Open Access

ARTICLE

Faster Than Real Time Stochastic Fire Spread Simulations

A.R.Ervilha^1,2, F.A.Sousa¹, J.M.C.Pereira¹, J.C.F.Pereira¹

1 Instituto Superior Técnico, Technical University of Lisbon, Av. Rovisco Pais, 1, 1049-001, Lisbon, Portugal
2 Email: eita.ervilha@ish.utl.pt

Computer Modeling in Engineering & Sciences 2012, 89(5), 361-387. https://doi.org/10.3970/cmes.2012.089.361

Abstract

Faster than real time stochastic fire spread predictions are reported using a Non-Intrusive Spectral Projection (NISP) method based on Polynomial Chaos expansion and Graphic Processing Units (GPUs). The fireLib BEHAVE model together with a raster surface fire growth algorithm was implemented using the Compute Unified Device Architecture (CUDA) programming language. The uncertainty generated by the four random variables considered (wind speed, wind direction, fuel moisture, and fuel load) is quantified in the stochastic solution. Stochastic simulation of an idealized vegetation fire in a realistic complex terrain is obtained with speed-ups as high as 176 when compared to Central Processing Unit (CPU) and two orders of magnitude faster than real time fire propagation. The results include the fire front location and its error bar area, based on a 95% confidence interval, as well as temporal Probability Density Functions at selected points that quantify the uncertainty on the fire spread.

---

Keywords

---