Artificial Neural Network Model for Rainfall-Runoff Relationship

Conceptual models have been widely used and are considered to be the best choice for describing the runoff process in a watershed. In most cases, the solution accuracy is mainly based on the topographic and hydrologic information subject to certain requirements for model calibration. Thus, these types of model are inappropriate for watershed area with little hydrologic data. Artificial neural network (ANN) has become an alternative approach to model the runoff process in situations where explicit knowledge of the internal hydrologic processes is not available. ANN has a flexible mathematical structure which is capable of identifying complex nonlinear relationships between the sets of input and output data. In this paper, ANN is proposed as a tool to predict runoff hydrograph from only one hydrologic station for the Mae Tun River in Omkoi District, Chiang Mai Province which is located in the northern part of Thailand. The watershed area to be considered here is approximately 503 square kilometers. It has distinct hydrologic features with insufficient information on topography and rainfall runoff data. In this study the problem is viewed as a time series prediction problem. A feed-forward artificial neural network is trained by using back-propagation algorithm. The training and testing data were collected during years 2000 to 2004. Since the original collected data contain only discharge and rainfall data, it is rather difficult to obtain accurate results. To improve the accuracy, the amount of water excess was calculated from the collected data and used as an additional input for training the network. The input to the network consists of discharge, rainfall measurement and discharge excess during the last 24 hours. The performance comparison is presented by statistic evaluation of percentage errors and correlations. From the simulations, it is found that random input pattern gives more accurate results than the well-order input pattern.