- 學位論文
考量不同颱風路徑之類神經網路暴潮預測模式
Typhoon Surge Forecasting by Artificial Neural Networks in consideration of different tracks
摘要
影響颱風暴潮的眾多因素都呈現高度的非線性關係，而類神經網路則具有處理非線性能力。因此本研究以實測颱風及海象資料作為輸入資料，藉由類神經網路非線性最佳化及學習演算能力，建立颱風暴潮的預測模式。 本文以宜蘭縣梗枋測站為研究對象，收集1993到2007年間第一路徑到第六路徑共42場颱風資料，並分析颱風特性間之關係；影響颱風期間暴潮偏差量之因素眾多，其中又以颱風至測站之距離與其關係最為密切，因此相同颱風規模下而因其侵台路徑不同，形成之暴潮偏差也會有所不同，第四路徑與第五路徑之颱風，由於距離梗枋站較遠，因此不論在中颱與輕颱其形成的暴潮偏差量，都低於其他四個路徑。 而訓練樣本數的部分藉由本文測試結果可得知，訓練樣本數對於類神網路的訓練有相當影響，訓練樣本在4場以下的話，由於訓練樣本數不足，推估上的精準度則相當低，5至11場時訓練樣本較足夠，推估能力以相當優異但準確度仍有起伏，而訓練樣本數15至38場時訓練樣本相當足夠，且具有高度的推估能力。 本文模式與前人不考量路徑所建立出之模式相比，分路徑後使樣本數減少因而使第二、五路徑網路訓練不足，而使建立出的第二、第五路徑預測模式的精準度與前人相比並未提升；而第三路徑訓練樣本數較足夠，因此第三路徑預測模式準確度較前人優異；第六路徑的訓練樣本數也較足夠，但由於選取之驗證颱風並非標準該路徑之颱風，因此第六路徑預測模式的準確度低於前人的預測模式。
並列摘要
The phenomenon of storm surge shows highly nonlinear characteristic. The physical numerical models or empirical formulas are accordingly inapplicable in estimating the deviation of the typhoon surge. Neural network, on the other hand, has been proved to be an appropriate method in solving nonlinear problems. In the present study, a typhoon-surge forecasting model was developed with a back-propagation neural network (BPN) method. The typhoon’s properties, local meteorological conditions and typhoon surges were used as input data of the model to forecast typhoon surges at the following time. The 42 typhoon events from 1993 to 2007 were used to analyze different charateristics of typhoon surges, Among numbers of factors influecing surge storm value during typhoon period, the distance from typhoon to the tidal station was found the most significant one. That is to say, given that identical scales of typhoons, the value of storm surge may vary substantially due to diverse tracks. In contrast to other typhoons, those that followed fourth and fifth track were relatively distant from Geng-fang tidal station, and therefore caused lower surge storm value whether the scale of typhoon was weak or medium-strength. Meanwhile, we also conclude in this study that the numbers of training samples have considerable influences on the training of neural network. Once the training sample numbers are less than 4, which is considered a threshold number, inaccurate estimators might therefor be resulted; 5 to 11 training samples may lead to relatively accurate, but still fluctuate, outcomes; when the numbers increase to 15 to 38, favorable estimators with highly explanatory and predicting power can be expected. Comparing to previous storm surge prediction models, unlike their ignorance of tracks factors of typhoons, our model includes these differences but in the meantime decreases numbers of training samples. Because of insufficiency in training samples in second and fifth track, no marked contrast have we discovered between estimators derived from previous models and ours; the training samples are comparatively sufficient for the third track and therefore more accurate estimators are produced; there are enough training samples for sixth track, however, the selected subjects are not routed on standardized track, and hence the estimators are inferior to those of previous models.