Research On Data-driven Intelligent Control Methods For Urban Wastewater Treatment Process

Nowadays,urban wastewater treatment has been an effective way to prevent and control water pollution,promote the recycling of water resources,and improve the urban ecological environment.Moreover,urban wastewater treatment process（WWTP）control is critical to improve the operation efficiency of wastewater treatment and ensure the effluent quality meets the discharge standard.Nevertheless,the activated sludge treatment processes,which involve complex biochemical reactions and physical treatment processes,are widely used in urban wastewater treatment plants.Consequently,the urban WWTP control faces many challenges,which can be described as the following four aspects:1)Due to the complex microbial chemical reactions,the urban WWTP is characterized by strong nonlinearity and dynamics,so it is difficult to achieve an accurate mathematical model.2)The key effluent indexes（such as ammonia nitrogen concentration,total phosphorus concentration,etc.）of urban WWTP are difficult to be accurately measured online.Hence,the operation status of some key processes cannot be accurately identified in real-time.3)The measurement noise and outliers are inevitable.in urban WWTP with a harsh measurement environment.Hence,process data show non-Gaussian characteristics,which challenges the reliability of data-based modeling and control methods.4)The strong uncertainties in the influent flow and pollutants,as well as the coupling phenomena among multi-unit processes,lead to great challenges to the accurate and stable.control of urban WWTP.To solve the above problems,based on the detailed analysis of the mechanism and operating characteristics of urban WWTP,the multi-layer perceptron-based features selection model for key effluent quality;the offline modeling method for the effluent ammonia-nitrogen concentration;the online modeling method for key effluent quality prediction;the self-organizing control method for process variables are studied.The main research and innovations of the thesis are as follows:（1）Feature variable selection of key effluent quality in urban WWTPIt is difficult to select feature variables of the key effluent qualitiy in urban WWTP.To solve the problem,a feature variable selection model is developed based on the backwards-selection multi-layer perceptron（MLP）.Firstly,several backward selection criteria are designed by evaluating the input feature attributes,which can effectively measure the impact of input features on MLP output.Secondly,the input feature pruning algorithm is designed for MLP,which can automatically select the relevant features in the MLP training process.Thirdly,the initial model input variables are selected based on the mechanism analysis of key effluent quality.Then,the feature selection model for key effluent quality is established.The experimental results show that the proposed feature selection model can automatically obtain the feature variables of key effluent quality,effectively reduce the model complexity,and play an important role in the accurate measurement of the key effluent quality.（2）An offline modeling method for effluent ammonia-nitrogen prediction in urban WWTPIt is difficult to estimate the effluent ammonia nitrogen（NH₄-N）online.Moreover,the mean square error（MSE）-baserd modeling methodst are vulnerable to the nonzero-mean noise in the process data.To address these issues,the modeling-error probability density function-based fuzzy neural network（PDF-FNN）is proposed for estimating the effluent NH₄-N.Firstly,the optimal error-PDF shape criterion is generated for adjusting the model parameters.Secondly,an adaptive parameter learning method is designed to improve the model performance.Then,the convergence of PDF-FNN is analyzed in theory,which improves the reliability of the model.Finally,the feature model for effluent NH₄-N concentration is constructed offline,which realizes the real-time estimation of the effluent NH₄-N concentration in urban WWTP,and improves the robustness of the feature model to nonzero-mean noise.（3）An online modeling method for key effluent quality prediction in urban WWTPConsidering the nonstationarity in urban WWTP,a robust online self-organization fuzzy neural network（ROS-FNN）is proposed to mitigation the model degradation and outlier-disturbance for predicting the key effluent quality.Firstly,the structure learning algorithm is proposed based on the online error compensation method and correntropy criterion.Specifically,fuzzy rules of ROS-FNN are automatically generated by compensating large errors at the extreme points,thus improves the modeling accuracy;meanwhile,redundant rules are deleted by evaluating online modeling performance,thus improves the compactness the network structure.Secondly,the parameter adaptive learning method is proposed based on a modified correntropy-induced criterion.Moreover,the correntropy-induced criterion,which is effective to suppress the outliers,is useful to improve the model accuracy.Finally,based on the ROS-FNN,taking the effluent NH₄-N concentration as an example,the effluent quality feature model is established online.The results show that the proposed model can adaptively adjust its structure and parameters.Moreover,the proposed method performs better than other involved methods in terms of modeling time,accuracy and robustness for predicting effluent NH₄-N concentration.（4）A self-organizing control method for dissolved oxygen concentration in urban WWTPTo deal with the nonlinearity,uncertainty and non-Gaussianity of urban WWTP,the data-driven online self-organizing control method is proposed for controlling the dissolved oxygen concentration.Firstly,the correntropy-based fuzzy neural network（CSOFNN）controller is designed.Specifically,the structure and parameters of CSOFNN can be automatically created or pruned based on the correntropy and rules-contribution criteria.Secondly,the compensation controller and parameter adaptive laws are developed,which makes full use of the ability of correntropy to suppress non-Gaussian noise,and can effectively reduce the uncertainty of the system.Thirdly,the stability of the proposed control method is theoretically analyzed to guarantee its feasibility in practice.Finally,the experimental results show the effectiveness of the proposed control method.（5）Robust multivariable self-organizing control method for urban WWTPFor urban WWTP,due to the nonlinearity,uncertainty and coupling between multi-unit processes,it is difficult to ensure the stability and accuracy of multivariable control.To tackle the above issues,a data-driven multivariable self-organizing control method is proposed in the paper.Firstly,the FNN,which has the ability of nonlinear approximation and dealing with uncertain information,is selected to build the multivariable controller.Secondly,considering the coupling problem of multivariable control,a structure adaptive adjustment algorithm is designed for the multi-input multi-output fuzzy neural network（MIMO-FNN）.Based on the internal and external information of the MIMO-FNN controller,the network structure is self-adjusted to ensure the adaptability of the multivariable control system.Then,to deal with the external disturbance,the correntropy-induced criterion,which is able to restrict large errors,is designed for the multivariable control system.Moreover,the parameter adaptive laws and multivariable compensation controller are developed based on the correntropy-induced criterion.Furthermore,the stability analysis of the control system is given.Finally,the experimental results based on BSM1 show that the proposed method can effectively improve the control performance of dissolved oxygen concentration and nitrate concentration,and realize the stable.and accurate multivariable control.