Automation of high impact weather forecasts is sometimes considered to be more challenging than the automation of routine weather. We compare the official forecasts of fire danger to forecasts based on an automated output. Official forecasts are prepared by operational meteorologists and issued to the public. Automated forecasts are based on a bias-corrected, poor-man’s-ensemble of NWP models. Operational meteorologists often alter the parameters that go into the forecast fire danger indices to increase the severity of the fire danger forecasts, aiming to predict the daily maximum fire danger, particularly when conditions are expected to be more extreme. In comparison, the current automated forecast guidance attempts to predict the on-the-hour fire danger.
Statistics such as a mean square error are useful to give an overall idea of forecast performance, however they may hide the ability of different forecast systems to predict severe fire danger. To understand the abilities of each forecast to predict warning conditions, we calculate the relative economic value for various thresholds. Additionally, we examine the potential relative economic value, which shows the value of forecasts when calibrated to a user's specific sensitivities.
Relative economic value curves allow us to understand how well the two different forecast sources predict warning conditions for a range of users with different sensitivities. Initial results suggest that the automated forecasts provide good value on days well below warning thresholds and may support decisions such as whether conditions are suitable to perform a prescribed burn. However, across Australia, the official forecasts generally provide more value for distinguishing between days that are above or below warning thresholds. Relying on the current un-calibrated automated forecasts would not be good for decisions such as whether to issue a fire ban.
We will present some of the results and discuss the implications regarding automation of fire danger forecasts.