Applications Of Support Vector Machine In Intelligent Modeling And Model Predictive Control

Model predictive control (MPC) has become one of the most successful advanced control techniques in the chemical and petrochemical process control industry recently. MPC refers to a class of control strategies, in which a dynamic process model is established to predict the future behavior of the unknown system and the system performance is further optimized according to the model. With the expansion of production scale and the increase of manufacturing complexity, mechanism-modeling approach has become more and more difficult. As a result, it is necessary to identify the unknown plant by using the available experimental or manufacturing data. In recent years, support vector machine (SVM) has been successfully applied in pattern recognition, system identification, etc. Therefore, SVM based intelligent modeling approach and the corresponding multi-step-ahead optimizing predictive control strategy for a class of complex plants with nonlinear behavior is studied in this dissertation.The major contributions of this thesis are summarized as follows: 1. For nonlinear multi-input/multi-output (MIMO) systems, the structure and algorithm of SVM with quadratic polynomial kernel function based one-step-ahead nonlinear model predictive control (NMPC) is proposed. By considering the coupling among the input and output components, each output component is modeled independently by employing SVM technique. And then all components' models are integrated into a vector form to derive the analytical equations related to the optimal one-step-ahead input components and the known inputs/outputs according to predictive control mechanism. And finally theanalytical equations are solved through numerical algorithm.2. For nonlinear MIMO systems, the structure and algorithm of parallel SVMs with quadratic polynomial kernel function based multi-step-ahead NMPC is proposed and the analytical equations with respect to the optimal multi-step-ahead input components and known inputs/outputs are derived. No error accumulation occurs because each SVM based multi-step-ahead predictive model is independent of each other, and a novel feedback correction strategy, suitable for parallel multi-step-ahead predictive models based MPC, is presented.3. SVM multi-classification based multi-model switching .strategy is proposed and consequently parallel SVMs based MPC (SVMs-MPC) is generalized to control the complex plants, which operate in multiple operating environments. Firstly, SVM modeling is implemented for each operating condition and the corresponding SVMs-MPC algorithm is developed, and then SVM multi-classification model is constructed for multiple operational conditions. In real-time control, SVM multi-classification model identifies the current environment and then activate the corresponding SVMs-MPC controller. The presented control strategy can provide a novel approach to the complex processes, whose operating conditions usually jump, by combining SVM classification with SVM regression well.4. The robust stability of parallel SVMs with linear kernel function based multi-step-ahead MPC is analyzed. According to SVM model, the sufficient and necessary stability condition for SVMs-MPC closed-loop is given. And then the constraint sets, which can guarantee the above stability is robust for model/plant mismatch within some given bounds, are derived by applying small-gain theorem. For the unknown linear and weak nonlinear plants, a closed-loop with a larger stable margin can be achievedby adjusting predictive control parameters according to the stability condition and the robust constraint sets.5. SVM with multi-kernel based model structure of nonlinear systems and its specific identification method is proposed, which is composed of a SVM with linear kernel function followed in series by a SVM with Spline kernel function. With the help of model structure, NMPC can be transformed to linear model predictive control (LMPC) and then a unified analytical solution of control law of multi-step-ahead predictive control is derived. This algorithm doesn't need online iterative optimization and be suitable for real-time control with less calculation.6. Short-term load forecasting based on SOM and SVM is proposed. SOM based clustering for the predicted day before SVM predicting can incorporate the advantage of SOM with that of SVM and consequently the precision and speed of the algorithm is improved. The influences of conventional data preprocessing and the SVM parameters on model precision are discussed.