以深度學習神經網路推估西北太平洋颱風強度

Estimating Typhoon Intensity in the Northwest Pacific Using Deep Neural Networks

林旭信(Shiu-Shin Lin) ； 溫慶亞( Qing-Ya Wen) ； 朱楷洋(Kai-Yang Zhu) ； 何御東(Yu-Dong Ho)

《農業工程學報》 69卷1期 (2023/03) Pp. 29-40

https://doi.org/10.29974/JTAE.202303_69(1).0003

摘要

本研究以深度學習神經網路（Deep Neural Network, DNN）架構為基礎預報西北太平洋颱風強度，建立未來五天每十二小時之DNN颱風強度預報模式，並以多元線性迴歸（Multiple Linear Regression, MLR）建立MLR颱風強度預報基準模式，分析比較颱風強度預報之效能改進。以西北太平洋為研究區域，蒐集西元2000年～2012年總共13年西北太平洋颱風之SHIPS資料（SHIPS Developmental Data），利用原始SHIPS資料之大氣海洋環境因子，分別為DNN颱風強度預報模式與MLR颱風強度預報基準模式之輸入，以RMSE（Root-Mean-Square Error）為模式評估指標建立DNN與MLR模式。颱風強度預報結果與桑達颱風個案研究結果顯示：（1）DNN模式相較於MLR模式之改進百分比皆有提升，尤其提前時刻超過84小時，改進百分比逐漸增加，其最大改進提升幅度甚至近10%，所有提前時刻皆以DNN模式表現較佳，平均改進百分比有4.53%，（2）MLR模式低估與高估情況均高於DNN模式，（3）桑達颱風個案研究顯示，提前時刻12小時到48小時DNN模式改進百分比變異性較大，提前時刻24小時與48小時，DNN模式改進百分比相對最大。整體而言，DNN模式對於颱風強度預報優於MLR模式。

關鍵字

西北太平洋 ； 颱風強度預報 ； 機器學習 ； 深度學習神經網路 ； 多變量線性迴歸

並列摘要

Based on the Deep Learning Neural Network (DNN) architecture, this study estimates the typhoon intensity of the Northwest Pacific Ocean, establishes a DNN typhoon intensity forecast model every 12 hours for the next five days, and uses multiple linear regression (Multiple Linear Regression), MLR) to establish the MLR typhoon intensity forecast benchmark model, to analyze and compare the efficiency improvement of typhoon intensity forecast. Taking the Northwest Pacific as the research area, collect the SHIPS data (SHIPS Developmental Data) of the Northwest Pacific typhoon from 2000 to 2012, and use the atmospheric and marine environmental factors of the original SHIPS data as the DNN typhoon intensity forecast model and the MLR typhoon respectively. The input of the intensity prediction benchmark model, the DNN and MLR models are established with RMSE (Root-Mean-Square Error) as the model evaluation index. The results of the typhoon intensity forecast and the case study of typhoon Sanda show that: (1) Compared with the MLR model, the improvement percentage of the DNN model has increased, especially when the lead time is more than 84 hours, the improvement percentage has gradually increased, and the maximum improvement has even increased by nearly 10 %, the DNN model performed better in all advance times, with an average improvement percentage of 4.53%. (2) The underestimation and overestimation of the MLR model were higher than those of the DNN model. (3) The case study of typhoon Sanda showed that the advance time from 12 hours to 48 hours, the DNN mode improvement percentage has a large variability. The advance time is 24 hours and 48 hours, and the DNN mode improvement percentage is relatively the largest. Overall, the DNN model is better than the MLR model for typhoon intensity forecast.

並列關鍵字

Pacific northwest ； Typhoon intensity forecast ； Machine learning ； Deep learning neural network ； Multiple linear regression

參考文獻

DeMaria, M. & Kaplan, J. (1999). An Updated Statistical Hurricane Intensity Prediction Scheme (SHIPS) for the Atlantic and Eastern North Pacific Basins. Weather and Forecasting, 14, pp. 326-337.

DeMaria, M., Mainelli, M., Shay, L. K., Knaff, J. A. & Kaplan, J. (2005). Further Improvements to the Statistical Hurricane Intensity Prediction Scheme (SHIPS). Weather and Forecasting, 20, pp. 531-543.

DeMaria, M., Sampson, C. R., Knaff, J. A. & Musgrave, K. D. (2014). Is Tropical Cyclone Intensity Guidance Improving? Bulletin of the American Meteorological Society, 95, pp. 387-398.

Duchi, J., Hazan, E. & Singer, Y. (2011). Adaptive Subgradient Methods for Online Learning and Stochastic Optimization. Journal of Machine Learning Research 12.

Jarvinen, B. R. & Neumann, C. J. (1979). Statistical Forecasts of Tropical Cyclone Intensity for the North Atlantic Basin.