颱風降雨是臺灣重要的水資源，卻也是主要的致災因素。因此，颱風預報的良莠與改善都十分重要。自發展數值天氣預報以來，不論是單一決定性預報或多成員系集預報，一個難解的問題，是個別預報均有其不確定性，因此該預報情境發生的機率高低，在事前無法得知、或至少無從確定。因此，隨著人工智慧的發展，本研究建構與測試一機器學習模型，在事前提供客觀預報指引，以幫助吾人判別每個颱風預報降水情境的可信度，藉以改善預報。確切而言，本研究使用2.5－km雲解析風暴模式，對10個西行準侵臺颱風每6h的八天差時系集預報結果，選取共105個預報參數，以機器學習模型針對預報颱風在影響期間（中心距離臺灣陸地在300 km以內）總累積降雨的相似性技術得分（Similarity Skill Score，簡稱SSS）進行預估。此得分由本文定義與使用，其值的高低與該雨量預報的可信度（即參考價值）成正比。所有評估的預報，其初始時間的颱風中心均尚未靠近臺灣到300 km以內，因此期限多在短期預報以外（＞72 h）。由此10個颱風個案的評估結果顯示，在大多數的情況下，機器學習對逐次預報所預估的SSS值，的確可以適當掌握未來真實SSS的上升下降趨勢，亦即提早告訴吾人，那些預報的可信度較高、那些可信度較低，而在事前提供有效的客觀預報指引。在本研究裡，當預估SSS的50百分位數達0.6以上時，其實際SSS有71％也大於0.6，而颱風行進方向的修正也有71％是正確的。特別對在前置時間長、不確定性高的預報初期，提高其參考價值。但是，因機器學習可視為複雜的統計方法，當某個颱風的行為與大部分輸入訓練的資料相左時，其效果也會受到限制，對此，本文提出了幾個可能的改進方向。

Typhoon rainfall is both an important water resource and potential disaster in Taiwan, so its forecast quality and improvement are important. An issue of all numerical weather predictions, regardless deterministic or ensemble, is that whether its scenario will occur or its exact probability, is not known in advance. Nowadays, this issue may be solved through artificial intelligence. In this study, therefore, we have developed and tested a model through machine learning that provides objective guidance to indicate the credibility of each quantitative precipitation forecasts (QPFs) for typhoons once it is made and thus help improve forecasts. Specifically, time-lagged forecasts (out to 8 days) every 6 h for 10 westward-moving typhoons affecting Taiwan by the 2.5-km Cloud-Resolving Storm Simulator (CReSS) are used. A total of 105 parameters are selected from each forecast and data from nine typhoons are fed into the learning model to, after training, predict the similarity skill score (SSS) of total accumulated rainfall during the period when the storm moves within 300 km from Taiwan in each of the forecasts for the tenth typhoon. As a measure to the overall quality of the QPFs, the predicted SSS thus serves as guidance for forecast credibility. At the initial time of forecasts included, the typhoon center is still at least 300 km away, so many are at ranges beyond the short range (≥72 h). Results from these 10 cases indicate that the machine learning model can capture the tendency of the actual SSS (computed using observed rainfall) for most cases, thereby informing the forecasters which QPFs are more trustworthy and which other ones are less so before the event. Such guidance is especially valuable at longer lead time, when the forecast uncertainty is relatively high, and thus our results are highly encouraging. Nevertheless, as machine learning can be viewed as a complicated statistical technique, when certain typhoon behaves differently from those that serve as the training data, the outcome would be less useful. Some possible directions for further improvement are also offered and discussed.