Herman is currently a Ph.D. candidate in atmospheric science at Colorado State University, advised by Dr. Russ Schumacher. Greg graduated from the University of Washington with a B.S. in atmospheric science, computer science, and physics, and defended his M.S. research at CSU in autumn 2015.  His research primarily concerns the application of machine learning towards the improvement of probabilistic high-impact weather forecasts at short-to-medium range timescales, with particular emphasis on extreme precipitation forecasting. In particular, Greg has developed machine learning-based real-time forecast products for forecasting cloud ceiling and visibility at select airports, warm-season convection over northeastern Colorado, and severe weather and extreme precipitation across the contiguous United States in analogous fashion to the Storm Prediction Center’s convective outlooks and Weather Prediction Center’s (WPC’s) excessive rainfall outlooks. He is currently working with colleagues at WPC to transition the latter forecast product into operations. Greg has also performed several forecast verification studies of extreme precipitation in order to make better informed machine learning model design choices and better contextualize the performance of those models. As a graduate student, Greg has also acquired extensive field project experience from numerous different field programs across the country. Furthermore, he has also engaged in interdisciplinary collaborations between National Weather Service forecasters, social scientists, and atmospheric researchers to better understand the meteorological and communication challenges associated with concurrent and collocated tornado and flash flood hazards. Currently in the process of beginning to assemble his dissertation, Greg anticipates graduating in autumn 2018.

Tuesday, February 20, 2018
1:30-2:30
FL2-1022, Large Auditorium 

Building:

Room Number: 
1022

Type of event:

Will this event be webcast to the public by NCAR|UCAR?: 
Announcement Timing: 
February 7, 2018 to February 20, 2018
Calendar Timing: 
Tuesday, February 20, 2018 - 1:30pm to 2:30pm