New Perspective\nfor the Prediction of Pollutant Removal\nEfficiency in Constructed Wetlands: Using a Genetic Algorithm-Assisted\nMachine Learning Model

Constructed wetlands (CWs) are widely used for wastewater\ntreatment,\nbut their performance is difficult to predict due to varying factors\nlike local weather, hydraulic conditions, vegetation, and wastewater\ncomposition. Here, we proposed a model method for predicting CW processing\nefficiency based on published literature simulations using machine\nlearning methods. Through data mining, we divided the obtained variables\ninto six different categories and proposed different data repair strategies\nfor each category. To improve the model performance, a genetic algorithm-assisted\ndatabase dimensionality reduction method was introduced in the model\ndestruction. After model selection and hyperparameter optimization,\nthe random forest algorithm was selected as the final algorithm, and\nthe model performances for all four predictions (ammonia nitrogen,\ntotal nitrogen, total phosphorus, and chemical oxygen demand removal\nefficiency) were 0.9405, 0.8277, 0.8136, and 0.8877, respectively.\nGenerally, the magnitude of the influence of the different categories\nis listed in the following order: meteorology/location > hydraulic\ncondition > substrate property ≈ water quality ≈\nreactor\nsize > vegetation. Based on this work, the future design and operation\nof CWs might find an efficient and environmentally friendly method\nthat could ideally maximize pollution control and economic benefits\nat the same time.