A Study of Dispersed, Thermally Activated Limestone from Ukraine for the Safe Liming of Water Using ANN Models

• Published in: Energies (Basel) Vol. 14; no. 24; p. 8377
• Main Authors: Trach, Yuliia, Trach, Roman, Kalenik, Marek, Koda, Eugeniusz, Podlasek, Anna
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.12.2021
• Subjects: Alkalinity; Aquatic plants; artificial neural network; Back propagation; Chemical oxygen demand; Contamination; Dispersion; Environmental protection; Fruits; Heavy metals; Interdisciplinary research; Interdisciplinary studies; Lakes; Limestone; Liming; Metal concentrations; Mineralization; modelling; Neural networks; Nitrates; Nitrogen; Phosphates; Phosphorus; Rivers; Salinity; Sensitivity analysis; Surface water; Variables; water pH; Water pollution; Water quality; Water supply; Water temperature; Water treatment; Water treatment plants
• ISSN: 1996-1073; 1996-1073
• DOI: 10.3390/en14248377

Liming surface water is a fairly popular method of increasing the pH values and decreasing the concentration of phosphates and heavy metals. According to the Environmental Protection Agency (EPA) recommendations, the increase of water pH should not exceed 1.5. If surface water is the source of water supply, liming is a process that reduces water contamination. This should prevent the creation of an additional load for the water treatment plants in urban settlements. This article is an interdisciplinary research study aiming to (1) determine and compare the doses of new dispersed, thermally activated limestone and natural limestone, (2) find the relation between dose value and initial water parameters (pH, Eh and total mineralization), and (3) create an artificial neural network (ANN) model to predict changes in water pH values according to EPA recommendations. Recommended doses were obtained from experimental studies, and those of dispersed, thermally activated limestone were lower than the doses of natural limestone. Neural networks were used to predict the changes in water pH values when adding different doses of limestone with different initial water parameters using the ANN model. Four ANN models with different activation functions and loss function optimizers were tested. The best results were obtained for the network with the ReLU activation function for hidden layers of neurons and Adam’s loss function optimizer (MAPE = 14.1%; R2 = 0.847). Further comparison of the results of the loss function and the results of calculating the quality metric for the training and validation dataset has shown that the created ANN can be used to solve the set research issue.