Search

Subsurface constructed wetland (SCW) appears to be an economical and environmental-friendly practice to treat nitrogen-enriched (waste) water. Nevertheless, the removal mechanisms in SCW are complicated and rather time-consuming to conduct and assessment the efficiency of new experiments. This work mined data from literature and developed the machine learning models to elucidate the effect of influent inputs and predict ammonium removal rate (ARR) in SCW treatment. 755 sets and 11 attributes were applied in four modeled algorithms, including Random forest, Cubist, Support vector machines, and K-nearest neighbors. Six out of ten input features including ammonium (NH4), total nitrogen (TN), hydraulic loading rate (HLR), the filter height (i.e., Height), aeration mode (i.e., Aeration), and types of inlet feeding (i.e., Feeding) have posed pronounced influences on the ARR. The Cubist algorithm appears the most optimal model showing the lowest RMSE i.e., 0.974 and the highest R 2 i.e., 0.957. The contribution of variables followed the order of NH4, HLR, TN, Aeration, Height and Feeding corresponding to 97, 93, 71, 49, 34, and 34%, respectively. The generalization ability to forecast ARR using testing data achieved the R 2 of 0.970 and the RMSE of 1.140 g/m 2 d, indicating that Cubist is a reliable tool for ARR prediction. User interface and web tool of final predictive model are provided to facilitate the application for designing and developing SCW system in real practice.

This study presents possible applications of machine learning (ML) methods for estimating water levels without a throughout understanding of hydrological processes and complex databases of irrigation systems. The Bac-Hung-Hai catchment, the biggest irrigation and drainage area in Vietnam, is selected as a case study due to the large database on this case consisting of 3348 samples drawn over a 21-year monitoring period. The state-of-the-art Gradient tree boosting (GTB)-based model was developed and is compared with eight other common ML methods. The proposed GTB-based model consistently showed the best performance, with the lowest value of mean-squares-error and the greatest values for R 2 and adjusted R 2 in all case studies. Moreover, over 91% of the total samples had an error rate of below 10% between the predicted and the observed values. The results suggested that the GTB model can predict water level with high accuracy, thus helping researchers and policy-makers devise proactive strategies for hydraulic regulation and sustainable water management.