Performance prediction of horizontal flow constructed wetlands by employing machine learning

This study is aimed to aid in the horizontal flow constructed wetland (HFCW) design by analysing their treatment dynamics using secondary datasets (n = 1232). Prediction of effluent Biochemical oxygen demand (BOD), Chemical oxygen demand (COD), Ammonium-Nitrogen (NH4+-N), Total Nitrogen (TN), and Total phosphorous (TP) was carried out via machine learning algorithms - multiple linear regression and support vector regression (SVR). Out of the two models, SVR resulted in a better prediction of all the effluent parameters in mg/l as well as in g/m3-d. The prediction of NH4+-N and TN (g/m3-d) could be performed with the highest accuracy (R2 and root mean square error 0.847 and 0.44 %; and 0.947 and 0.18 % respectively). The classification of the dataset according to organic loading rate (OLR) enhanced the performance of SVR for BOD (high OLR), COD (low OLR), and TP (low OLR). The performance of all the trials was acceptable in the training and validation stages in 3-fold cross-validation which was helpful in reducing the mean square error value by 68.19 % and increasing the R2 value by up to 16.13 %. Areal removal rate (K, m/d) values computed by the reverse P-k-C* approach using the predicted effluent concentrations were found to be highly correlated with actual K values for BOD, COD, and TN (R2 0.954, 0.948, and 0.819, respectively). These K values can be used for the customized design of HFCWs for organics and nitrogen removal and might also help to achieve the targeted discharge standards in HFCWs.