Please use this identifier to cite or link to this item:
Nearshore wave height hindcasting at an arbitrary point by using a combined numerical-ANN model during typhoons
|Authors: ||Hsieh, Chia-Jung|
|Contributors: ||NTOU:Department of Marine Environmental Informatics|
typhoon wave height;artificial neural network;numerical mode
|Issue Date: ||2020-01-20T06:20:28Z
|Abstract: ||臺灣位於太平洋熱帶及副熱帶的交界處，身處大洋與高溫的條件下，在夏、秋兩季經常有颱風來訪，臺灣東北部往往首當其衝，迎來最強大的風雨，加上多為平原地形不受阻擋，颱風可以肆無忌憚地侵襲，且臺灣東北部不如西部有中央山脈的阻隔，起風降雨時間較早，往往成為颱風來襲的第一道防線。而臺灣目前設立的海上浮標測站有限，無法確切得知無浮標區域的波浪高度值，欲求得任意點的颱風期間波高，為此加入網格化數值模式模擬的波高結果做為預測的輔助資料。數值模式的發展已經非常進步，預測颱風時期的波浪高度也趨於精準，但缺點是數值模擬所需的時間較長，無法做短時間的即時預測。為此，本研究設計三階段的研究流程，發展一套類神經統計結合數值模式的方法，能同時具有數值模擬的準確性以及類神經模擬的效率，以達到即時預測颱風侵襲時臺灣東北部任意點波浪高度的目的。本研究資料範圍為2005年至2015年，第一階段利用機器學習的方法建立颱風警報時期波浪預測模式，此階段收集中央氣象局的氣象站氣候資料、浮標觀測資料以及颱風警報資料，建立類神經網路(Aritficial Neural Network, ANN)統計模式；第二階段使用發展較成熟的SWAN(Simulating Waves Nearshore)數值模式，進行颱風波浪的模擬，利用大範圍的計算結果輸出小範圍波浪邊界條件，再進行小範圍的計算，模擬颱風波浪於近岸處的分佈情形；第三階段進行任意點颱風波高預測，假設沒有實測資料供類神經網路模式的學習目標值，使用 SWAN數值模式的波高模擬結果，發展同時具有SWAN數值模式利用物理意義模擬颱風期間波浪高度與類神經網路方法快速運算的模式，以達到預測任意點颱風波高的目的。研究結果顯示：(1)龍洞區域的預測誤差較龜山島區域小且相關性較高，可能原因為龜山島浮標所屬位置受地形影響，使波浪高度較高，因此容易造成預測低估的現象，使整體誤差較大；(2)將波浪高度小於2 m定義為小浪；波浪高度介於2 m至3 m間定義為中浪；波浪高度3 m以上定義為大浪，龍洞區域小浪及大浪的預測較佳；龜山島區域小浪預測較佳；(3)比較三種預測模式即類神經網路模式、SWAN數值模式以及混合模式，類神經與實測值的整體相關性較大且誤差較小，其次為SWAN，最差的為混合模式，由此可知使用類神經網路方法的模式能較準確預測下一小時的波浪高度，而SWAN模式本身的誤差將會影響後續混合模式的準確性。|
Taiwan is located at the junction of tropical and subtropical Pacific Ocean. Typhoons often occur in summer and autumn under the conditions of ocean and high temperature. The northeastern of Taiwan often bears the brunt and comes the most powerful storms. Because there is no mountain barrier in northeastern of Taiwan like western, typhoons can invade with impunity. However, the current buoys are seldom in Northeastern Taiwan and the information of wave height without buoys is limited. Thus, the wave heights of typhoons at an arbitrary point are determined difficultly. Therefore, adding the wave height simulated by the numerical model as a supplementary data for prediction. The development of the numerical model has been very advanced. The prediction of the wave height during the typhoons also tend to be accurate. However, the disadvantage is that the numerical simulation requires a long time and it can’t be used for real-time prediction. This study designed a three-stage research process and developed a method of artificial neural network (ANN) combined with numerical model that can simultaneously have the numerical simulation accuracy and the efficiency of ANN to achieve real-time wave height forecasting of an arbitrary point at northeastern of Taiwan in typhoon period. The collected data are from 2005 to 2015, including the weather data of the weather station of the Central Meteorological Bureau, the buoy observation data and the typhoon warning information. In the first stage, we used the machine learning methods to establish the wave forecasting model during the typhoon periods. In the second stage, we used the SWAN numerical model to simulate typhoon waves. The SWAN model used a wide range of calculations as boundary conditions, and then did small-scale calculations. Then, the SWAN model is used to simulate the typhoon waves of the nearshore. In the third stage, we predicted the typhoon wave height at an arbitrary point. Assuming that there is no measured data for the learning target of the ANN model. The wave height results simulated by SWAN model were used to develop the combined ANN with SWAN models. The merits of the combined model are having physical meaning of SWAN numerical model and calculating ANN rapidly. The findings revealed that: (1) The prediction error of Longdong area was smaller and had a higher correlation than that of Guishan Island. (2) Wave height less than 2 m was defined as small wave; Wave height between 2 m and 3 m was defined as medium wave; Wave height above 3 m was defined as large wave. Prediction of small wave and large wave in the area of Longdong area were better. Prediction of small wave in the area of Guishan Island were better. (3) The overall correlation of ANN model between the predicted value and the measured value was relatively large and the RMSE was small, followed by SWAN model, and the worst is the coupled model.
|Appears in Collections:||[海洋環境資訊系] 博碩士論文|
All items in NTOUR are protected by copyright, with all rights reserved.