A random forest model for inflow prediction at wastewater treatment plants

Influent flow of wastewater treatment plants (WWTPs) is a crucial variable for plant operation and management. In this study, a random forest (RF) model was applied for daily wastewater inflow prediction, and a new probabilistic prediction approach was, for the first time, applied for quantifying the uncertainties associated with wastewater inflow prediction. The RF model uses regression trees to capture the nonlinear relationship between wastewater inflow and various influencing factors, such as weather features and domestic water usage patterns. The proposed model was applied to the daily wastewater inflow prediction for two WWTPs (i.e., Humber and one confidential plant) in Ontario, Canada. For the confidential WWTP, the coefficient of determination (R2$$\varvec{R}^{2}$$) values for training and testing were 0.971 and 0.722, respectively. The R2$$\varvec{R}^{2}$$ values at the Humber WWTP were 0.957 and 0.584 for training and testing, respectively. In comparison with other approaches such as the multilayer perceptron neural networks (MLP) models and autoregressive integrated moving average models, the results show that the RF model performs well on predicting inflow. In addition, probabilistic prediction of daily inflow was generated. For the Humber station, 93.56% of the total testing samples fall into its corresponding predicted interval. For the confidential plant, 78 observed values of the total 89 samples fall into its corresponding interval, accounting for 87.64% of the total testing samples. The results show that the probabilistic approach can provide robust decision support for the operation, management, and optimization of WWTPs.