影響暴潮水位變化之相關因素眾多,颱風暴潮與一般規律升降之潮汐運動不同,且影響因素與水位變化間存在著複雜之非線性關係,因此不論以物理模式或經驗公式來推估颱風暴潮皆存在相當之困難度。基於類神經網路對於非線性函數關係有不錯之適用性,本文應用類神經網路法與時間序列分析法,利用颱風資料與當地氣象資料建立颱風暴潮預測模式。 本研究使用5個不同演算法之倒傳遞類神經網路建立颱風暴潮預測模式,利用過去颱風特性因子資料與潮位測站氣象資料,推估未來颱風期間潮位測站之颱風暴潮。本文以宜蘭梗枋測站過去16場颱風資料進行模擬訓練,並使用5種不同之倒傳遞類神經網路演算法進行預測模式比較分析,再以4場颱風資料進行驗證,作為模式評選依據,用以評選較佳之演算法建立預測模式。依綜合評估結果顯示,以Levenberg-Marquardt演算法建立之預測模式較佳。 在時間序列分析法中,因颱風暴潮本身就是時間序列資料,受到前一時刻甚至前幾小時之暴潮影響甚大。本研究使用單變量時間序列法來建立颱風暴潮預測模式,利用自我相關函數及偏自我相關函數等與ARIMA(p,d,q)之參數組合,經由測試得出ARIMA(2,1,2)模式之預測效果最好。 類神經網路預測模式與單變量時間序列預測模式之比較,預測模式隨著預測時間增加預測值誤差也隨之增大。然而,在未來一小時預測上,單變量時間序列優於類神經網路,而在未來二小時及未來三小時預測上,類神經網路預測效果優於單變量時間序列。 單變量時間序列模式,在短時間之預測結果較佳,但其模式預測穩定性不佳,就單變量時間序列分析而言,是根據時序資料本身的變化規律預測未來變化。因此單變量時間序列在短時間之預測上雖優於類神經網路模式,但其預測值誤差隨預測時間增大,預測水準大幅下降,顯示單變量時間序列模式預測穩定性低於類神經網路模式。
Typhoon surge is different from tide phenomenon, because it has a non-linear characteristic and more complexes impact factors. It is hard to predict by using traditional numerical analysis and experience formula. In this study we applied artificial neural networks and time series analysis with to establish the typhoon surge prediction model. The study uses five different algorithms to build typhoon surge forecasting model with the application of artificial neural networks. The data including typhoon’s characteristics and meteorological conditions at the tidal station were used as the input data to the model for the forecasting of typhoon surges in the next few hours. Twenty typhoon surge data at Geng-fang tidal station were collected, sixteen of them were used in model’s calibration and the others were used in model’s verification. A general evaluation index was used to evaluate five different algorithm of typhoon surge forecasting model. The results show that: the model developed with Levenberg-Marquardt algorithm was the best of five different algorithms. In the time series analysis, typhoon surge itself is time-series data; it may influenced by typhoon surge one hour or several hours ago. In this study aims to use autoregressive moving average model modeling typhoon surge forecasting. In order to evaluate model’s parameters, the study used autocorrelation function、partial autocorrelation function and try and error to test model's parameters combination. After try and error test the results indicated that ARIMA (2, 1, 2) predicted performance was best than other parameters combination test. In comparison typhoon surge forecasting model between artificial neural networks and single variable time series model, as observed the time step increase the prediction error increase. However, in predict of next one hour typhoon surge, the results indicated that single variable time series model was better than artificial neural networks of Levenberg-Marquardt algorithm. But in predict of next two and three hour typhoon surge, as observed typhoon surge forecasting model of artificial neural networks of Levenberg-Marquardt algorithm was better than single variable time series model. In the prediction of next one hour typhoon surge forecasting, single variable time series model simulated was the best. When it comes to the stability of typhoon surge forecasting model, single variable time series wasn’t as good as artificial neural networks. In the time series analysis, typhoon surge forecasting model according to variation of time series data to predict future typhoon surge. Although single variable time series model better than artificial neural networks in the next one hour prediction of typhoon surge, but with time step increase prediction ability drop down. The results indicated that the stability of single variable time series model worse than artificial neural networks of Levenberg-Marquardt algorithm model.