Projecting forecast quality before events through machine learning: Preliminary results of cloud-resolving quantitative precipitation forecasts in Taiwan for westbound typhoons

A long-standing problem of all numerical weather predictions, regardless deterministic or ensemble, is the more accurate assessment in probability (or likelihood) for the predicted scenario to occur, especially at longer lead times due to typically larger errors. The rapid development of artificial intelligence today may offer an effective method to tackle this issue. In this study, a neural-network machine-learning model is developed to, after training, project the expected value of the similarity skill score (SSS) of predicted total rainfall distribution in Taiwan for westward-moving typhoons during their influence period, thus serving as an objective guidance for the quality of the prediction. Ten typhoons are included, and a total of 105 parameters linked to rainfall are used from time-lagged forecasts (out to 8 days) every 6 h by a cloud-resolving model, when they cover the entire influence period (inside 300 km from Taiwan) with enough lead time. For each typhoon, only data from the other nine cases are used to train the model.

The results indicate that machine learning can capture the tendency of the actual SSS (calculated against observed rainfall) for most cases (eight out of ten), thereby informing the forecasters which quantitative precipitation forecasts (QPFs) are more trustworthy and which other ones are less so beforehand. Such guidance is particularly valuable at longer lead times, when the forecast uncertainty is relatively high. Thus, our results are highly encouraging. Nevertheless, if a typhoon behaves differently in forecasts from those that serve as the training data, the outcome would be less useful. Possible directions to remedy this issue and make further improvement are also offered.