Comparison of Artificial Neural Network, Genetic Programming, Genetic Algorithm, and Multiple Linear Regression for Water Quality Modeling

In a watershed framework, the selection of a particular type of water quality model depends on several factors such as complexity of process being modeled, input data requirements, modeling objectives, and model applicability. For most applications, process-based simulation models or mechanistic models are routinely used to quantify the response of different hydrologic and water quality processes occurring in a watershed. In a complex watershed, the modeling objectives may require the use of multiple models of varying complexity. For instance, both a watershed-scale loading as well as a receiving water model may be needed for a watershed of sufficient complexity in which both point and non-point sources of pollution are being modeled. Recently, inductive or data-driven models are increasingly used for applications in watershed management. Examples of inductive models range from simple linear regression models to more complex nonlinear models based on artificial neural networks. Both linear and non-linear inductive models can be used to fit a mathematical model to a given data set in order to represent a process. Inductive or data-driven models are becoming more and more popular due to their ease of use and simplicity as substitutes for more process-based models in a number of applications. For instance, inductive models may be preferred where 1) computational expense is a critical issue, 2) the process-based deductive models are over parameterized and cannot be adequately calibrated, 3) budgetary constraints do not allow for a complex deductive model, and 4) quick and simple models are needed for integration into an optimal management framework for evaluating multiple scenarios in a relatively short period of time. Both explicit inductive or implicit inductive models can be developed in such applications. While implicit inductive models require output from a calibrated mechanistic model of the watershed, explicit inductive models can be easily developed using raw data collected for the process being modeled. More recently, inductive models derived using evolutionary and biological principles are becoming increasingly popular. These include artificial intelligence-based models such as artificial neural networks, genetic algorithms, and genetic programming. This paper will compare these techniques among themselves as well as with a simple baseline technique such as multiple linear regression models for application to water quality modeling in a watershed management framework. Example applications include modeling water quality parameters such as pathogens, dissolved oxygen, total nitrogen, and total phosphorus in an urban watershed.