結合潮汐理論與類神經網路在潮汐預報上之研究

Abstract

本文以引潮理論加入類神經網路來預測時序列的潮汐水位。若僅輸入短期潮位資料，純以神經網路的基本潮位預報模式會造成大潮的高低潮位學習不到的現象發生，本文提出在基本潮汐預報模式加入二維與三維的潮汐理論，使神經網路有大小潮變化的訓練與學習的機會。 由本文針對各類潮型的潮汐預報結果，則加入引潮理論的模式均較基本的模式有較佳的預報精度，尤其加入三維理論的模式最優良。若給予神經網路訓練的資料為小潮或日潮不等現象不嚴重時，加入三維理論的模式較基本網路模式在每時刻預報的平均誤差可降低約20%以上，高低潮點的峰值誤差可降低約40%以上。若訓練的資料為大潮時，則三維理論的模式較基本網路模式的平均誤差約降低5%~15%，峰值誤差可降低約10%~30%左右。

It is a good way to use artificial neural network (ANN) to predict hourly tidal level, especially when tidal of only few days are available to be input. However, a simple ANN model still has a drawback that the simple ANN model are poor to predict high and low tidal level when short-time neap tidal data are input. Therefore, we propose a complemental input of tidal envelope obtained from 2D or 3D tidal theory into the simple ANN model to promote its prediction ability. Five sets of tidal data of three kinds of different tidal type were collected from five stations. Comparisons of predicted peak error and mean error of three models show that the ANN model combined with 3D tidal theory is the best to predict tidal levels for all chosen data and that the simple ANN model is the poorest model. When hourly neap tidal data of one day of Penghu station are input in three models, hybrid ANN model with 3D tidal theory has smaller mean error than the simple ANN model by 20% in general and reduces peak error by 40% from the simple ANN model. It also proven that the ANN with 3D tidal theory is prior to harmonic analysis method to predict tidal levels when fifteen-day data were used.