Machine Learning and Statistical Analysis in Groundwater Monitoring for Total Dissolved Solids Assessment in Winkler County, Texas

Abstract

This research aims to develop reliable models using machine learning algorithms to precisely predict Total Dissolved Solids (TDS) in wells of the Permian basin, Winkler County, Texas. The data for this contribution was obtained from the Texas Water Development Board website (TWDB). Five hundred and ninety-three samples were obtained from two hundred and ninety-eight wells in the study area. The wells were drilled at different county locations into five aquifers, including Pecos Valley, Dockum, Capitan Reef, Edward Trinity, and Rustler aquifers. A total of fourteen different water quality parameters were used, and they include Potential hydrogen (pH), Sodium, Chloride, Magnesium, Fluoride, TDS, Specific Conductance, Nitrate, Total Hardness, Calcium, Temperature, Well Depth, Sulphate, and Bicarbonates. Four machine learning regression algorithms were developed to get a good model to help predict TDS in this area: Decision Tree regression, Linear regression, Support Vector Regression, and K-nearest neighbor. The study showed that the Decision Tree produced the best model with attributes like the coefficient of determination R2 = 1.00 and 0.96 for the training and testing, respectively. It also produced the lowest score of mean absolute error MAE = 0.00 and 0.04 for training and testing, respectively. This study will reduce the cost of obtaining different water quality parameters in TDS determination by leveraging machine learning to use only the parameters contributing to TDS, thereby helping researchers obtain only the parameters necessary for TDS prediction. It will also help the authorities enact policies that will improve the water quality in areas where drinking water availability is a challenge by providing important information for monitoring and assessing groundwater quality.