利用類神經網路於潮汐長期預報之研究

Abstract

本文引入潮汐理論當為類神經網路於長時期潮汐預報之輸入參數之一，以提升預測精度，同時測試較佳的網路架構，測試項目包括學習速率與慣性因子、輸入變數組合、隱藏層神經元個數、學習次數以及學習時數。找出三種潮型各一個測站的最佳預測模式；半日潮新竹測站最佳網路架構為I5H6O1、混合潮後壁湖測站為I5H12O1而全日潮鼻頭角測站為I5H12O1，輸入層輸入變數皆為輸入前兩個及前一個實測潮位與前兩個、前一個和當時之理論潮位，隱藏層神經元個數分別為6、12、12個，最佳學習次數為500、1000、2000次，最佳學習時數均為720小時，學習速率與慣性因子皆設定為0.1及0.8。 進行長時期潮位預測時，季節性的潮位變化會降低預測精度，本文利用360小時的移動平均法修正實測值再進行預測，結果顯示，新竹測站的精度提升5.7﹪，後壁湖測站提升34.3﹪，鼻頭角測站則提升45.9﹪。

That tidal levels obtained from tidal theory are added as inputs in artificial neural network model is found to improve prediction ability for tidal levels in this paper. The optimum structure of the present artificial neural network model for each station is set up from examining the learning rate, moment factor, input parameters, numbers of hidden layer, learning times and input length. The optimum ANN models for three kinds of tidal types also have five inputs that are two observed tidal levels and three theoretical tidal levels and have learning rate of 0.1 and moment factor of 0.8, respectively. The optimum model for semi-diurnal type at Hsian-Chu station is I5H6O1 with 500 learning times. The optimum model for both mixed type at Hou-pi-hu station and full diurnal type at Pi-tou-chiau station is I5H12O1. The observed tidal data have seasonal deviation from mean water level because of temperature and are deseasonalized by moving Gaussian average with a length of 360 hours. The ANN models have better long-term forecasting for deseasonalized tidal data.