11A.4 A Fuel Moisture Model for WRF-SFIRE from HRRR and RAWS Data by a Physics-Initialized Recurrent Neural Network

Wednesday, 31 January 2024: 2:30 PM
345/346 (The Baltimore Convention Center)
Jan Mandel, Univ. of Colorado Denver, Denver, CO; and J. Hirschi, A. Kochanski, A. Farguell, D. V. V. Mallia, B. Shaddy, A. A. Oberai, K. A. Hilburn, and J. Haley

The WRF-SFIRE modeling system integrates the Weather Research Forecasting (WRF) model with a wildfire spread model using the level set method, and a fuel moisture content (FMC) model. FMC plays a pivotal role in wildfire behavior, influencing its diurnal variability, rate of spread (ROS), severity, and plume development. Presently, WRF-SFIRE's FMC model computes FMC equilibria based on atmospheric variables (such as temperature, relative humidity, and rain) and employs a time-lag model. During its training phase, the model assimilates RAWS FMC measurements using an augmented extended Kalman filter.

Our research pivots from the Kalman filter approach. Instead, we employ a Recurrent Neural Network (RNN) to estimate RAWS FMC data directly from a time series of the HRRR product. A dedicated layer within the network eases the transition from an air moisture absorption or drying regime to a wetting by rain scenario. Moreover, initializing the network weights to emulate a physical FMC model proves crucial for the training to converge to a satisfactory solution. In its predictive capacity, the RNN can be utilized to estimate FMC either from the HRRR forecast or directly from WRF variables.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner