Artifical neural network model for rainfall-runoff relationship

Artificial neural networks (ANNs) have emerged as an alternative approach in modeling the runoff process in which the explicit form of the relationship between the variables involved is unknown. This thesis focuses on the implementation of artificial neural network to forecast the three hourly discharge in the small watershed area with limited hydrologic data. Data for model calibration and validation are obtained from only one available hydrologic station (P.64) at the Mae Tuen River in Om Koi District, Chiang Mai Province located in the northern part of Thailand. The watershed is small with approximately 503 square kilometers. It has a distinct hydrologic feature with relatively little information on topography and runoff data. A feedforward backpropagation ANN is used to model and forecast the three hourly discharge. Rainfall-runoff relationship in this studied area was previously investigated by Pukdeboon (2001). He used the Tank model, originally proposed by Sugawara (1974), to forecast the three hourly discharge for the wet- and dry- periods. Based on the same set of hydrologic data, comparisons of the predicted discharge from the Tank model and ANN are presented. The results showed that the average relative error computed from the ANN (10.17%) substantially decreased when comparing with the average relative error computed Tank model (41.35%). The performance evaluation of these two models, based on various statistics, are presented and discussed.当所涉及变量间的显式关系未知时，人工神经网络（Artificial Neural Networks，ANNs）已成为径流过程建模的可选方案之一。本论文聚焦于利用人工神经网络，对水文数据有限的小流域开展逐三小时流量预报。模型校准与验证所用数据，仅取自泰国清迈府翁桂县湄屯河上唯一可用的水文站P.64。该流域面积约503平方公里，规模较小，虽具有独特的水文特征，但地形与径流数据相对匮乏。本论文采用前馈反向传播人工神经网络（Feedforward Backpropagation ANN）进行径流建模与逐三小时流量预报。该研究区域的降雨-径流关系此前已由Pukdeboon（2001）开展过研究，其采用杉原（Sugawara，1974）最初提出的水箱模型（Tank model），针对丰水期与枯水期开展逐三小时流量预报。本论文基于同一套水文数据，对比了水箱模型与人工神经网络的预报流量结果。结果显示，人工神经网络的平均相对误差为10.17%，相较水箱模型的41.35%平均相对误差大幅降低。本论文基于多项统计指标，对这两种模型的性能评估结果进行了呈现与讨论。