A Neural Network Model to Estimate the Clogging of Ash Mixed Sand Filter

Abstract

Infiltration filters of LID facilities are frequently confronted with clogging of pores by contaminants. To avoid malfunctioning of LID facilities, filters should be maintained according to the degree of clogging. In this study, clogging was characterized by curves representing the correlation between infiltration rate (<i>Q</i>) and infiltration quantity (<i>IR</i>). To estimate the degree of clogging, a regression neural network model was developed to estimate the label (infiltration quantity) corresponding to the selected features (ash mix ratio, infiltration rate, and variation of infiltration rate). According to the results obtained from the evaluation of accuracy and applicability of the final model trained over the whole training dataset, the following conclusions were drawn: 1) the training of the final model was verified to be correct according to mean absolute errors of 0.00445 and 0.00491 for the training and test datasets, respectively; 2) the final model could be applied to estimate multiple, multivariate clogging curves for arbitrary ash mixing ratios and trained range of time series in terms of infiltration rate; 3) given that the final model could not estimate the clogging curve for future time steps for which the model was not trained, a CNN model or LSTM-RNN model are suggested as an alternative.