File attachment:
| Title | Sensitivity Analysis of PHOENIX RapidFire |
| Publication Type | Report |
| Year of Publication | 2013 |
| Authors | Chong, D, Tolhurst, KG, Duff, TJ, Cirulis, B |
| Date Published | 05/2013 |
| Abstract | An analysis of the sensitivity of the outputs of PHOENIX Rapidfire (PHOENIX) to a range of inputs and simulation parameters was undertaken. This was done using two separate methods; assessment of model response in an artificially generated idealised landscape and assessment using case-studies of real fires. The ideal landscape was used to evaluate model sensitivity in response to temperature, relative humidity, wind speed, fuel type and wind direction relative to slope. The model was evaluated under two sets of weather conditions, mild (representing moderate fire spread potential) and extreme (representing high fire spread potential). Each scenario was evaluated for each of two fuel types, forest and grass. Sensitivity was evaluated in terms of the gross area burnt when the input of interest was systematically changed while all other inputs were held constant. For all evaluations except relative wind direction, model sensitivities were compared to an equivalent area burnt using the corresponding McArthur Forest or grassland fire danger meter (assuming an elliptical fire shape). The combination of wind direction and slope resulted in simulated fires that were not elliptical, so comparisons with shapes generated with the fire danger meters were not valid. PHOENIX predictions differed from those generated using point estimates for some circumstances; however without further investigation it is unclear on what is causing these differences. Differences in predictive performance are not necessarily representative of model error, as there are a number of differing assumptions between the systems used. However, specific situations have been flagged for follow up work. Two case study areas were used for PHOENIX model sensitivity evaluation; Wangary and Kilmore. Case study fires were simulated using observations from the day that the fires occurred with one input systematically varied. Three inputs were evaluated using the case studies; simulation resolution, start time (simulating fire ignition to occur earlier and later than observed) and start location (varying the ignition location in space). Sensitivity was evaluated by considering the change in the Area Difference Index (ADI, an index of the ratio between incorrectly predicted burnt area and the correctly predicted burnt area) from the baseline scenario (simulation resolution of 180m, ignition location and time as observed. Predictive performance varied wide with changing inputs. In general as the difference in input value to the ‘best estimate’ increased, predictive performance degraded. |
