Verification of short-range ensembles for fire threat days over the Northeast U.S

High resolution temporal and spatial information about atmospheric winds, temperatures, and moisture are required for effective fire-weather forecasting. Since existing observations cannot readily diagnose the relevant small-scale weather variations across highly variable land-surface and terrain characteristics, numerical weather prediction (NWP) models are vital for developing fire-weather forecasts that help predict changes in fire danger and the potential for extreme fire behavior. Over the Northeast U.S., mesoscale NWP models have been used to better understand the atmospheric structures and evolution of fire events; however, there has been limited verification of these models over a large sample of fire weather days. Furthermore, the strengths and weaknesses of utilizing short-range (0-48h) ensembles have not been quantified for fire events over the Northeast U.S.

Stony Brook University (SBU) has been running a 13-member ensemble numerical weather prediction system down to 12-km grid spacing over the Northeast U.S. since 2006 using the Weather Research and Forecasting (WRF — 6 members)) and Penn-State/NCAR Mesoscale Model (MM5 – 7 members). Both models are run with multiple initial and boundary conditions derived from different operational centers (GFS, NAM, CMC, and NOGAPS), as well as different microphysics, cumulus parameterizations, radiative, and planetary boundary layer (PBL) schemes. From March 2006 through May of 2008, the performance of the Stony Brook University (SBU) ensemble was evaluated for fire event and threat days over the Northeast U.S. for critical fire weather ingredients such as low-level temperature, low-level relative humidity, surface wind speed and direction. This model verification was accomplished by interpolating the model to the available NWS surface and sounding stations, surface Remote Automatic Weather Stations (RAWs), and some AWS Weatherbug stations.

Most of the ensemble members have a low-level cool and moist bias during the fire threat days over the Northeast U.S., which are generally drier than normal. The largest cool bias is associated with the Mellor-Yamada-Janic PBL. This PBL parameterization does not include entrainment at the top of the PBL, which may be important in mixing dry air to the surface during some fire events. It is hypothesized that the surface cool bias is related to the soil moisture analyses being excessively moist during these generally warm and dry periods. The soil moisture analyses will be varied for specific cases using WRF to test this hypothesis. Since these biases lead to an underdispersed ensemble, a simple bias correction and Bayesian Model Averaging (BMA) will be shown for surface temperature forecasts over the Northeast U.S. for fire threat days. BMA calibrates the ensemble by weighting each member based on its past performance. BMA has been shown to improve probabilistic forecasts over the Pacific Northwest, but it has not been tested for a large ensemble over the Northeast U.S., especially for fire event days.