Primary Exploration Of Neural Network Based Adaptive Control For Wastewater Treatment

There exist various kinds of sewage treatment processes,however,how to ensure the effluent to be kept under certain standards and regulations is a common problem for all wastewater treatment systems.This problem is related to the influent fluctuations,the complex biochemical reactions,the technology limits for on-line monitoring and the excessive dependence on human interference.To handle these problems,researchers have developed expert systems,mechanism models and data-driven models,in order to promote fault diagnosis,real-time control and performance imporvement,and finally to achieve the stable and efficient operation of the whole system.Yet,the effictive applications of the above-mentioned models are linked to the deep understandings of the controlled process or the long-term accumulation of monitoring data,both of which limit further applications of these approaches.Focusing this restricts,this paper developed an reinforcement learning method based on Deep Q-Network(DQN)for adapative control,with the support of Gate Recurrent Unit(GRU)neural netwroks to weak the inflow disturbance.The main results of the paper are listed as below:(1)Benchmark Simulation Model NO.1(BSM1)simulatin platform is developed based on Python.This platform was just a series compiled scripts written in Python and covering the biochemical reactions,various kinds of variables and parameters in Actived Sludge Model NO.1(ASM1)and the mass balance in BSM1.The construction of the platform laid the foundation of the simualtion experiments related to DQN based model.(2)GRU recurrent neural networks are constructed to predict the sewage inflow.The design and training of the recurrent networks are based on the module called Keras and compared with certain traditional time series analysis methods in two cases:BSM1 inflow prediction and leachate yield forecasting.Results show that,compared to the traditional methods,the GRU recurrent neural networks show better performance due to their ability of non-linear fitting and long-term memory.(3)The DQN based reinforcement learning model is designed for self-adaptation control of wastewater treatment systems.for activated sluge treatment process,this model learns the inherent relationships between control actions and changes of the system state,by interacting with the sewage treatment system gradually,and finally achieve the sable operation of the whole system by providing optimal control of dissolved oxygen in aeration tanks and reflux flow.In the simulation experiment on the BSM1 platform,after a learning period around 150 days,the model can spontaneously control the dissolved oxygen set points and internal reflux to ensure concentrations of ammonia,total nitrogen and Chemical Oxygen Demand(COD)in effluent meet the requirements.Compared to the fixed control strategy,under the control of this model,the average dissolve oxygen in the forth tank as reduced by 33%,the average of internal reflux as reduced by 32%,but the average ammonia in effluent is lower and more stable.In the study of self-adaptive control of laboratory scalled reactor for SFAO~4 micro-aeration treatment process,the model can provide realible guidence for the ditermination of dissolved oxygen set points and external reflux after a learning period around 80 days.Under the control of this model,the reactor operated stably:the effluent COD concentration was less than 250 mg/L;the ammonia concentration was less than 40 mg/L;the total nitrogen was less than 80 mg/L.