Research On Modeling Of Spatiotemporal Distributed Systems Based On Deep Learning

Process data in industrial systems is highly complex,substantially nonlinear,and spatiotemporally coupling.Spatiotemporal distributed systems(SDSs)are a concept that researchers have developed to evaluate and track such manufacturing processes.The first principle,which states that sufficient physico-chemical knowledge must be attained before developing a model,is the basis of the conventional modeling methodology.However,it is challenging to find realistic models,particularly for processes with complicated structures,as the complexity of process production flows increases.Data-driven modeling techniques have increasingly become more popular as a result of advancements in computer technology.This paper proposes a deep learning-based modeling method for spatiotemporal distributed systems and conducts research in this area to address the current linear or non-linear statistical learning methods,which still have the flaws of hard-to-determine parameters and weak generalization ability with low accuracy.This work has the following aspects:Prior to figure out the mathematical partial differential equation models of SDSs,we first conduct a thorough background analysis based on reference material and the current state of domestic and international research.Then,we introduce the theory of spatiotemporal decomposition,model reduction,sequence prediction,and spatiotemporal reconstruction.The main steps of the current applied methods and their advantages and disadvantages are summarized,and ideas for improvement are explored for the above problems.Secondly,a joint learning approach of sparse autoencoder coupled with gated recurrent units for modeling SDSs is proposed for the temporary systems with high spatiotemporal data dimensionality and strong coupling of spatiotemporal properties.The proposed method is shown to be able to resolve overfitting,achieve accurate modeling,and obtain lower errors in comparison experiments with traditional principal component analysis,non-linear principal component-radial basis functions,and the same type of autoencoder recurrent neural networks.Lastly,an approach for modeling SDSs that combines long short-term memory networks and convolutional neural networks is suggested.The convolutional operation differs from the previous fully connected layer in that it has a small perceptual domain,allowing the network to focus less on the overall connection and more on the influence of the local domain of space,which is more in line with the characteristics of spatiotemporal distributed systems,and with the long short term memory,a more accurate model is eventually obtained,and the comparison with traditional machine learning also shows that the proposed method are efficient and accurate.Finally,two typical examples of chemical industrial processes,a catalytic rod and a tubular reactor with cycling,are used for simulation and validation,showing that deep learning combined with industrial process modeling has realistic implications and excellent application development prospects.At the same time,this dissertation also points out some directions for future work to improve and hopefully continue to extend this research.