Study On Kernel Function Modeling Method For Unsteady Process Of Wastewater Treatment

Protecting the environment and preventing corresponding pollution is one of the major challenges of China.This means that the wastewater treatment standards need to be continuously enhanced and the treatment processes need to be continuously updated,implying that higher requirements will be posed for the online detection and optimal operations of wastewater treatment plants(WWTPs).However,existence of many difficult-to-measure but quality-relevant variables adds much complexity of WWTP management.To deal with this problem,soft-sensors are established to predict important quality variables,such as total nitrogen and total ammonia by using existing process variables that are easily measurable,in such way that it can guide the process control for water quality and the fault prediction of wastewater treatment,which can further optimize wastewater treatment process management.However,the WWTPs are always a typical nonlinear,multivariable,unstable and time-varying system.Furthermore,lack of data and strong variable coupling make it difficult for modeling,usually resulting in rapid degradation of predictive performance.These make a huge challenge to build a soft sensor model properly.Therefore,the soft sensors in WWTPs have been widely concerned by academic and industrial communities from both domestic and overseas.The main purpose of this paper is to study on the kernel-based soft sensor modeling.The main contributions of this paper are that Multi-kernel learning is introduced into the relevance vector machine model and the adaptive technologies are introduced into the multi-output soft sensor model to predict difficult-to-measure quality variables online in a dynamic process.Meanwhile,a multi-output multi-step prediction model applied in wastewater treatment is studied for the first time.The main research contents of this paper are shown as follows:1.Learning soft sensors using time difference-based multi-kernel relevance vector machine with applications for quality-relevant monitoring in wastewater treatment.Considering the time-varying,uncertain and nonlinear properties of the WWTPs,a multi-kernel relevance vector machine(MRVM)is proposed for soft-sensor modeling.The model is not only suitable for little amount of sample modeling but also has a good generalization ability to overcome the problems of under-fitting and over-fitting effectively.A particle swarm optimization(PSO)algorithm is further used to optimize multi-kernel weights and kernel parameters.Meanwhile,the time difference(TD)method is introduced to improve the dynamic characteristics of MRVM.Results showed that the proposed model achieved better prediction accuracy.The robustness of the models is discussed in the context of data drifting and abrupt changes.2.Study on adaptive multi-output soft-sensers with applications in wastewater treatment.In the current,multi-output soft-sensing modeling of wastewater treatment is limited to offline and online test modes.Some negative effects like missing data,abrupt changes,drifting errors and measurement delays may lead to prediction degradation sequentially.Thus,some adaptive strategies,including just-in-time learning(JITL),time difference(TD),and moving window(MW)methods have been proposed in this paper to enhance multi-output soft sensor models to ensure online prediction for a large amount of hard-to-measure variables simultaneously.3.Research on online adaptive multi-output modeling feature extraction method: for the selection of adaptive multi-output input variables,an in-depth study is made to improve the prediction accuracy of multi-output models,and a time difference based on kernel canonical correlation analysis-instant multi-output Gaussian is proposed process regression soft-sensing model(KCCA-TD-JIT-MGPR)predict difficult-to-measure quality-related variables in wastewater treatment plants.Kernel canonical correlation analysis maximizes the correlation between the input and the corresponding target and is used to select the most relevant input variable set for the target at the same time.The combination of jet lag and real-time learning methods can not only weaken the negative impact of uncertainty but also effectively improve the robust performance and predictability of the model.In order to select the best parameters of MGPR,a simulated annealing firefly algorithm was introduced to optimize the covariance function of the model and help to select relevant parameters.The results show that the proposed model can effectively predict many difficult-to-measure variables of unsteady wastewater treatment process.4.Research on the application of multivariable multi-step multi-output prediction in WWTP.In order to deal with the problem of equipment failure in the sewage treatment process,resulting in inaccurate data and abnormal changes,the collected abnormal data will be followed by subsequent monitoring,causing greater errors.An online multi-output multi-step prediction model is proposed to directly The recursive strategy is combined with the multi-output model to perform multi-step prediction on multiple variables that are prone to failure.The results show that the multi-output multi-step prediction model can effectively predict the steady-state process.Finally,the work is summarized,and the future research on soft sensors for the wastewater treatment process has prospected.