Research On Data-driven Spatiotemporal Modeling Methods Of Distributed Parameter Systems

Distributed parameter systems(DPS)exist in a wide range of engineering fields,including semiconductor manufacturing,biomedical Engineering,chemical synthesis,materials engineering,etc.The energy and mass of these systems distribute in space widely and change in time continuously,which brings in challenges for modeling the DPS in parameter prediction,controlling design and process optimization.Traditionally,these systems are modeled by partial differential equations.However,due to the large scale of the system,the complex working mechanism and the large number of control variables,it is difficult to obtain the partial differential equation models of the DPS accurately.Currently,it is known that data-driven modeling methods do not need knowledge of mechanism and internal structure of the industrial processes,and rely on the collected industrial process data only.Data-driven methods are black-box modeling methods which become effective means of modeling DPS.In this paper,the time coefficient model and time/space separation method for modeling of complex nonlinear DPS are studied by using the data-driven method and time/space separation method.The main work of the thesis are as follows:The performance of the time coefficient model of DPS is usually affected by the redundancy of input variables and model structure.To that end,a time coefficient model with the double compression of the variable and structure is proposed.By combining the advantages of feature selection of least absolute shrinkage and selection operator(LASSO)and nonlinear character of multilayer perceptron(MLP),the redundancy of the feature and structure of the model is reduced and the model accuracy is improved.The method is validated by data from simulation examples as well as an industrial application.To deal with the problems with dual nonlinear structures and dynamic-changing characteristics,a dynamic dual-structure time coefficient modeling method based on neural network is proposed in the process of spatiotemporal modeling of DPS.The novel dynamic hybrid time series model works by connecting MLP with long short-term memory(LSTM)neural networks in parallel to deal with the two inherently coupled nonlinearities,where the influence of historical information on the predicted values is also considered.The effectiveness of the proposed model is validated by the data from a snap curing oven thermal process.The satisfactory agreement of the current model shows that the spatiotemporal model based on the proposed method is reliable for estimating the temperature distributions during the thermal process.A two-scale time/space separation method considering both local and global characteristics is proposed to deal with the problems of missing the nonlinear characteristics in the complex DPS.The proposed method is a global two-scale nonlinear spatial basis function learning method,which can preserve nonlinear structure characteristics on both local and global scales during model reduction.The reliability of the spatiotemporal model based on the local and global two-scale time/space separation method is validated by the thermal process experiments of lithium-ion batteries.For DPS with the characteristics of large range and multiple conditions,the traditional single global spatiotemporal models are not able to simulate the whole process of this kind system accurately.To address this problem,the multi-model time/space separation method based on density peaks clustering(DPC)is constructed,where the time and space variables could be divided by the DPC algorithm into several subspaces representing local characteristics of time and space of the original system.Then,all the local spatiotemporal models are established respectively,and the integrated model is obtained by the weighted sum of the local spatiotemporal models using the LASSO regression algorithm to approximate the original system.In order to validate the model,the heat exothermic catalytic reaction process of the catalytic rod is employed to carry out experimental simulation and analysis,and the reliability of the multi-model time/space separation method based on DPC is verified.