PREDICTION THE AVERAGE ANNUAL CONCENTRATION OF PHOSPHATE IN RIVERS USING ANN AND MLR MODELS

Clean and quality water is one of the main aims in achieving the seventeen adopted Sustainable Development Goals. That means taking legislative measures at national and international level for the preservation of water quality, prevention of pollution and continuous monitoring. According to global and regional cooperation which is stated by numerous legal regulations, all member countries have obligation to monitoring water quality and submit annual report. For that reason it is very important to obtain accurate and precise data of all monitored water pollutants. In developing countries, because of poor and underdeveloped environmental infrastructure, obtained data for pollutants could have a certain percentage of uncertainty. In order to avoid that problem it is very important the existence of a larger number of models to estimate concentration of water pollutants, including concentration of phosphate. Due to the increasing influence of anthropogenic sources on the content of phosphorus, first in the soil and then in water sources, and negative effects caused by the eutrophication process, it is crucial to monitor the state of water in terms of the presence of phosphate. In the present study, in order to obtain one optimal model to predict the average annual concentration of phosphate in rivers was used two different approaches. In the first approach, was used artificial neural network (ANN), two different ANN architecture: Multilayer Perceptron (MLP) and Radial Basis Function (RBF). In the second approach were used two type of Multiple Linear Regression (MLR) technique, simultaneous and stepwise MLR. Available data used in this study were obtained from Eurostat-the statistical office of the European Union, for 19 European countries for the period from 2004 to 2012. For all created models was used the same dataset. As input variables were used sustainable development indicators, in total 10. All selected inputs were chosen on the basis of their influence on phosphate concentration in rivers and also reviewing of adequate literature. As model performance indicators in this study was used Mean Absolute Error (MAE) and coefficient of determination (R2). Results of all created models first at all showed that comparing two ANN models, better performances on training and test dataset shows MLP model. On test dataset, MLP model gives satisfactory prediction with value of coefficient of determination 0,657. On the other side, developed MLR models show significantly poorer results. Taking into account performances of created models, it is clearly that MLP model shows the best predictive results. On the basis of results of MLP model, could be recognized that it can be an alternative model for prediction annual concentration of phosphate in rivers, and significant tool in water pollutant monitoring.