Systems/Process Monitoring, Diagnostics and Control
Advanced (AI-Based) Nonlinear Controllers for Industrial Processes
The overall objective of this research is to explore and demonstrate the utility of artificial intelligence (AI) technologies for developing effective controllers for dynamical industrial processes. Research results obtained with AI technologies, in the form of artificial neural networks (ANNs) and fuzzy logic, have shown the potential for creating nonlinear multivariable controllers for complex industrial processes, without explicitly representing the underlying physical principles that govern process behavior. In addition, this approach allows for ease in handling state-space constraints, and it does not require linear approximations of the system performance, a requirement of most classical approaches that often distorts and fails to truly represent the real problem. The methods being investigated are expected to have very broad applicability because they permit the development of computer-based controllers even in cases where the process is poorly understood or far too complex to be modeled from first principles.
The research initiative includes the development of generalized, inductive, input/output data-driven algorithms for creating models of nonlinear dynamical processes, the development of AI-based suboptimal and optimal control algorithms, and the comparison of these algorithms with standard PI and PID controllers, as well as with classical optimal control methods to establish their efficacy. To date, ANNs have served as the nonlinear models of the process, and either ANNs or fuzzy logic methods have been used for the nonlinear process controllers.
High-fidelity ANN models of the process to be controlled are developed first by "training" them to accurately map process inputs onto process outputs on the basis of measured process data. With successful training, the ANN process model can accurately reproduce the measured process dynamics, even with minimal understanding of first principles. That model is then used for developing and tuning the AI-based controller--again by "training" the controller to manipulate the model inputs in a way that yields the desired model outputs. For real-world control, after the controller has been properly tuned, the process model is replaced by the actual process. The schematic below illustrates one of many possible neural network representations for closed-loop control of discrete-time dynamical processes. This representation has two basic components: a neutral network controller and a neural network model of the process. Given the desired process setpoint of target T and an arbitrarily selected initial process state y1, the controller provides a sequence of control actions u1,u2,...,uk,...,uK-1 that drives the dynamical system, represented by an ANN model of the process from y1 to yK=T. This representation offers the advantage that training of the neural controller requires only the provision of the desired process setpoint T, rather than the (generally unknown) control law uk. Training of the neural controller is achieved by minimizing the difference between T and yK.
This general approach was tested for industrial-scale plant control in an application that has the goal of limiting emissions from operating fossil power plants. Two projects which were conducted in this area had the purpose of developing controllers that improve coal power plant efficiency while reducing environmental impacts, such as the emissions of oxides of nitrogen (NOx) and carbon monoxide (CO). The objective of the first project, funded internally through Argonne's Laboratory Directed Research and Development Program, was to reduce NOx emissions through improved control of furnace tilt and air injection. Initial proof-of-principle results1-3 are encouraging regarding the feasibility of using ANNs following the representation shown above for both modeling the dynamic formation of NOx emissions in the furnace and for closed-loop control of NOx emissions with the routine industrial control variables. Future research on this project will develop control systems as illustrated in the schematic, except that the control actions will be provided by fuzzy-logic-based controllers instead of neural controllers. This will allow intercomparisons of the efficacy of the two AI control technologies. Further efforts will extend the scope of the research to development of optimal control algorithms, where the controllers will be required to reach the given process setpoint while at the same time optimizing a prespecified objective function related to operating cost.
In a second project, conducted in collaboration with a small consulting company and funded by a U.S. Department of Energy grant under the Small Business Technology Transfer Program (Phase I), the goal was to optimize the injection of natural gas in coal-fired power plants.4 When injected into the upper region of the plant's furnace, natural gas will "re-burn" the NOx in the flue gas, converting it to atmospheric nitrogen and thereby reducing NOx emissions. Natural gas is more expensive than coal; thus, the objective of this project was to develop controllers that achieve NOx reduction while at the same time optimizing the gas injection amplitude and spatial distribution to reduce average gas consumption and limit CO generation.
- J. Reifman, J. E. Vitela, E. E. Feldman, and T. Y. C. Wei, "Recurrent Neural Networks for NOx Prediction in Fossil Plants," Proceedings of the Computer Simulation Multiconference, New Orleans, Louisiana, April 8-11, 1996.
- J. Reifman and E. E. Feldman, "Identification and Control of NOx Emissions in Fossil Plants," Proceedings of the Computer Simulation Multiconference, Atlanta, Georgia, April 6-10, 1997.
- J. Reifman and E. E. Feldman, "Identification and Control of NOx Emissions Using Neural Networks," Journal of the Air & Waste Management Association, May 1997.
- J. Reifman, E. E. Feldman, T. Y. C. Wei, R. W. Glickert, J. M. Pratapas, and K. E. Wanninger, "An Intelligent Emissions Controller for Fuel Lean Gas Reburn," POWER-GEN International Conference, Orlando, Florida, December 9-11, 1998.
Last Modified: Thu, April 21, 2016 4:54 AM