Quantitative Precipitation Forecasts Using Numerical Models: The Example of Taiwan

The quantitative precipitation forecastQuantitative Precipitation Forecast (QPF) (QPF) is a challenging area in modern numerical weather predictionNumerical Weather Prediction (NWP) but is crucially important and demanded by the society, especially those of heavy-rainfall events due to their high potential to cause hazards like flooding, landslide, and debris flow. During the past decade, heavy-rainfall QPFs have been shown to improve considerably by using a cloud-resolving modelCloud-Resolving Model (CRM) at a grid size of 2.5 km in Taiwan, where its steep topography enhances the rainfall from two major sources: typhoonsTyphoon (July–October) and Mei-yuMei-yu (May–June). Within the short rangeShort range of day 1 (0–24 h) to day 3 (48–72 h), the threat scoreThreat Score (TS) (TS) of 24-h QPFs for all typhoon events is about 0.7 at the lowest threshold of 0.05 mm (per 24 h), but stays ≥ 0.2 (indicating certain skill) up to almost 500, about 450, and over 250 mm on days 1–3, respectively. Moreover, for the top 7% rainiest events, the short-range QPFs have TSs close to 1.0 at 0.05 mm on days 1–3 and about 0.3 higher than those for all events up to 50 mm, changing to 0.2 higher at 100 mm, and 0.1 higher at 350 mm. Thus, as previously found, the skill level is higher for larger rainfall events, in contrast to the common beliefs of many. Although not as high as for typhoonsTyphoon, the skill of Mei-yuMei-yu QPFs is also improved, and the TSs for the largest 4% of events with high potential for impacts on days 1–3 are around 0.95 at 0.05 mm and stay ≥ 0.2 up to 200 mm on days 1–2 and 150 mm on day 3, respectively. At lead timesLead time beyond the short rangeShort range, the time-lagged approach using the cloud-resolving modelCloud-Resolving Model (CRM) has been demonstrated to be a feasible and effective method for ensemble with high-quality QPFs for typhoons. At the longer range with high forecast uncertainty, it produces various realistic scenarios associated with different storm tracks, among which the worst-case scenario is very useful for early preparation. This scenario tends to be produced earlier than five days before landfall. As the typhoon approaches and the uncertainty decreases, the predicted tracks converge toward the best track, and the most-likely scenario emerges for the authorities to make proper adjustment. The above results are likely applicable to many regions in East Asia due to similar characteristics in rainfall and topography to Taiwan. In the future, continuous improvement will take place toward more members with a cloud-resolving capability, and new methods from artificial intelligenceArtificial intelligenceand machine learningMachine learning can also contribute and help make better QPFs and model forecasts in general.