Application Of Support Vector Machines In Chemical Process Modeling

Chemical process modeling is an important method to study these processes and plays a fundamental role in process simulation, optimization and control. Modeling by mechanism usually describes the process by some mathematical equations based on analyzing process principles and making some assumptions and simplifications. However, most chemical processes have very complex mechanism, a lot of influencing factors and high severe nonlinearities, so modeling by mechanism is difficult to describe these processes. In these cases empirical modeling is adopted. As a typical kind of empirical modeling method Artificial Neural Networks (ANN) has been applied to many chemical problems for its good performance in solving nonlinear problems. But ANN has some disadvantages such as overfitting, local minimum, etc. because its theory is based on Empirical Risk Minimization (ERM) principle. Support Vector Machines (SVM) is a new learning method based on Statistical Learning Theory (SLT). SVM based on Structural Risk Minimization (SRM) principle overcomes ANN's inherent disadvantages and greatly improves models' generalization ability. In this thesis we discuss the application of SVM in modeling chemical processes in detail.Firstly, knowledge about process modeling is introduced and then ANN's characteristics, structure and implement are reviewed. After analyzing ANN's performance SVM's theoretical basis, computing process and optimization algorithms are demonstrated detailedly.This thesis mainly studies SVM's application in chemical processes. A model to predict final concentration of citric acid in fermentation process is set up by SVM and the model parameters' effect on the model performance is carefully studied. In addition this model is compared with the traditional one based on ANN. SVM is applied to predict the biomass state in the alcohol batch fermentation process. Combing the process mechanism reflected by kinetics equations and high generalization ability from SVM, Hybrid Support Vector Machines (HSVM) modelis proposed in serial and serial-parallel approaches. These two kinds of hybrid models are compared with black-box SVM model, simple and approximate kinetics model and the corresponding hybrid ANN model. The experimental results show SVM is able to model fermentation process effectively and efficiently and HSVM has a good performance.Finally a summary of this thesis is given and a further research about S VM's application in modeling chemical processes in the future is prospected.