應用類神經網路進行台灣地區颱風系集降雨機率預報校正

颱風定量降雨預報是氣象防災的重中之重，建立在系集預報所提供的大數據基礎上，利用類神經網路技術建立颱風定量降雨預報校正模型，將是一個值得期待的趨勢。然而，由於雨量特殊的統計特性以及颱風個案稀少，因而使得應用類神經網路技術建立預報校正模型時將面臨相當的挑戰。本研究旨在探討類神經網路方法應用在台灣地區颱風雨量預報校正的可行性，以及在不同颱風雨量下的預報校正成效；特別是在雨量極大時，類神經網路校正模型是否能改善原始系集降雨機率預報。以2015年蘇迪勒颱風為例，本研究建立3個淺層類神經網路與1個深層類神經網路，進行颱風未來24小時累積降雨預報校正，評比結果顯示淺層類神經網路有較佳的校正成效。本研究同時考量颱風快速變化的特性，使用非均等權重方法來建構模型，結果顯示非均等權重方法比均等權重方法所建立之淺層類神經網路模型有較佳的校正成效。本研究也將結果應用到2013-2015年間侵台的其他11個颱風，同樣發現非均等權重方法的淺層類神經網路能校正颱風原始系集降雨機率預報。

關鍵字

校正；系集預報；類神經網路；定量降雨機率預報；颱風雨量

並列摘要

Focusing on typhoon precipitation, this study adopts the neural network (NN) technique to calibrate the bias of 0-24-h probabilistic quantitative precipitation forecasts (PQPF) over Taiwan produced by the WRF ensemble prediction system (WEPS) developed by the Central Weather Bureau. The goals are to explore whether the NN post-processing technique is suitable for the calibration of typhoon precipitation forecasts and to compare the calibration effects among different structures of NN models. Four NN models (three are shallow NNs and one is a deep NN) are developed and applied to calibrate the precipitation forecasts of Typhoons Soudelor, which brought torrential rain during the period of its passing through Taiwan. Evaluation results indicate that the post-processing procedures improve the performance of calibration of ensemble precipitation forecasts, especially for heavy precipitation. The simple shallow NNs perform better than the deep learning NN model. In addition, we compare the forecast performance of the NN models when equal temporal weights (unadjusted) or unequal weights are used for the training data. The results with unequal temporal weights have lower forecast errors than the model with equal weights. This model is then applied to other typhoons that attacked Taiwan during 1013-2015, and the finding remains that the unequal weights of simple shallow NN can improve the calibration. However, it is noted that the NN models in this study can correct only the values of forecast probability, not the precipitation pattern.

並列關鍵字

calibration ； ensemble forecast ； neural network ； probabilistic quantitative precipitation forecasts ； typhoon precipitation

主題瀏覽