Determination of Suitable ANN Architecture for Groundwater Fluoride Prediction

Abstract

Using soft computing methods is considered as an innovative approach for constructing a computationally intelligent system, having humanlike expertise in a specific domain. They learn the system and adapt themselves with the varying pattern and at the end they take decisions for future trends. The applications of these techniques, for prediction of trends in environmental problems, are increasing day by day. In the present study, the ANN model was used for groundwater fluoride prediction in western parts of Jharkhand district, India. MATLAB software is used for artificial neural network and graphical user interface (GUI) toolboxes. The performance criteria used for the measurement of efficiency and accuracy were root mean square error (RMSE) and regression coefficient (R²). The 6-year dataset for fluoride level in groundwater is used for prediction. The input parameters used were depth, groundwater level, electrical conductivity (EC), pH, chloride (Cl), bicarbonates (HCO₃), calcium (Ca), magnesium (Mg), and sodium (Na). The search of suitable architecture was done by the evaluation of predicted output with the actual fluoride level data. Simulation results reveal that three-layered feed forward backpropagation type of network with Levenberg–Marquardt (LM) training algorithm is the promising architecture for groundwater fluoride prediction.