Machine Learning Application for Nutrient Removal Rate Coefficient Analyses in Horizontal Flow Constructed Wetlands

Abstract

Land area optimization for horizontal flow constructed wetlands (HFCWs) with a low organic loading rate (OLR) needs special considerations as the microflora changes dramatically with the OLR. The P-k-C* approach does not lead to an accurate calculation of k-values in these wetlands. In this research, nonlinear machine learning models [Support Vector Regression (SVR), Random Forest (RF), and Artificial Neural Networks (ANN)] are applied to predict realistic k-values. Data from 37 low-OLR HFCWs (n = 544) were analyzed, and the k-values calculated for these wetlands were found to vary markedly (0.059–0.249 with an average of 0.113 ± 0.090 m/day). The classification of k-values based on the OLR, applied loading rate, and media depth leads to the reduction in standard deviations (SDs) from 83.40 to 35.27%. k-values with the least SDs are needed for optimal design for low-OLR CWs. The SVR, RF, and ANN models were tested, and the best prediction efficiency on testing datasets was achieved through the ANN model with R2(kTKN)= 0.768 (RMSE = 0.067) for total Kjeldahl nitrogen (TKN), R2(kTN)= 0.835 (RMSE = 0.043) for total nitrogen (TN), and R2(kTP) = 0.723 (RMSE = 0.087) for total phosphorus (TP). The outcome was validated using primary data from HFCWs, which also confirmed the superiority of the ANN-based model, which can be used for design customization of low-OLR HFCWs.