Predicting physico-chemical parameters of Barekese reservoir using feedforward neural network

Abstract

The Barekese reservoir which is an earth-filled dam impounding about 35.3 million m³ of water provides potable water to Kumasi and its environs. With most of the population depending on this resource, it is extremely important to be mindful of the quality of water produced as anything below ideal quality could lead to catastrophic public health issues. This study evaluates the suitability of a Feedforward Neural Network (FNN) in predicting six vital water quality parameters: pH, turbidity, temperature, total dissolved solids (TDS), alkalinity, and nitrate concentration. Historical water quality data spanning 2010–2021 were obtained from the Ghana Water Company Limited (GWCL). A backpropagation FNN was trained, validated, and tested using a 70–15–15 % split strategy. The Six (6) parameters were predicted using 6 distinct optimal FNN models derived from Bayesian optimization. The optimization defined the optimal number of neurons and Layers needed for predicting the physio-chemical Properties of the reservoir. Model Performance metrics such as the Mean Square Error (MSE), Average Absolute Percent Relative Error (AAPRE), Standard Deviation, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of determination (R²). The FNN models developed performed exceptionally for predictions of pH, Turbidity, and Nitrate, this was seen with the least errors and measures of accuracy greater than 0.99. FNN models developed for Temperature and Alkalinity prediction were also good but slightly less precise comparatively. The worst performing FNN model was that for TDS prediction which show the highest model variability defined by high errors relative to other models in this work. This study provides an effective data-driven approach and basis for real-time water quality monitoring