Artificial Neural Networks Combined with Sensitivity Analysis as a Prediction Model for Water Quality Index in Juru River, Malaysia

Abstract

The manuscript describes the application of artificial neural networks (ANNs) for the series modeling of surface water quality prediction in Juru River, Malaysia. This is based on water quality data from twelve monitoring stations of Juru River (6 January 2003 until 5 November 2007) provided by Department of Environment (DOE), Malaysia. Thirty physicochemical parameters were involved in this analysis as input variables and water quality index as output variable. Three models were proposed to identify the most effective model in attempt to predict the WQI. Sensitivity analysis (SA) was carried out by using leave one out approach in order to indentify the most significant input-output relationship. The ANNs developed was successfully trained and tested using the available data sets and the performance of ANNs models was determined by coefficient of determination (R 2 ), coefficient of correlation (R) and root mean square error (RMSE). Results show that ANN-1 gives the higher value of R 2 (0.9942) and RMSE (2.8966), however this model was only trained and tested using all available parameters. The second model (ANN-2) gives R 2 value (0.9839) slightly higher compare to the third model, ANN-3 (0.9811). This is supported by the RMSE values which indicate that ANN-2 has a lower value compared to ANN-3 which is 2.1877 and 2.411. Hence, this study will trigger DOE to use ANNs in order to predict WQI other than using conventional method (WQI equation) that is currently being used by DOE. In addition, the ANNs managed to show remarkable prediction performance to predict the WQI in Juru River.