An artificial neural network model of a generalised channel network R. K. Price, J.N. Samedov and D. P. Solomatine International Institute for Infrastructural, Hydraulic and Environmental Engineering, Delft, The Netherlands ABSTRACT: Artificial neural networks (ANNs) are widely used for modelling complex phenomena by con- necting input and output data without requiring any detailed knowledge of the phenomena. Increasingly ANNs are being used to replicate complex physically-based models. This paper explores the possibility of replicating a particular class of model, namely the hydraulics of flow in a network of channels. Instead of connecting network inputs to the output(s) the concept is one of using an ANN to replicate the performance of a single element in the network, namely a channel. What is then of interest is the compilation of a network out of the individual ANNs for each channel component. The ANNs replicate the routing of a kinematic wave along a single channel. The results confirm that ANNs can be successfully used to replace complex, time consuming simulation models by fast-running replicates for such applications as real-time control. 1 INTRODUCTION In recent years artificial neural networks (ANNs) have been applied widely to a number of different fields (Fausett 1994). ANN technology was origi- nally inspired in the 1960s by neurological theories of interconnection and parallelism that describe cer- tain aspects of the functioning of the brain. Cur- rently ANNs are being used for such activities as pattern recognition and data description and model- ling, especially when no accurate physically-based model can be built either because the processes are not known or too complex to reproduce. The traditional way of implementing a neural network is to train it on the input and output for a given system. Data describing the system behaviour has either been monitored from the physical system or generated by a mathematical model. An ANN can then be trained on a particular set of the input and output data for the system and then verified in terms of its ability to reproduce another set of data for the same system. The hope is that the ANN can then be used to predict the output given the input. Such techniques have been applied, for example, to the rainfall-runoff process (Mason et al 1996; Hall and Minns 1996) and to combined hydrau- lic/hydrologic models (Solomatine and Torres 1996). A disadvantage of these models is that a separate neural network has to be constructed and trained for each particular catchment or river basin. In practice, this problem could be addressed by in- cluding the generic parameters describing the catchment or river basin as input to the ANN. Such an approach requires however, data from a number of different catchments. The advantage of using an ANN to model rainfall- runoff is that it avoids the need to consider in detail the complex physical processes involved. A disad- vantage is that the resulting ANN tells us nothing about the several, individual sub-processes. An al- ternative approach, which to the authors’ knowledge has not yet been tried, is to model each sub-process with its own ANN and to link them together to form an integrated model for the entire rainfall-runoff process in a given catchment. This requires the identification of each sub-process and the links with other sub-processes. There is yet another approach. Consider the clas- sification of a conceptual, lumped model of the rain- fall-runoff process as having a single structural ob- ject defining the physical domain (catchment) and many sub-processes. A physically-based distributed model of the rainfall-runoff process can be regarded as having many structural objects defining the physical domain (sub-catchments, cells or grid points) with many sub-processes. Each of the structural objects can be defined in such a way that it has the same defining characteristics. It might be possible to construct an ANN for each structural object and to hide the sub-processes within the ANN. Proc. of the 3 rd International Conference on Hydroinformatics, Copen- hagen, Denmark, 24-26 August 1998. Balkema, Rotterdam, pp. 813-818