In this study, 12-h quantitative precipitation forecasts (QPFs) by a cloud-resolving model for six heavy-rainfall events selected from the Mei-yu seasons (May–June) of 2016–2019 in Taiwan are examined. For each case, 30 forecasts from five experiments were made at six lead times 12 h apart (up to 60 h) using different initial fields, some with data assimilation (DA) of those on the Global Telecommunication System (GTS data) and GTS data plus sounding retrievals from polar-orbiting satellite (SAT data) to study their impacts on QPFs.

Among these six cases, three occurred when strong prefrontal southwesterly flow at low levels impinged on the island (SWF cases), with rainfall maxima of 193–307 mm in mountain interiors in southern Taiwan. For these SWF cases, the predictability was relatively high, and the QPFs also benefited from initial fields produced using a finer grid. Around 70 % of instances, the DA of additional data also exhibited positive impacts on QPFs, both over the entire Taiwan and at local rainfall centers. In the other three cases, Mei-yu fronts were present near Taiwan and heavy rainfall (peaking at 227–635 mm) from organized convective systems occurred over low-lying plains as well. Due to the high nonlinearity of deep convection, the predictability of such events was much lower with less-ideal overall performance. While the overall gain was only minimal, better QPFs for the frontal cases were still achievable at roughly 1/3 of times through DA, particularly when key information near Taiwan was ingested, typically immediately upstream from the heavy rainfall region.