Deep Learning for High-Resolution Wildfire Modeling

摘要: We demonstrate the use of a deep learning (DL) approach for representing the behavior of a high-resolution physics-based wildland fire spread model. The ultimate objective is being able to efficiently use the DL model for intensive simulations of large fires while retaining fidelity to the fine-scale physical processes. We begin with a fire model that reduces the spatial domain of the fire spread problem to one dimension (1D). The 1D model explicitly resolves cm-scale fuel variations, heat transfer and heating/drying dynamics of individual fuel particles and burning behavior of the bed. We then ran the fire model for 78,125 factorial combinations of fuel, weather, and topographic conditions as training data for the DL algorithm. The results of the DL analysis show overall agreement of 96% of the variation in fire behavior as represented by steady state rate of spread, flame length and flame zone depth. Exceptions to the DL regression indicate areas where more work is required in refining the resolution in training cases and use of advanced methods of embedding the fire model inside the DL algorithm loop.