類神經網路應用於光纖雷射與聲波杜卜勒測速儀渠流速度剖面量測之研究

Abstract

研究複雜水流流場的水力特性首重平均流速剖面之研析，因而，本論文透過於臺北自來水事業處直潭淨水場內流量量測資料蒐集，以及楊（1998）的實驗成果，試圖以具有強大學習能力及歸納推演的類神經網路（Artificial Neural Networks，簡稱ANN），應用於所蒐集的實驗量測資料，進行平均流速剖面的模擬分析與推估比對。以期對複雜水流紊動現象能有進一步的掌握，從而能更精確的推判輸水設備之流量及大幅減輕日後相同檢測實驗上所需花費的經費。 直潭淨水場內淨水設備的檢測實驗係採用聲波杜卜勒測速儀（Acoustic Doppler Velocimeter，簡稱ADV），進行八組不同定量流之平均速度資料蒐集，每組流量於量測斷面共測取八條不同位置之平均速度剖面，每條剖面量取十個不同位置之速度值，全部實驗共含640個平均速度值。實驗寬深比介於1.388與1.438之間；渠道為混凝土邊壁，且底床坡度固定為0.1%；雷諾數涵蓋4000005與800000之範圍；福祿數則介於0.068與0.118，所有實驗流況均屬於亞臨界流況（subcritical flow）。楊（1998）的實驗則使用光纖雷射杜卜勒測速儀（Fiber-optic Laser Doppler Velocimetry，簡稱FLDV）進行陡坡光滑床面的明渠流流場特性之研究。實驗寬深比介於3.79與11.36之間；光滑底床坡度變化分別為0.3%、1%及2%；雷諾數涵蓋10000與80000之範圍；福祿數則介於0.96與2.72，大部份實驗流況屬於超臨界流況（supercritical flow）。 研究中，類神經網路選擇最具代表性與應用最為廣泛之倒傳遞網路（Back-Propagation Network）進行平均速度剖面模式之建構與應用。倒傳遞神經網路則使用三層架構之網路，分別為輸入層、隱藏層以及輸出層。輸入項主要有寬深比、坡度及速度剖面量測位置等；輸出層為平均速度值。模式建構之過程，將資料分為訓練、驗證與測試三階段，以分別獲得最佳訓練模擬與推估的情形。研究結果顯示，倒傳遞神經網路對於模擬與推估陡坡光滑實驗水槽及淨水場實驗輸水幹渠之平均速度剖面，皆具有相當良好之成效。由實驗之成果可知，應用類神經網路模式，可有效降低日後實驗檢測之花費，成果值得推廣。除此之外，倒傳遞神經網路所模擬與推估之平均速度剖面與斷面積相乘後累加，可推求相對應的流量，進而建立水深-流量關係曲線，日後藉由水深的量測即可迅速精確得知渠道之流水量。此結果對於灌溉排水等固定渠道之流量量測，提供一便利且精確之方法。

The Artificial Neural Network (ANN) has great capability for solving various complex problems, such as function approximation. The main objective of this study was to evaluate the applicability of the ANN for simulating and estimating mean velocity profiles and the discharges accordingly. The data obtained from the open channel of the Chihtan purification plant, Taipei and experimental on steep open channel flow (Yang, 1998) were used to train and verify the proposed ANN. An Acoustic Doppler Velocimeter (ADV) was adopted in this experimental study to measure the mean velocity profiles in the open channel of the Chihtan purification plant, Taipei, at fixed measuring section with eight different discharges and ten different depths. The total number of experimental data sets was 640. In this study, the aspect ratio varied from 1.388 to 1.438, the Reynolds number varied from 400000 to 800000, and the channel bed slope of 0.1% was selected. The Froude number ranged from 0.068 to 0.118, and all of the experiments belonged to subcritical flow. Yang (1998) used the Fiber-optic Laser Doppler Velocimetry (FLDV) to investigate the characteristics of the steep open channel flow over a smooth boundary, that the aspect ratio varied from 3.79 to 11.36, the Reynolds number varied from 10000 to 80000, and the channel bed slope of 0.3%, 1% and 2% were selected. The Froude number ranged from 0.96 to 2.72, and most of the experiments belonged to supercritical flow. The backpropagation (BP) algorithm was applied to construct the neural network. The structure of the BP neural network included input, hidden, and output layers. The input layer contained B/H, S, and Z, while the output layer has only one node representing the velocity value in the specific location of y/H on the vertical. The experimental data were split into three sub-sets, training, validation and testing sets to train and to verify the built ANN. The results demonstrated that the constructed neural network models could be embedded as a module for estimating or generating the profiles of mean velocity for turbulent open channel flows. The model could be used as a powerful tool to simulate and estimate the flow profiles for the similar flow conditions to reduce the cost of the experimental work. The trained model could also be used to provide the flow profiles for missing data, and estimated the discharge for a given specific depth.