Artificial Neural Network for Prediction of Total Nitrogen and Phosphorus in US Lakes

Abstract

Modeling is an important aspect of water quality management because it saves material and labor costs. The nonlinearity of water quality variables due to the complex chemical and physical processes in a body of water makes the modeling process difficult. This study used an artificial neural network (ANN) approach, a powerful computational tool for nonlinear relationships, to develop a model that estimates the summer concentration of total nitrogen (TN) and total phosphorus (TP) in US lakes using interrelated and easily measurable water quality parameters. Two ANN models, using regional and national data sets, and one linear regression model were trained, tested, and validated using three inputs (pH, conductivity, and turbidity) that were statistically correlated with the outputs. The prediction accuracy of the ANN models consistently outperformed the linear regression model. The statistical accuracy of the ANN models for regional data sets was superior to that of the national data set. A sensitivity analysis showed that pH was the most predictive parameter for nutrients. These results indicate that the ANN modeling technique can be a screening tool for an overall estimation of nutrient concentrations in regional lakes.