Potable Water Quality Assessment Using Machine Training Methods

Continuous water pollution and environmental degradation make it difficult to maintain the quality of potable water acceptable for safe consumption. To get potable water of the required quality, water from a water body is treated in a water purification system. After purification, the quality indicators of the purified water are measured. When the received parameters go beyond acceptable limits (emergency situation) the water supply system starts receiving drinking water stored in tanks, and the water treatment system itself is temporarily stopped in order to determine the reasons for failure. The water purification process is adjusted, if necessary. To minimize the time to eliminate emergency situations, an immediate response to potable water quality violation is necessary. Here we get a problem of predicting possible violations based on the results of investigation of water source physical and chemical parameters, using appropriate mathematical models and information technologies. The initial parameters in the problem under consideration are the results of monitoring the indices of water coming from a water source, as well as the quality of potable water leaving the treatment system within a definite period. It is necessary to predict the values of potable water quality parameters, and if the quality violation is predicted, to assess the necessity to correct the reagent dosages. Currently, the most common methods of water quality monitoring based on statistical data are regression analysis, group method of data handling, artificial neural networks, etc. It is worth noting that water quality prediction is rarely applied to the existing water treatment systems, and where it is applied the prediction error is too high (>10%). In this paper, the following approaches are proposed to solve the problem. Firstly, standard regression model is created, and its quality is assessed. If the accuracy of the constructed model is insufficient, machine learning methods are used: gradient boosting, random forest or support vector machine with the hyperparameter adjustment using various algorithms. If the decision to change the reagent dosage is made, fuzzy logic is applied. The methodology of mathematical models is illustrated on the example of predicting one of the important parameters, namely, residual chlorine in the potable water.