Optimization of the Wastewater Treatment Plant Recycle Flowrates Using Artificial Neural Networks

Process control in wastewater treatment plants (WWTPs) is, in a varying extent, relying on the operator's experience and is often based only on the quality of the effluent water. An alternative to such not always efficient approach is the development of artificial intelligence tools for plant performance prediction and operation optimization. This work focuses first on developing accurate artificial neural network (ANN) models capable of predicting effluent quality and energy performance indices for the WWTP. Second, these models are further utilized in operation optimization by manipulating the nitrates and activated sludge recirculation flowrates, in association to aeration control. Four categories of ANNs were considered, two of nonlinear autoregressive networks with exogenous inputs (NARX), one of Radial Basis (RBNN) and one of Generalized Regression (GRNN) types. The training dataset was obtained from simulations with an analytical model that was calibrated with real plant data. Topologies of NARX type networks were found to be optimal for predicting performance indices, both for single output and multiple output network structures. The ANN models showed high accuracy, as their mean absolute percentage error (MAPE) values between predicted and targeted outputs were ranging from 0.85% to 3.50%. Optimization using the developed ANN models showed similar results to those obtained with the use of the analytical model. The meaningful difference in the optimization computation time was revealed when comparing it to the optimization performed with the analytical model. The ANNs required four orders of magnitude less computation time, proving the efficacy and potential of the proposed method to be used for real time optimization applications and digital twins implementation.