Abstract The quality of groundwater in the source area of pollution depends on many factors, including the weather and hydrogeological conditions within the given area. Anassessment of water quality can be carried out based on data obtained from sensors placed in boreholes. This research examined the influence of air and water temperature, groundwater table position and precipitation on the value of electrical conductivity in groundwater in a selected piezometer belonging to the monitoring network of the Quaternary aquifer in the area of a waste landfill site in Tychy-Urbanowice in southern Poland. The influence of individual factors was checked by using twenty neural network architectures of a Multilayer Perceptron Model (MLP). Each of these indicated factors were selected as input variables. Ultimately, three neural networks were selected, which were characterized by the smallest validation and test errors and showed the highest learning quality. The significance of individual variables for the effectiveness of the model was checked using a global sensitivity analysis. Three selected MLP models contained seven to nine neurons in the hidden layer and used a linear or exponential function as the hidden and output activation. The maximum test quality was 0.8369, while the smallest test error was 0.0011. The results of the sensitivity analysis highlighted the important role of water temperature and water table position on the conductivity value. The obtained goodness of fit results of the models to the input data allowed us to conclude that the MLP was applicable to such forecasts and can be extended by the analysis of further factors.
Read full abstract