Developing a new approach for design support of subsurface constructed wetland using machine learning algorithms.

Knowing the effluent quality of treatment systems in advance to enable the design of treatment systems that comply with environmental standards is a realistic strategy. This study aims to develop machine learning - based predictive models for designing the subsurface constructed wetlands (SCW). Data from the SCW literature during the period of 2009-2020 included 618 sets and 10 features. Five algorithms namely, Random forest, Classification and Regression trees, Support vector machines, K-nearest neighbors, and Cubist were compared to determine an optimal algorithm. All nine input features including the influent concentrations, C:N ratio, hydraulic loading rate, height, aeration, flow type, feeding, and filter type were confirmed as relevant features for the predictive algorithms. The comparative result revealed that Cubist is the best algorithm with the lowest RMSE (7.77 and 21.77 mg.L for NH-N and COD, respectively) corresponding to 84% of the variance in the effluents explained. The coefficient of determination of the Cubist algorithm obtained for NH-N and COD prediction from the test data were 0.92 and 0.93, respectively. Five case studies of the application of SCW design were also exercised and verified by the prediction model. Finally, a fully developed Cubist algorithm-based design tool for SCW was proposed.