Research On The Data Driven Soft Sensor For BOD Of The Wastewater Treatment Process

Since the 21st century,along with the rapid development of national industrialization,the discharge of industrial sewage and domestic sewage has also been growing continuously,which made the water pollution problem of our country to develop very serious,therefore,the problem of sewage treatment has also been paid attention to gradually by people.How to effectively treat sewage is related to the survival and development of people in the future.Biochemical Oxygen Demand（BOD）is through to reflect the total amount of oxygen required for the conversion process of organic matter to inorganic salts in the water body,which is an important indicator of water pollution and is also an important indicator for the protection and governance of water environment.The accurate measurement of BOD is necessary to ensure the normal operation of sewage treatment system and the quality of effluent up to the standard.Therefore,this thesis intends to study a data-driven soft sensing method for BOD to solve the problem of accurate measurement of BOD.First of all,an improved Grey Wolf Optimization based on mixed strategies（GWOM）and the Stacked Denoising Auto Encoder（SDAE）is proposed to carry out feature dimension reduction and feature extraction for auxiliary variables in the sewage treatment process,and secondly the extracted features with certain weights are input into the Elman neural network model optimized by GWOM to achieve accurate prediction of BOD.The research work in this thesis mainly includes the following points:（1）After in-depth study of the search mechanism of Grey Wolf Optimization（GWO）algorithm,Liebovitch chaos initialization was introduced to make the initial population distribution in a more reasonable way,aiming at the shortcomings of GWO algorithm such as lack of population diversity,insufficient local search ability and unstable performance.A convergence factor based on the decay of power index function is proposed,which makes the algorithm search smoothly in the early stage,converge quickly in the late stage and refine the precision step by step.The spiral search strategy of whale optimization algorithm was introduced,and then the adaptive factor was added to propose the variable spiral search strategy.Then combined with the hunting strategy of GWO,the mixed hunting strategy based on the improved variable spiral search strategy was obtained.The combination of Cauchy perturbation strategy and OBL reverse learning strategy is introduced to perturb the optimal solution,which can help the algorithm to escape from the local optimal to a certain extent.So the hybrid strategy based grey wolf optimization（GWOM）.The performance of the proposed algorithm is verified by several standard test functions of the stability and effectiveness of the GWOM algorithm.（2）The high-precision feature extraction strategy of data under sparse constraints was studied.The parameters of the Activated Sludge Model 1（ASM1）optimized by GWOM algorithm are compare with the optimization results of GWO,WOA and ASOA algorithms.The simulated data set generated by the optimized ASM1 model and the real sparse water quality data set of the Indian Ganges river were combined to train the SDAE feature extraction model.Finally,the operational data of an industrial wastewater treatment plant in Gansu Province was used to make BPNN-based regression prediction by comparing various feature selection algorithms,and the feature extraction capability of the SDAE model trained under sparse constraints was tested.（3）Aiming at the problem that BOD5,an important effluent parameter in the wastewater treatment process,is difficult to be directly measured,a combined prediction model based on SDAE-Elman is proposed,while the GWOM is used to optimize the connection weight of the SDAE-Elman neural network.When dealing with the fitting relationship between BOD₅concentration and auxiliary variables with high complexity and nonlinear,the model takes advantage of the accurate reconstruction of high dimensional data by SDAE deep learning model,and combines with the advantages of GWOM algorithm in the stable search in the early stage,rapid convergence in the late stage and gradual refinement of accuracy to extract the deep correlation features of data and eliminate invalid redundant information.Finally,the compact-structured Elman neural network with the well designed GWOM algorithm performed the high precision prediction of BOD5 concentration.Compared with the back propagation neural network（BPNN）,the Elman neural network and the SDAE-Elman model with average weight connection,the SDAE-Elman model optimized by GWOM for weight distribution effectively improved the prediction accuracy of BOD5 concentration.