Neural Network ABAC with Dropout Layer for Activated Sludge System

Abstract

Due to the expensive operation of the activated sludge process and more stringent effluent requirements of wastewater treatment plant (WWTP), the wastewater treatment operator has been forced to find an alternative to improve the current control strategy, especially for those operating using an activated sludge system. The study aims to reduce the energy usage of a WWTP and to increase the effluent quality to meet the requirements of state and national laws by using the aeration control technique. The goals are achieved by varying the dissolved oxygen concentration in the benchmark plant's fifth tank according to the real ammonium measurement, a technique known as Ammonium-based aeration control (ABAC), which produced less nitrogen, resulting in better effluent and lower energy consumption. The simulation model Benchmark Simulation Model No. 1 (BSM1) was used to analyze ABAC in this study. The neural network (NN) model is used to design the ABAC controller, and simulation results were compared to the Proportional Integral (PI) controller of the BSM1 and PI ABAC control configurations. A dropout layer was added during the training process to improve neural network generalization. The dropout layer in the NN ABAC has improved the performances in terms of total nitrogen effluent violations by 4 percent less than the PI-ABAC and by 36 percent less than the PI. The NN ABAC LM dropout has been proven to be more effective in terms of energy efficiency by significantly reduced by 25 percent, effluent quality by successfully improved by 1 percent, and successfully reduced the total overall cost index by 5 percent when compared to PI-ABAC control. The study has illustrated that the NN ABAC could be used to improve the performance of the activated sludge system.