Researches On Hybrid Modeling Of Chemical Processes

Process modeling is the basis of process control and process optimization. The becoming severely market competition and environmental requirements force companies to improve their productivity and efficiency, which poses new requirements on process control and process optimization. As a result, more severe requirements are posed on process modeling. Since conventional black-box modeling approaches rely simply on single data resource, their ability to describe the real process is limited. Hybrid modeling algorithms seek to combine every useful information in a uniform framework to improve the capacity of the hybrid model. This dissertation focuses on new identification algorithms, new hybrid model structures and their applications. 1. A neural-network-based fuzzy system (NNFS) is proposed. It is a self- organizing neural-network which can partition the input spaces in a flexible way based on the distribution of training data set in order to reduce the number of rules without any loss of modeling accuracy. Associated with the NNFS is a two-phase hybrid learning algorithm, which utilizes a nearest neighborhood clustering scheme for both structure learning and initial parameters setting and a gradient descent method for fine tuning the parameters of the NNFS. By combining the above two methods, the learning speed converges much faster than the original back-propagation algorithm. Simulation result suggests that the NNFS has merits of simple structure, fast learning speed, few fuzzy logic rules and relatively high modeling accuracy. The NNFS is applied to the construction of a soft-sensor for an industrial propylene-propane distillation column. 2. If the training data are highly noisy, the constructed model will fit into noise. To deal with this situation, the PLS (partial least squares) algorithm is introduced into radial basis function (RBF) networks to construct MIMO nonlinear models. The PLS algorithm projects the correlated basis functions and outputs down to a number of principal factors to construct a parsimonious model. Examples of nonlinear modeling and identification are used to demonstrate better generalization and noise tolerable performance of the ~1 VI proposed algorithm than the full RBF network model trained by the conventional least squares algorithm. 3. A nonlinear adaptive predictive control strategy (NAPC) based on fuzzy-linear combined models is proposed fur 5150 systems. The fuzzy-linear combined model is used to represent the forward system dynamics. The model structure and its on-line recursive identification algorithm are discussed. It is proved that it can approximate the dynamics of any plant to arbitrary accuracy. A constrained nonlinear optimization approach using simplex search is used to generate the optimum control law. Simulation result of pH neutralization process demonstrates its ability to outperform nonlinear adaptive predictive control strategy based on fuzzy models and predictive control strategy based on linear models, especially in the case of sparse training data. At last, NAPC is applied to a nonlinear liquid level control of a laboratory tank. Experiment result indicates that NAPC can exceed the performance of conventional P11) controller. 4. A grey-box fuzzy modeling approach is prc1~osed. Using fuzzy set theory, it tempts to combine the expert experience, local...