運用倒傳遞類神經網路於河川定點水位預報

Use BPN to predict river’s water level of the fixing point

鮑繼勇(Ji-Yung Bau)

指導教授 ： 蔡丁貴

國立臺灣大學/工學院/土木工程學研究所/碩士(2006年)

https://doi.org/10.6342/NTU.2006.02743

摘要

過去洪水預報系統主要是針對河川中的流量，但河川流量的量測資料，準確度並不高，且需由率定曲線(rating curve)將流量轉換成水位後，方可判斷是否超出警戒水位。若能直接計算出未來發生的定點水位，作為河川洪水預報模式上游的邊界點輸入條件，對於洪水預報系統之發展，實有相當大的貢獻。因此本文擬利用倒傳遞類神經網路(back-propagation neural network,BP)，找出臨前水文觀測資料與未來水位之間的關係，進而預報未來一段時間後，水位的變化情形。
在建立類神經網路的預報模式之前，必須考慮影響未來水位變化之臨前水文因子，傳統上是以試誤法找出集流時間(time of concen-trateion)，再以集流時間內之逐時水文資料為模式之輸入單元，此法不但費時且效果不佳，因此本文在決定模式輸入單元之前，先將臨前各時段的水文資料與未來水位做相關性分析，取各臨前水文資料與未來水位變化相關係數最高的時段作為模式的輸入單元，以此建立類神經網路預報模式，不但可節省許多測試的時間，且較傳統以試誤法決定輸入單元合理。而由實例演算的驗證結果得知，以此方式建立之預報模式，其準確度亦有明顯的改善。

關鍵字

倒傳遞類神經網路 ； 相關性分析

並列摘要

In the past, the focus of flood forecasting model is on the flowrate in the river. It is unfortunate that the field measurements of the flowrate are not very accurate. Futhermore, the discharge measurement has to converted to water level by using the flowrate-level rating curve. It is not convenient and reliable to use the predicted water level in the flood warning system. It is intended to predict water level at a outlet of a watershed, to prouide upstream boundary condition, for river flood forecasting model. The technique of back-propagation neural network is employed to establish relationship between prior hydrological observations and water level. The developed relationship is then used for water level prediction when hydrological information is provided.
Tradictionally, the concept of concertration time of runoff in a watershed is assumed to relate future runoff flowrate to prior hydrological data. Based on the time-step hydrological input, the watershed runoff can be obtained. This is a tedious process and not reliable. In this paper, correlation analysis are performed to determined the highest correlation coefficient of prior hydrological events. The neural network technique is then employed to develope a prediction model for water level at the outlet of a watershed. It is demonstrated to be more efficient and consistent then tranditional approach. The advantages are illustrated by using historical events. Better perdiction accuracy is observed.

並列關鍵字

back-propagation neural network ； correlation analysis

參考文獻

1.Abrahart, R. j. and See, L., “Comparing neural network and autoregressive moving average techniques for the provision of continuous river flow forcasts in two contrasting catchments,” Hydrological Processes, 14, 2157-2172, 2000.

“A comparison between neural-network forcasting techniques-case study: river flow forecasting,” IEEE Transactions on Neural Networks, 10(2), 402-409, 1999.

3.Baba, K. et al. “Explicit representation of knowledhe acquired form plant historical data using neural network,” IJCNN-91, III, 155-160, 1991.

4.Cameron, D., Kneale, P., and See, L., “An evaluation of a traditional and neural net modeling approach to flood forecasting for an upland catchment,” Hydrological Processes, 16, 1033-1046, 2002.

5.Campolo, M., Andreussi, P., and Soldati, A., “River flood forcasting with a neural network model,” water Resources Research, 35(4), 1191-1197, 1999.