Time-Lagged High-Resolution Typhoon Ensemble Forecasts for Taiwan by the Cloud-Resolving Storm Simulator (CReSS)

High-resolution deterministic forecasts from a single model and ensemble forecasts from a group of relatively low-resolution models are two different but complementary approaches in our modern strategy to produce forecasts used to inform the public and for hazard prevention and reduction. Each approach has its strengths and shortcomings, and there is often a tradeoff between them under limited computational resources. For typhoon emergency action and hazard reduction in Taiwan, realistic rainfall scenarios (amount and distribution) associated with different tracks are the most important and useful information, and can only be provided by high-resolution models at preferably cloud-resolving grid sizes. However, it is generally regarded that deterministic models provide no probability information to quantify forecast uncertainty (as an ensemble system does), and they also give inadequate lead-time. Thus, how to improve these two major drawbacks of high-resolution models becomes an important question in our forecast strategy. To tackle the above issues, eight-day experimental forecasts using the Nagoya University Cloud-Resolving Storm Simulator (CReSS), at a grid spacing of 2.5 km and a domain of 1860 km by 1360 km, were carried out at real time during the typhoon season for Taiwan since 2012, initially only once daily (at 0000 UTC). The National Centers for Environmental Prediction (NCEP) Global Forecast Systems (GFS) operational analyses and forecasts are used as initial and boundary conditions. Using these 8-day forecasts, the author demonstrates that such a strategy can produce highly realistic rainfall scenarios needed for emergency action, for some typhoon cases with a lead-time of 4-5 days or more when the track errors remain relatively small. Moreover, ensemble information is also available through multiple time-lagged runs. If executed 4 times a day, the probability information generated by such a time-lagged ensemble from a single model is roughly comparable to a low-resolution ensemble system with 25 members making 3-day forecasts, using also the same computational resources (about 1600 cores) but giving 8-day forecasts with more realistic rainfall scenarios, longer potential lead time, and overall improved intensity and track due to the high resolution.