Intelligent Control of Large-Scale Industrial Systems

Background

The EPSRC have granted a 3-year CASE industrial studentship to the Electrical and Electronic Engineering Department at Napier University. It is a collaboration with Performance Improvements (PI) who are an Aberdeen-based company providing process and control engineering consultancy to the oil and gas industry. PI engineers have identified, through discussion with major North Sea operators, a requirement, and niche market, for "smart products" to improve existing plant performance. This efficiency will lead to:

Improved oil recovery; Increased transportation efficiency; and, Make the processes more environmentally friendly.

PI require to develop these ideas, through research and development, primarily in the areas of application of artificial neural networks and fuzzy logic, to improve oil and gas production and transportation. They have identified potential operators who have expressed an interest in these products and may be willing to work with PI and Napier University in providing expert knowledge for their existing production facilities. Napier University has been involved with many oil and gas companies in the past and has developed systems which optimise their control systems, including extensive work with British Gas Transco. Using this experience, Napier and PI will investigate fuzzy logic and artificial neural network control of large industrial control systems and over the period of 3 years will develop a range of software systems which will automatically learn industrial control environments. A major objective of the system is to provide an extremely safe working environment and also one which will work with existing on-site systems. This developed system will automatically generate a set of rules in which to control the plant. Initial work has already been conducted on fuzzy logic control of Turbocompressors and has proved successful. The projected saving in operating costs on the Turbocompressor optimisation project is £600,000 per annum with a reduction of 10,000 tons of C02. If this cost saving is projected to other oil and gas systems then the savings, in general, would be considerable in terms of cost and environment emissions.

ABSTRACT

This paper reviews the application of control strategies for large-scale industrial systems using Artificial/Computational Intelligence (AI/CI) methods. The application of Fuzzy Logic (FL) and Neural Networks (NN) control methods to non-linear control is rapidly developing field. In many applications these still have to be proved as robust technologies, as existing control strategies are based on discrete, distributed controllers. Most currently implemented systems are run in tandem with existing control strategies, which allows plant operators to select which control action to follow. Due to computational requirements, implementing FL control and NN technologies presently imply a reversion to large SCADA-based, centralised control systems.

Background

This paper describes a programme of research which is being undertaken through the Department of Electrical and Electronic Engineering at Napier University, Edinburgh. Funding has been granted through an EPSRC Case Award No. 97595937, and is entitled "Neural Network Control of Large-Scale Industrial Systems". Some of this work is commercially sensitive and certain applications cannot be divulged at the present.

The work carries on from previous work on fuzzy logic control applied to large-scale industrial plant optimisation [1]. In this, fuzzy logic rules were applied to a large-scale industrial system for the supply of conditioned gas. The implemented system was successful and had main advantages, such as:

• Increased fuel efficiency. A small increase in control system efficiency leads to large reductions in emissions and fuel usage.
• Simple rule-based system. The rules were generated through extensive discussions with plant operators (the experts), and were considerably easier to understand than the traditional control methods. This allowed the system to be easily understood by plant operators, technicians and engineers.
• Fixed rules. Rules were constructed so that safety was the strongest rule and plant optimisation was the weakness. From this safety always overrules plant optimisation.

The disadvantages of such the system where:

• The system was based on fixed rules, which did not change with plant operation. On a large industrial plant many factors can change, over time.

Thus further efficiencies could be achieved by some method of on-line, intelligent control, possibly using neural networks. This would allow the system to learn optimum operating points and make adjusts for changes in the system.

Introduction

Control systems are often required to deal with complex applications, sometimes having several process variables which require non-linear solutions. Such applications may have been traditionally solved using skilled operators or complex model-based software, as illustrated in Figure 1. Both solutions are costly, as the former requires operator training and the presence of a skilled operator and the latter requires sophisticated and costly software, and specialist support. Previous research [1] of this nature has been undertaken at Napier University.

Figure 1: Plant control

Botros [2], identified that advances in industrial control would be through, principally, "…incremental improvements in existing technology", such as the implementation of different computational techniques. These could include the application of Neural Networks and Fuzzy Logic. The learning properties of Neural Networks and heuristics of Fuzzy Logic based controllers allows the control of plant where formal knowledge of the plant process may be limited, but where data describing plant performance may be abundant. Typical applications may include:

• Load balancing of energy generation and transmission plant.
• Predicting and controlling plant responses to changes in load demand.
• Selection of optimal plant operating point.

