Research And Application On Hybrid Modeling Approach For Chemical Processes

The success of model based control and optimization of chemical process is dependent on good process models. The main process modeling approaches include first-principle model, empirical model and hybrid model. Much for reasons of the difficulties experienced in the chemical dynamics and poorly known or unknown properties of a process, building first-principle models often requires a large amount time and resources. Empirical models can generally be developed very quickly without requiring detailed insight into the process. However, they are usually not as reliable as first-principle model for limitations on 'black box'. The combination of partially known process knowledge and 'black box' modeling results in hybrid model that is also called 'grey box' model. The dissertation focuses on research and application on hybrid modeling approach in chemical processes. The content is arranged as follows:Firstly, a model of a 5-kw air-hydrogen proton-exchange membrane fuel cell stationary power system is developed using a combination of mechanistic and empirical modeling techniques. Utilizing the relationship between fuel cell stack and single cell, the paper details the mechanistic descriptions of single cell and development of fuel cell stack including the thermodynamic equilibrium potential, activation overvoltage, ohmic resistance losses and empirical expression for resistively of single cell. A series of experiments were designed to obtain parameters in activation over-voltage equation. The model provides estimates of polarization curves of the fuel cell stack under different conditions. Its effectiveness has been proved through model validation and simulation studies.Secondly, a new hybrid neural networks modeling approach, named Structure Approaching Hybrid Neural Network (SAHNN) is proposed. The characteristics and structure of this approach are introduced in details. The approach which combines first-principles model with neural networks fully utilizes structural information of known nonlinear system, makes neural networks to be 'grey-box' , describes and explains the consequence relation of system variables in better way. A detailed analysis of SAHNN modeling has been performed. Considering delayed measurement of the process output, the approach to chemical batch reactor has been performed and makes comparison to two types of Hybrid Neural Network (HNN).The results of simulation and comparison illustrate that the approach is a promising tool to model complicated nonlinear system effectively and can be utilized as a vehicle to control and optimization of chemical reactors.Thirdly, an intelligent modeling approach, named Virtual Supervisor -Artificial Immune (VS-AI) algorithm, was developed for chemical processes with partial unmeasured states. VS algorithm is utilized to solve training problems of neural networks that some supervisors are not available for partial unmeasured states. By optimizing VS and weights of neural networks at the same time, the precision of modeling is improved dramatically. Immune population was used to explore new better solutions and avoid wrong direction in training. Detailed process of modeling and training of hybrid neural network was described in modeling of batch reaction of producing accelerant for sulfuring rubber. The simulation result proves that the approach is effective.Fourthly, a novel extended integral square error index is proposed. A typical complex exothermic batch reactor model was developed by Structure Approaching Hybrid Neural Networks (SAHNN). The optimal reactor temperature profiles are obtained via the PSO-SQP algorithm solving maximum product concentration problem based on recurrent neural network (RNN). Considering model-plant mismatches and unmeasured disturbances, a novel extended integral square error index (EISE) was proposed, which introduced mismatches of model-plant into the optimal control profile. The approach applies a feedback channel for the control and therefore dramatically enhanced the robustness and anti-disturbance performance. The stability analysis of the One-Step-Ahead control strategy through SAHNN-based model is described based on Lyapunov theory in details. The result fully demonstrates that the proposed optimal control profile is effective.Finally, the summary and perspectives of hybrid modeling approach are addressed.