Water Resources Management 13: 219–231, 1999. © 1999 Kluwer Academic Publishers. Printed in the Netherlands. 219 A Genetic Programming Approach to Rainfall-Runoff Modelling DRAGAN A. SAVIC ⋆ , GODFREY A. WALTERS and JAMES W. DAVIDSON The Centre for Water Systems, School of Engineering and Computer Science, Department of Engineering, University of Exeter, Harrison Building, North Park Road, Exeter EX4 4QF, United Kingdom (Received 9 October 1998; accepted 10 May 1999) Abstract. Planning for sustainable development of water resources relies crucially on the data avail- able. Continuous hydrologic simulation based on conceptual models has proved to be the appropriate tool for studying rainfall-runoff processes and for providing necessary data. In recent years, artificial neural networks have emerged as a novel identification technique for the modelling of hydrological processes. However, they represent their knowledge in terms of a weight matrix that is not access- ible to human understanding at present. This paper introduces genetic programming, which is an evolutionary computing method that provides a ‘transparent’ and structured system identification, to rainfall-runoff modelling. The genetic-programming approach is applied to flow prediction for the Kirkton catchment in Scotland (U.K.). The results obtained are compared to those attained using two optimally calibrated conceptual models and an artificial neural network. Correlations identified using data-driven approaches (genetic programming and neural network) are surprising in their consistency considering the relative size of the models and the number of variables included. These results also compare favourably with the conceptual models. Key words: artificial neural networks, genetic programming, identification, rainfall-runoff model- ling. 1. Introduction Planning for sustainable development of water resources relies crucially on the data available for planning activities. Streamflow forecasting and the use of flood fre- quency analysis in the design of dams and bridges or for control of water supply and flood control systems cannot be performed without extended records of streamflow data. Continuous hydrologic simulation based on parametric (conceptual) models has proved to be an effective tool for studying rainfall-runoff processes and for providing the necessary data (Todini, 1988). The main characteristic of these mod- els is that they have a simple, assumed structure with model parameters that are physically relevant to large sub-areas of the catchment or to the whole area. Figure 1 shows a schematic representation of such a simple model with the unknown parameters K i (Nash, 1958). However, these parameters have to be identified either ⋆ E-mail: D.Savic@exeter.ac.uk.