倒傳遞類神經網路應用於台灣北部水庫懸浮固體濃度即時分析與預測之研究

Application of Back Propagation Artificial Neural Network to Real Time Analysis and Prediction of the Total Suspended Solids in Northern Taiwan Reservoirs

張郁麟(Yu-Lin Chang)

指導教授 ： 范正成

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2007年)

https://doi.org/10.6342/NTU.2007.00160

摘要

水庫集水區的治理、開發與操作，常會遭遇地表土壤沖蝕所產生的非點源污染。為了能夠有效防止此類災害的發生，隨時監測集水區的整治情況，以及建立完備的懸浮固體濃度即時監測系統是必要的。本研究以中華民國行政院環境保護署新山水庫、翡翠水庫、石門水庫、寶山水庫、永和山水庫、明德水庫水質監測數據查詢資料庫中1993-2005年間的資料來進行分析。從資料庫中所選許的水質參數有比導電度、溶氧、酸鹼值、濁度、溫度、採樣月份、葉綠素α、總磷、總硬度及透明度。然後利用水質之間的群集分析、測站之間的顯著性分析、水庫之間的相關性分析，進一步選取合適的水質參數和測站。再利用類神經網路架構來進行訓練、驗證網路即時推估懸浮固體濃度。經過一系列分析及觀察，發現類神經網路可以由水質參數推估懸浮固體濃度，但其推估的準確度依地理位置及土壤分布的不同而有所改變。結果亦顯示以類神經網路模式在一些條件下可利用數種容易量測的水質資料來推估不易量測的懸浮固體濃度。此外，以石門水庫水質資料利用倒傳遞類神經網路來預測懸浮固體濃度並作驗證，其結果顯示，預測與實測值的迴歸式係數達到0.90，表示推估趨勢十分良好；且網路輸出與期望輸出的判別係數R2達到0.63，顯示以本研究所提出之方法和石門水庫各項水質參數應用在其懸浮固體濃度之推估上，可預測到各個峰值，且可相當準確的預估其變化趨勢。

關鍵字

倒傳遞類神經網路倒傳遞類神經網路 ； 懸浮固體濃度 ； 非點源污染 ； 水質監測

並列摘要

In the management of reservoir, non-point source pollutions caused by surface soil erosion are frequently encountered. In order to prevent this kind of problems, it is necessary to continually monitor the watershed of the reservoir as well as to real-time monitor the total suspended solid(TSS). The data of the water quality of Xin-Shan reservoir, Feitsui reservoir, Shimen reservoir, Baoshan reservoir, Yonghe-Shan reservoir, and Mingd reservoir used in the study were provided by Environmental Protection Administration of the Executive Yuan, R.O.C.. These data included electrical conductivity, dissolved oxygen, pH value, turbidity, temperature, month, chlorophyll-α, total phosphorus, total hardness, and transmissivity, in the period from 1993 to 2005. Suitable water quality parameters and observation stations were further chosen from the statistical results by cluster analysis of the water quality, dominance analysis of the observation stations, and correlation coefficient of the reservoirs. Back propagation artificial neural network was applied to real time analysis and prediction of the total suspended solids. However the estimation accuracy would vary with locations and soil types. From the results, it was also found that the nural network model may be used to estimate the concentration of suspended solids, which is difficult to be real time measured, by using several parameters of water quality, which are easier to be measured, under some specific conditions. When back propagation network was modified to predict the real time total suspended solids in Shimen reservoir, the results showed that the predicted variation tendency of total suspended solids in network output agrees well with that in expected output, the R2 can reach 0.63, the regression coefficient can reach 0.90. It could be concluded that the method of back propagation artificial neural network and water quality can be used to rapidly and accurately estimate TSS.

並列關鍵字

back propagation network ； total suspended solids ； non-point source pollution ； water quality monitoring

參考文獻

1.白子易、江舟峰、蔡嘉和、蔡曜聲、朱校興、廖婉君(2002)，「以類神經網路預測三種型態污水處理廠出流水水質之研究」，朝陽學報，82:253-267。

3.行政院環境保護署(2002)，「台灣地區環境水質監測年報」，pp.28-31。

10.范正成、張郁麟、楊文仁、劉哲欣(2006)，「倒傳遞類神經網路應用於石門水庫懸浮固體濃度之即時分析與預測」，中華水土保持學報，37(4):367-376。

14.楊文仁、范正成(2003)，「運用類神經網路建立紋溝間土壤沖蝕推估模式之研究，中華水土保持學報」，34(3)：271-279。

19.Baruah, P.J., M. Tamura, K. Oki, and H. Nishimura (2002), “Neural Network Modeling of Surface Chlorophyll and Sediment Content in Inland Water from Landsat Thematic Mapper Imagery Using Multidate Spectrometer Data,” [w:] Gilbert, G.D., Frouin, R.J. (Eds.), Ocean Optics: Remote Sensing and Underwater Imagery, Proceedings of SPIE the International Society for Optical Engineering, Seattle, 4488:205-212.