Understanding probabilistic forecasting and verification in tropical cyclones

Ensemble forecasting of tropical cyclones (TC) has advanced rapidly in both research and operational community in recent years. Global model ensemble forecasts of TC tracks are routinely available from many operational centers worldwide. The National Hurricane Center provides probabilistic forecasts of surface wind speed exceeding certain thresholds using the official track, intensity and storm size forecast uncertainty from previous 5 years. However, verification of probabilistic forecasts in TCs is much more difficult than that of other continental weather systems, due mostly to their infrequent occurrence. The lack of high quality long-term observations, especially surface wind and rainfall distributions associated with TCs, makes it even more difficulty to evaluate probabilistic forecasts of TC wind and rainfall fields. This study aims to develop and test a methodology using the TRMM satellite estimated rainfall and the NOAA H*WIND data for verification of TC rainfall and surface wind forecasts.

Major sources of uncertainty in TC forecasts are from initial conditions and model errors. Ensemble forecasts can be generated by perturbing model initial conditions or model physics. To understand the uncertainty due to physical parameterizations of microphysics and planetary boundary layer, we first generate probabilistic forecasts using the 5th generation PSU-NCAR mesoscale model (MM5) with multi-nested model grids of 12, 4, and 1.3 km resolution. The ensemble members consist of various microphysical and PBL parameterizations. Probabilities are calculated at each verification time (12 hourly) over a 5-day forecast period (or until a TC dissipation at landfall). The ensemble forecasts are evaluated using the Brier Score (Ferro 2007, Hamill 2008) and reliability diagrams as described by Kay and Brooks (2000). The Brier Score is defined as BS = 1/n ∑ (P_i – O_i)² where n is the number of cases or grid points, P_i is the probability for each case, and O_i is one or zero depending on whether or not an event occurred. A typical reliability diagram shows the skill of forecasts (e.g., under or over-predict the reality). Both satellite estimated rainfall and H*WIND data are interpolated to 0.05^ox0.05^o resolution grids over a 10^ox10^o area centered around a storm. The probabilities of specific model forecast rainfall accumulation and wind speed thresholds (i.e. tropical storm or hurricane force winds of 34 and 64 kts, respectively) are compared with the observed frequency at each grid point.

Preliminary results from Hurricanes Ike and Paloma (2008), as well as Typhoon Choi-wan (2009) indicate that a major part of the forecast uncertainty in surface wind and rainfall are due to variations in track forecasts. A conditional probability is computed by using storm-relative distributions of rain and wind from the ensemble forecasts. The ensemble probabilities tend to over predict rain. This is partially due to the rain probabilities being spread over a larger area due to the track spread. There is less of an over-prediction with the conditional probability. Verification of the surface wind probabilities is currently ongoing. Very preliminary results show some over prediction of the surface wind. Ensemble forecasts from perturbation of initial conditions will be added to in the near future.