|
|
ADVANCED PROCESS CONTROL GOES NEURAL
Alex Conradie
The process industries (chemical, steel and mineral processing) face ever-increasing demands for higher production rates and greater product quality, while consuming minimal resources and ensuring minimal damage to the environment. The process industries are under greater pressure than any other industry to ensure continuous improvement, as economies of scale make small improvements in processes translate to large monetary savings.
There are generally three courses of action that may ensure company growth. In the short term, mergers and acquisitions offer a means of bringing about company growth. In the long term, research and development (R&D) offers a hedge over the competition, that may ensure positive net present values. Between these two time frames, the implementation of advanced process control in existing process facilities allows for significant growth in the medium term. Although the number of process units requiring some form of advanced control is small, these key unit operations may account for up to 40% of the gross revenue generated in the USA process industries (1998).
Most industrial controllers are designed for single variable processes. A single-variable or single-loop controller measures the one and only process variable (i.e., temperature), compares that measurement to the set point, and applies a corrective effort via a single actuator (i.e., a valve). For example, a chemical reactor's temperature may be measured, compared to the desired reaction temperature and a control valve may be utilised to increase or reduce the amount of cooling supplied to the reactor. Most single-loop controllers implement the linear PID (proportional-integral-derivative) algorithm to compute the desired control action. Linear PID loops control a majority of the automated processes in industrial facilities. The PID algorithm is both simple and reliable, and has been applied to hundreds of thousands of control loops over the last 50 years. However, not all industrial processes can be controlled with PID loops. Multivariable and non-linear processes all require more advanced control techniques.
Multivariable controllers are not new to industrial automation; only traditionally harder to use. Implementing a multivariable control system is a complex operation that has traditionally required expert assistance from the hardware vendor or a consultant. As a result, multivariable controllers have been relegated to high budget applications in control intensive industries such as petroleum refining, petrochemicals, pharmaceuticals, natural gas processing, and aerospace -- where single variable controllers just will not do. Considering the difficulties in implementing multivariable control, numerous control problems that require multivariable control are handled by many single-loop controllers that act independently. This independent control action by each PID controller may negatively impact on the desired control objective, through unwanted controller interference and uncoordinated control actions. A true multivariable controller, on the other hand, can balance the actions of several actuators that each effect several process variables simultaneously. A reliable and user-friendly means of developing a multivariable controller would thus be most beneficial to industry and positively affect the bottom line.
It is also well understood that that no linear controller (such as a PID controller) is able to match the performance of a rationally designed non-linear controller, controlling a non-linear process. The majority of chemical and mining processes are profoundly non-linear in nature. Along with this knowledge, it has also been realised that the current non-linear control theory unfortunately has been unable to provide effective solutions to modern day control challenges, making practical implementations almost impossible. This has brought about the scenario were even advanced process control implementations remain linear in their implementation, though the process may exhibit significant non-linear behaviour. Developing effective non-linear controllers with suitable robust performance in practical process environments, is thus highly sought.
A clear understanding of the economic advantages of both multivariable and non-linear process control, contributes to the conclusion that advanced process control serves as technology leverage for process companies wishing to achieve growth in the medium term. A Ph.D. student, Alex Conradie, and his supervisor and mentor, Professor Chris Aldrich, at the University of Stellenbosch have utilised evolutionary computing to further the search for the illusive easy-to-design combination of multivariable and non-linear control. Evolutionary reinforcement learning entails the use of a state-of-the-art genetic algorithm (optimisation algorithm) for the development of neurocontrollers (neural networks) for controlling multivariable and non-linear process plants. The SANE algorithm (symbiotic, adaptive neuro-evolution) allows for the development of optimal process controllers that are able to control existing process facilities with greater economic return than single-input-single-output, linear PID controllers.
All current advanced control products are based on algorithmic approaches (i.e., Model Predictive Control software), while the SANE algorithm is biologically motivated, using ideas from neuro-science (neural networks) and Darwinian evolution (genetic algorithms). The use of evolutionary reinforcement learning for neurocontroller development is thus a new process control technology. Although clearly superior to PID control, the SANE algorithm has also proven to be more effective than Model Predictive Control in rejecting process disturbances and operating in the face of greater process uncertainty. In terms of usability, Model Predictive Control requires a large degree of expert user intervention and engineering judgement during the controller development phase, whereas SANE automates the development process into a single comprehensive step.
The SANE algorithm has been coded into a user-friendly Windows 2000 software application. The software utilises COM technology, which allows for the software to communicate with various commercial modelling software packages, such as Matlab and Hysys. Matlab and Hysys are typically used to develop process models for control applications. The SANE software package takes full advantage of these developed process models, as opposed to linear controller development methodologies. In addition, process data from plant databases may be used to developed sophisticated non-linear dynamic models of the underlying process, using advanced time series analysis techniques. This enables process engineers to datamine vast plant databases, creating process knowledge from process information. The SANE algorithm thus ensures a plant-wide control approach, which allows for more effective control strategies and the integration of control operations.
|
