Evaluation of Mei-yu heavy-rainfall quantitative precipitation forecasts in Taiwan by a cloud-resolving model for three seasons of 2012–2014

Abstract

Abstract. In this study, the performance of quantitative precipitation forecasts (QPFs) by the Cloud-Resolving Storm Simulator (CReSS) in Taiwan, at a horizontal grid spacing of 2.5 km and a domain size of 1500×1200 km2, in the range of 1–3 d during three Mei-yu seasons (May–June) of 2012–2014 is evaluated using categorical statistics, with an emphasis on heavy-rainfall events (≥100 mm per 24 h). The categorical statistics are chosen because the main hazards are landslides and floods in Taiwan, so predicting heavy rainfall at the correct location is important. The overall threat scores (TSs) of QPFs for all events on day 1 (0–24 h) are 0.18, 0.15, and 0.09 at thresholds of 100, 250, and 500 mm, respectively, and indicate considerable improvements at increased resolution compared to past results and 5 km models (TS < 0.1 at 100 mm and TS ≤ 0.02 at 250 mm). Moreover, the TSs are shown to be higher and the model more skillful in predicting larger events, in agreement with earlier findings for typhoons. After classification based on observed rainfall, the TSs of day − 1 QPFs for the largest 4 % of events by CReSS at 100, 250, and 500 mm (per 24 h) are 0.34, 0.24, and 0.16, respectively, and can reach 0.15 at 250 mm on day 2 (24–48 h) and 130 mm on day 3 (48–72 h). The larger events also exhibit higher probability of detection and lower false alarm ratio than smaller ones almost without exception across all thresholds. With the convection and terrain better resolved, the strength of the model is found to lie mainly in the topographic rainfall in Taiwan rather than migratory events that are more difficult to predict. Our results highlight the crucial importance of cloud-resolving capability and the size of fine mesh for heavy-rainfall QPFs in Taiwan.