Modeling, estimation, and control of nonlinear time-variant complex processes

Mathematical modeling of physical phenomenon is an effective way to understand, analyze, and control complex physical processes in natural science and engineering disciplines. In reality, a majority part of physical processes is inherent nonlinear and time-variant. Linear dynamic models have been used pervasively in control engineering practice to model such processes, due to its simplicity. However, poor model predictions can be resulted in, if only linear models are used, since there may exist a big mismatch between the linear model and the real process. Unmeasured, time-varying external disturbances entering the system may even deteriorate the performance of the system.

In this research, the nonlinear modeling, estimation, and control problems for such nonlinear and time-variant processes have been studied. Two application studies are investigated. One is the control problem of a highly nonlinear pH neutralization process. The other one is the leak detection problem for a natural gas pipeline system using nonlinear state estimation methods. In the first application, the pH neutralization process is characterized by its high nonlinearity and time-varying properties. It has a wide operating range with large dynamics, which can be seen from its titration curves. A single nonlinear controller is not able to deal with all the possible operating conditions. Therefore, a multi-model predictive control (MMPC) strategy has been proposed to control the effluent pH of a pH neutralization tank at a setpoint value of 7. The recurrent neural network (RNN) models are used as the predictive models in the MMPC framework. A real-coded genetic algorithm with a novel simulated binary crossover operator (SBX-RCGA) is developed to tune the structure and the weights of the RNNs simultaneously. Such a novel design has the advantages of global accuracy and fast convergence. An offset-elimination measure has been incorporated into the MMPC strategy to compensate steady-state offsets. Simulation results show that the proposed strategy is able to reject all the severe disturbances (i.e., from pH=3 to pH=13), with an offset free performance.

In the second application, an efficient and reliable real-time leak detection system has been developed for a single natural gas pipeline system based on a dual unscented Kalman filter (DUKF) technique. Two first principle models of the pipeline flow are built. The first model, which has a leak introduced at one of the discretized nodes, is used to simulate the artificial measurement data. The second one, which does not include any leak, is nested in an online observer to predict the pressure and mass flow rate profiles under normal operating conditions. The discrepancies between these two date sets at the inlet and outlet ends of the pipeline are used to detect, locate, and estimate the leaks. Practically, modeling of a real process can never be 100% accurate.

The uncertainties in the process model need to be considered and addressed properly. Therefore, the DUKF technique is proposed in that regard. It provides state estimation, as well as parameter estimation. Comparison study with UKF shows the superiority of DUKF in increasing the leak detection and location accuracy.