WCD-20. Forests and Trees as Tools for Wildfire Power Modeling

Abstract
For this research, the new RAVE Fire Radiative Power (FRP) data has been combined with Rapid Refresh (RAP) numerical weather model data in machine learning random forests to explore combinations of variable inputs and modeling structures to model hourly FRP into the future. The RAVE dataset is a fairly new combined satellite product, July 2019, that provides hourly FRP measurements from both geostationary and polar orbiting satellites. Additionally, Eric James' Hourly Wildfire Potential (HWP) was experimented as a solo input as well as in combination with previously observed RAVE FRP observations. The goal of this work is to find a random forest or other decision tree machine learning model and input combination which improves upon the current way hourly FRP is modeled. By reducing errors in this variable in forecasts, corresponding smoke and aerosol models might also benefit with forecast improvements. Traditionally, controlled and wild fire behaviors are very different, even if both have detectible FRP. Given the rapid decline in controlled burn's FRP, these points have not been included in previous model datasets. This time around, however, models included FRP from both types of fire since both contribute aerosols into the atmosphere. The results and progress of this work will be shown here.