Applying Multi-Layer Artificial Neural Network and Mutual Information to the Prediction of Trends in Dissolved Oxygen

Abstract

Predicting {\color{red}{trends in water quality}} plays an essential role in the field of environmental modelling. Though artificial neural networks (ANN) have been involved in predicting water quality in many studies, the prediction performance is highly affected by the model's inputs and neural network structure. Many researchers selected water quality variables based on Pearson correlation. However, this kind of method can only capture linear dependencies. Moreover, when dealing with multivariate water quality data, ANN with the single layer and few numbers of units show difficulties in representing complex inner relationships between multiple water quality variables. Hence, in this paper we propose a novel {\color{red}{model}} based on multi-layer artificial neural networks (MANN) and mutual information (MI) for predicting the trend of dissolved oxygen. MI is used to evaluate and choose water quality variables by taking into account {\color{red}{the non-linear relationships between the variables}}. A MANN model is built to learn the levels of representations and approximate complex regression functions. Water quality data collected from Baffle Creek, Australia was used in the experiment. Our model had around 0.95 and 0.94 $R^2$ scores for predicting 90 mins or 120 mins ahead of the last observed data in the wet season, which are much higher than the typical ANN model, support vector regressor (SVR) and linear regression model (LRM). The results indicate that our MANN model can provide accurate predictions for the trend of DO in the upcoming hours and is a useful supportive tool for water quality management of the aquatic ecosystems.