Ty Dickinson - April 20

Weather and Climate Systems Forecasting Subseasonal Extreme Precipitation in the Contiguous United States: Examining the Utility of Random Forests Ty Dickinson Wednesday, April 20 03:00 PM Join Google Meet: https://meet.google.com/dai-kdtr-gtz Extreme precipitation across multiple timescales is a natural hazard that poses a significant risk to life, with a commensurately large cost

Start

April 20, 2022 - 3:00 pm

End

April 20, 2022 - 4:00 pm

Weather and Climate Systems

Forecasting Subseasonal Extreme Precipitation in the Contiguous United States: Examining the Utility of Random Forests

Ty Dickinson

Wednesday, April 20

03:00 PM

Join Google Meet: https://meet.google.com/dai-kdtr-gtz

Extreme precipitation across multiple timescales is a natural hazard that poses a significant risk to life, with a commensurately large cost through property loss. This study aims to quantify the utility of random forests in predicting subseasonal extreme precipitation events. We employ a database of 14-day extreme events to classify days between 1950 and 2018 as either extreme or not extreme, serving as the predictand for our models. The database was designed such that it identified periods with prolonged, persistent precipitation, likely to have relatively larger synoptic- and global-scale contributions to the extreme event. The database also categorized events into geographic regions; the present work will compare and contrast models built for the Central West Coast, the Central Plains, and the Ohio River Valley. Important precursor variables identified via lag composites including geopotential height, horizontal wind components, and integrated vapor transport from ERA5 reanalysis are used as predictors in the model design. Models for each region are fit using daily data from 1 January 1950 through 31 December 2000 and trained with data between 1 January 2001 and 31 December 2018. Select random forest hyperparameters are tuned using 10-fold cross validation. Results show that model performance varies from region to region with the source of model error also varying spatially. Although geopotential height is the best predictor in all regions, other predictors, such as sea-level pressure, have varying degrees of importance.