Assessment of water quality index in unmonitored river basin using multilayer perceptron neural networks and principal component analysis

This study describes the combination of neural networks and multivariate methods to develop a proper model for the forecasting of water quality index (WQI) in the Saf-Saf River using water quality parameters. The main objectives of this work were to determine the importance of different input variables and to assess the spatial and temporal water quality variation. MLP models were trained using three different algorithms and tested, these models were compared in terms of efficiency criteria and goodness-of-fit for WQI modeling. The results show that MLP BFGS model provide the best performance with small root mean square error value (RMSE = 0.007) and high coefficient of determination value (R 2 = 0.811) compared with the other types of MLP models. In the meantime, sensitivity analysis reveals that BOD 5 acts as the most contributor decreasing WQI. PCA/FA results show relatively spatial and seasonal changes in surface water quality, it generated three groups of sampling sites with similar characteristics. Group I (upstream sites), group II (midstream sites), and group III (downstream sites) correspond to a relatively low pollution, moderate pollution, and high pollution sites, respectively. Therefore, this approach can provide managers with the right tools to make decisions about the implementation of sustainable management practices.