How Well Can Machine Learning Predict Weeks 3 and 4 Vapor Pressure Deficit?

Modeling wildfire weather risk is important both on short as well as longer timescales. Depending on the forecasting timescales, different weather and/or climate and hydrological and other factors will weigh more or less into the likelihood of high-risk wildfire conditions. The goal of this project is to develop a Machine Learning (ML) model that can improve predictions on critical wildfire risk at weeks three and four.

To do this, we are attempting to design an ML model to predict three and four week vapor pressure deficit (VPD) classifications from the GEFS model output in combination with more climatological scalar variables, such as the ENSO index and drought monitor. The daily VPD is calculated from ERA5 data and then manipulated to represent a week three and week four value that is then classified to represent wildfire potentials. This presentation will cover preparing an ML-ready dataset that uniquely processes GEFS forecast data to be representative of three and four week timescales, as well as the targeted VPD variable, and show preliminary results from the ML experiments.