Multilevel split of high-dimensional water quality data using artificial neural networks for the prediction of dissolved oxygen in the Danube River

Abstract

In this study, a self-organizing network-based monitoring location similarity index (LSI) was coupled with Ward neural networks (WNNs) with the aim to create a more accurate, but less complex, multiple sites model for the prediction of dissolved oxygen (DO) content. This multilevel splitting approach comprises the LSI-based grouping of monitoring locations according to their similarity, and virtual splitting of processed data based on their features using WNN. The values of 18 water quality parameters monitored for 12 years at 17 sites on the Danube River flow thought Serbia were used. The optimal input combinations were selected using partial mutual information algorithm with termination based on the Akaike information criterion. LSI-based splitting has yielded two groups of monitoring sites that were modeled with separate WNN models. The number and types of selected inputs differed between those two groups of sites, which was in agreement with possible pollution sources. Multiple performance metrics have revealed that the WNN models perform similar or better than multisite DO prediction models published in the literature, while using two to four times less inputs and data patterns.