Evaluation of multi-member ensemble quantitative precipitation forecasts in the Taiwan Area Heavy-rainfall Prediction Experiment (TAHPEX) for dry-run Mei-yu events: higher predictability over mountains than plains in Taiwan

In this study, the skill of 24—h quantitative precipitation forecasts (QPFs) for nine verification periods in three Mei-yu events during dry-runs by the cloud-resolving multi-model ensemble (with grid sizes of roughly 1 − 3 km) in the Taiwan Area Heavy-rainfall Prediction Experiment (TAHPEX) is evaluated. Categorical statics of threat score (TS) and bias score (BS) at thresholds of 50 − 350 mm (per 24 h) are employed, for QPFs out to five days at most. Overall, the members show improvements in QPF skills compared to previous forecast verification studies, with TS reaching at least 0.22 at 0 − 24 h, 0.19 at 24 − 48 h, and 0.15 at 48 − 96 h at thresholds of 130 mm and below. The TSs are only slightly lower at 200 mm, but in general ≤ 0.1 at 350 mm. The lead time with some skill is also extended to beyond three days. Most members, however, under-predict heavy rainfall with BS < 1, more serious toward higher thresholds, at ranges close to 48 − 72 h and beyond, and over the plain areas (≤ 300 m in elevation) compared to the mountains (> 300 m). A key result of the present work is that the skill of QPFs and thus the predictability is significantly higher over the mountains in Taiwan than the plains, as a large component of mountain rainfall is phase-locked and stationary. Thus, the mountain regions consistently exhibit higher TSs, BSs closer to 1 (with less under-prediction), and higher probabilities of heavy rainfall derived from the ensemble runs across all members, lead times, and thresholds examined. Among the ensemble products, the exceedance probability from the most-rainy member often has higher TSs than all other products, including probability matching, because it has the least under-forecast and can over-forecast. It is more useful over the mountains. Finally, the finer 1—km members, executed only once per day, show good potential for further improvement to QPFs in heavy-rainfall scenarios.