River flow prediction: an artificial neural network approach

A multi-layer perception neural network trained with the backpropagation algorithm is adopted to make daily river flow forecasts at the Nakhon Sawan gauging station on the Chao Phraya River in Thailand. Lead times from 1 day to 28 days have been used. The predicted flow rates when compared with observed ones agree within acceptable limits for both training and testing data for most of the considered lead times. The study also demonstrates that different network structures for the longer lead times can enhance the prediction. K e y w o r d s artificial neura l ne tworks ; r iver f low predic t ion; cross-val idat ion; C h a o Ph raya River I N T R O D U C T I O N River flow prediction is an important topic of research for hydrologists for a number of reasons. For example, it is useful and necessary for flood disaster mitigation. It helps in the planning, management and operation of water resources development projects. The topic is however a difficult one because of the complexity of the physical processes involved in the generation of river flows. An approach based on physics is still far from being realized and researchers have therefore focused attention on the use of "data driven" techniques in the recent years. Data driven approaches do not lead to a better understanding of the process of the river flow generation, but they have their advantages. They are very practical for "data rich" but "theory weak" systems. Model formulations are often quite simple and nonlinearities in the system pose no problems. One disadvantage is that extrapolation of a model beyond the data range that has been used for calibration is not reliable if the system is nonlinear. In this study, one such data driven technique, the artificial neural networks (ANN) approach is pursued with the objective of predicting daily river flows using lagged variables. Artificial neural networks are capable of performing nonlinear modelling without a priori knowledge about the relationship between input and output variables. They learn from given data and capture the functional relationships among the data even if the underlying relationships are unknown. After learning from sample data, they can be used to approximate a continuous function to any desired accuracy (Hornik et al, 1989; Hornik, 1991) using unseen input data. These distinguishing features make ANNs a general and flexible alternative modelling tool for hydrological time series prediction in recent years (Dawson & Wilby, 1998; Coulibaly et al, 2000; Jayawardena & Fernando, 2000; Jayawardena et al, 2000). 240 A. W. Jayawardena & T. M. K. G. Fernando