Techniques for predicting total phosphorus in urban stormwater runoff at unmonitored catchments

Abstract

This paper investigates the applicability of using artificial neural network (ANN) and multilinear regression models to predict urban stormwater quality at unmonitored catchments. Models were constructed using logarithmically transformed environmental data. Violation of the assumption of data independence lead to the inclusion of insignificant variables when a straightforward stepwise regression was applied. To overcome this problem, cross validation was used to determine when to stop adding variables. Regression models calibrated using event mean concentration (EMC) as the dependent variable were more accurate than those using event load. Regression models developed on a regional subset of data were more accurate than the models developed on the entire data set. Even though regression and ANN models yielded similar predictions, regression modelling was considered to be a more applicable approach. Compared to ANN models, regression models were faster to construct and apply, more transparent and less likely to overfit the limited data.