The aim of implementing plant control using AI techniques is to improve business performance [3], this can be through improved process management, such as:

• Improved scheduling.
• Improved planning.
• Increased optimisation.

These allow for more reliable strategic decision-making and better-informed decisions, which should make for better business performance.

Generic Plant Control Structure

A hierarchical plant control structure can be represented as high, intermediate and low-level tasks. Each of these tasks can be grouped into five areas:

• Plant management.
• Planning and scheduling.
• Advanced control.
• Conventional control.
• Data acquisition and reconciliation.

These areas are interdependent, that is, an instrument failure could impact on higher activity level tasks such as maintenance planning or production targets. Similarly, managerial decisions may affect operating procedures and plant performance at intermediate or lower levels.

Figure 2 shows a typical hardware hierarchy. The stronger the information chain through these levels, the less the likelihood that plant may run sub-optimally. Hence, each level requires knowledge from the other. In addition, the higher levels require external knowledge such as sales and marketing information.

Figure 2: Plant hierarchy

Intelligent Control Functionality

In modern SCADA-based plant control systems there are areas of overlap between functions which were once implemented by discrete controllers. Now a PID function might be implemented using SCADA software. Fuzzy logic-based rules can be implemented in a PLC, as well as in SCADA software modules. Figure 3 shows the overlaps.

Figure 3: Technology overlaps

It can be seen from Figure 3 that the information becomes available to the operator at the SCADA operator interface. Plant performance is then very much dependent on the operator’s ability to assimilate information form many sources, Then to manipulate controllers to achieve production targets. As a secondary task the operator may then attempt to adjust the plant to run optimally.

For a given set of operating conditions the operator must have prior knowledge or experience of these conditions to know controlled variable settings for optimal operation of the plant. Figure 4 illustrates the learning or training process which the operator is likely to have gone through.

Differences between the actual and expected performances generate a deviation signal. This signal could be due to non-conformance of the plant behaviour after equipment failure, poor plant modelling or lack of understanding of plant processes. The deviation signal might imply change controller settings or re-training the operator.

Figure 4: Operator learning process

Knowledge against Data-Driven Control Strategies

Operator training results in a knowledge-based, expert system, that is, the operator. After training the operator is able to make causal connections between observed plant conditions and plant performance, usually in terms of production or yield performance. With experience of plant operation the operator is able to adjust key process parameters to influence or control plant behaviour. The data from the plant, fed into or calculated within the SCADA, can be used to predict or model plant performance as a function of key variables. Modelling can be achieved by applying mathematical laws to the data or using statistical techniques, such as multivariable regression. Highly non-linear and highly interactive processes are very difficult to model, which results in poor accuracy and the predicative qualities of models.

Neural networks are being increasingly used to predict non-linear plant performance, and are trained to implement control functions. The notion of training implies a supervised learning category of Neural Networks.

An overview of a Supervised Training Network

Generally a feed-forward network, using a back-propagation learning algorithm is used for control type applications. Figure 5 shows an outline of a Hierarchical Fuzzy-Neural controller.

The Neural Network is trained on a data set which consists of an input data set and corresponding training targets. The network is presented with an input data set and its corresponding output is compared with the training target. The difference between the actual output and the target is reduced using back propagation. This technique minimises the error between the current layer and preceding layer of the network in successive steps until the input layer is reached.

A feed-forward, back-propagation trained network is also referred to as a Multi-Layered Perceptron, MLP. More recently the use of clustered Radial Basis Function, RBF (for example, Gaussian) networks have been proposed.

Figure 5: Controller architecture

Generic Application of Neural Networks for Plant Control

A common implementation of Neural Network control is to use two networks, one to emulate the forward process and one to learn the reverse process. Plant inputs are fed into the emulator network and the network is trained to model selected plant parameters or performance descriptors. The reverse process network is trained using the performance descriptor output by the emulator to learn key set-points which will reproduce the performance descriptors modelled by the emulator network. This type of architecture allows the networks to be trained on-line, and is illustrated in Figure 6.

Figure 6: Operator Learning Process with Neural Networks

References