Fire spread prediction using a lagged weather forecast ensemble

Abstract

From October to early December 2014 the Bureau of Meteorology conducted a pilot trial of a Rapid Update Cycle (RUC) numerical weather model. The model was run at a horizontal resolution of approximately 1.5 km, with hourly updates to the forecast, and was nested within the operational ACCESS suite of weather models. At this resolution, convection can be resolved without parameterization and the influence of topography on near surface conditions should be better captured. As part of this trial the model data was made available to the NSW Rural Fire Service in order to gauge how useful it would be in an operational setting. We present a set of three case studies of wildfires that occurred during the RUC trial period. In each case a short overview of the fire is presented, and the broad synoptic scale weather pattern outlined. The 0000UTC RUC model run and the morning official Australian Digital Forecast Database (ADFD) nearest grid cells were compared with observations at a nearby Automatic Weather Station (AWS) location as a check on the broad scale performance of the model. The gridded weather at the fire ignition point was then compared with the gridded weather at the AWS site to assess any local variation in expected conditions, e.g. wind change timing. The RUC guidance and ADFD guidance were then used as input to the Phoenix fire spread model, with an examination of the relative performance of each model in predicting the resulting fire spread. For each case, a proof of concept lagged ensemble 'probability burnt' prediction map was produced using five consecutive (unweighted) RUC forecasts. RUC model runs were initiated hourly on the morning of each case study from 0700 local time (2000UTC on the previous day). All times are given below as UTC. One issue that we encountered with attempting to construct a lagged ensemble type fire spread product was that the intermediate RUC model runs (i.e. those falling between the major six hourly updates) had quite short forecast time domains, and an overlapping period of a few hours only was available. This was considered sufficient to assess the initial phase of fire spread, however, and comparisons with the actual burn extent generally become more difficult on longer time scales as active suppression becomes more of a factor. The selection of cases to examine was based on a combination of an interesting weather situation, a sufficiently large fire run, and the availability of aircraft linescan data at appropriate times to verify the actual fire spread. We also wanted to ensure that we examined at least one case in predominately grass fuels and one case in predominately forest fuels. None of the cases selected occurred under extreme weather conditions, as little extreme weather occurred during the RUC trial period, and uncertainty around the applicability of the fire behaviour models that underpin Phoenix in extreme conditions would make comparisons with the actual fire spread maps less certain. We also present a short discussion on the sensitivity of the Phoenix model to fuel state, ignition pattern and the weather forecast, and consider the relative magnitudes of these sensitivities with reference to one of the case study fires, showing that sensitivity to the weather forecasts is comparable to the other inputs.