01 October 2004
Meeting the challenge of complexity
Surging consumer, environmental, and economic demands compel plants to achieve a variety of technical, fiscal, and ecological objectives simultaneously.
Oftentimes these objectives conflict.
To illustrate, a plant must rapidly reach generation targets, remain within emissions limits, and generate power at minimum cost for extended periods of time. The emphasis on achieving several conflicting objectives in power plant operation poses a decision problem with multiple criteria.
A particular combination of plant inputs from an available operating space must come into play to achieve the best performance from a plantwide perspective. And, decisions must happen in real time to create value continually in the dynamic environment of electric power generation.
Control systems rooted in classical feedback techniques lack the power to solve these complex problems. Human operators with their experience and ingenuity can steer feedback control systems to the optimum point in the operating space.
Relying on people for this complex task is not the most productive approach because of human cognitive characteristics and constraints in workload. A prudent and popular approach is to deploy model predictive control (MPC) to direct the distributed control system (DCS).
MPC generally brings the capability to predict plant behavior, make preemptive control moves, and optimize plant performance. The MPC–DCS combination is under human supervision. Consequently, people perform decision and control functions.
An ensemble of sophisticated systems can certainly boost the "intelligence" in measurement and control as well as the capability of plant equipment. However, the value obtained from the ensemble is critically dependent upon three factors.
First is the ability of the MPC. For instance, an MPC that cannot effectively manage computational complexity, nonlinearity, and uncertainty is unlikely to add much value.
Second is the synchronized functioning of the individual systems. For example, an MPC cannot add any value if the DCS it is directing cannot maintain process variables at set point.
Third is the complexity people face while interacting with the systems. To illustrate, a high level of combinatorial or dynamic complexity can make people error-prone during operations.
Not all installations of MPC deliver the expected benefits. A primary cause for failure is the inability of the individual problem-solving agents or the combination of agents to manage the inherent complexity during operation. The origins of such inability usually reflect insufficient attention given to complexity during the design phase.
An optimal control solution to improve performance in multiple dimensions involves a higher level of mathematical and computational complexity. Obtaining the conditions for optimality requires mathematical manipulation of the system dynamic equations and the equation for performance measure. System dynamics relate controlled variables and manipulated variables. Naturally, these introduce constraints. The resulting equations are often nonlinear differential equations of second or higher order. Besides, they could be time varying. Furthermore, future information about the system's behavior and the performance measure is required. In other words, the reward must be clear before making the decision.
Evaluating performance on the basis of individual criteria can be relatively simple. But, determining what is best in a global sense presents greater complexity. Oftentimes a global performance criterion, such as cost expressed in mathematical form, is an aggregation of several distinct criteria, each of which epitomizes plant performance in a particular dimension. Some aspects of performance may be difficult to measure accurately or reliably. No solution may ever exist to entirely satisfy all the individual criteria. Compromises enter the picture. Ranking criteria influence the extent to which lower performance in one area is acceptable for higher performance in another.
Finding the right mix of plant inputs that will minimize or maximize a comprehensive performance measure involves combinatorial complexity, which quickly increases with the number of factors in decision-making. Only some combinations of plant inputs will be viable. Some others will be impractical. Still others will be ineffective. Similarly, some values of plant outputs and some plant states just can't happen. A further complication enters through the presence of global and local minimums in the search space. The challenges in "searching" for the best solution include identifying viable combinations, rejecting combinations that offer no improvement, and conducting the search to converge on a global minimum. Besides, the solution must exist within the bounds of physical realities. What's more, the solution must happen in real time.
People face increased dynamic complexity while interacting with different systems deployed in a hierarchy. Numerous relationships varying with time, frequent interactions over time, and outcomes that are difficult to predict, are the principal factors that contribute to dynamic complexity. Correlations between causes and effects may not be simple. Effects may be disproportional to causes or counterintuitive. And, effects may unfold over many time scales. System behavior can differ with operating point and time. Furthermore, a transition may force other transitions. Interdependency may propagate or amplify the effects of disturbances or failures. Emergent behavior is the result of a mélange of actions. Consequently, predicting overall behavior is much harder. And, static characteristics of a system offer little help to forecast dynamic behavior.
Work, people, and systems
Broadly, work in power plant operation includes monitoring plant performance, controlling the plant to achieve operating objectives, and enhancing plant performance. Work in plant operation demands consistency, precision, and speed. Identifying an abnormal situation as normal is totally unacceptable due to safety implications. Identifying a normal situation as abnormal is philosophically unacceptable even though plant safety is not threatened. It is also undesirable because later actions may lead the plant to suboptimal or even abnormal operation.
People and control systems function autonomously as well as cooperatively to perform the work essential to keep a plant running and to operate without incident. The two "orchestrate" work across time and space. They are spatially close and geographically dispersed. They influence each other and are interdependent. But their needs and capabilities with respect to communication, data processing and storage, decision making, and action–execution differ. Both are resource bound.
Control system behavior is primarily algorithmic. At a particular level of abstraction a collection of algorithms works together to achieve a higher-level objective. The "intelligence" in control systems comes from the algorithms input in advance by designers. It is adequate to manage operational tasks anticipated and addressed by designers. All too often machines may not cope with events unanticipated and not addressed by designers. For example, an MPC with no model between heat input and steam temperature is unlikely to exert control on steam temperature.
Humans, on the other hand, have superior abilities to cope with unanticipated situations. They accomplish operational tasks by automated behavior, conscious execution of procedures, and conscious problem solving. They can change strategies instantaneously when current strategy makes little sense, or new information refutes current beliefs. They can solve unusual problems by utilizing experience, intuition, imagination, and creativity. Put another way, they can enhance the control algorithm online in real time.
People as well as control systems work in a web of interrelationships and interdependencies. Interaction between machines is generally limited to data exchange. On the contrary, interaction between people almost always has a social dimension in addition to the exchange of information. Individual roles play out in a larger context of interconnected roles. People coordinate their actions, negotiate to resolve differences, and cooperate to achieve common objectives. Envy, selfishness, or altruism affects human decisions. In addition, organizational structure and culture influence human performance. As a result, conflicts and compromises occur.
Broad design activities
The focus here is control hierarchy as opposed to its components. Indeed, components provide a basis for overall control capability. For example, an MPC positioned at a higher level in the control hierarchy must manage higher levels of nonlinearity, computational complexity, and uncertainty compared to a DCS positioned at a lower level. That said, what must designers seek in an MPC? Key considerations include the following:
- Does the MPC have the means to capture nonlinear functions with a high degree of accuracy?
- Can the system capture nonlinearity without detailed specifications?
- How accurately can the mechanisms in the MPC represent plant behavior/structure?
- What is the MPC's tolerance to uncertainties originating from noisy data, disturbances, models, and novel situations?
- Is computing speed compatible with operational requirements?
- What adaptive capability does the MPC bring to the control hierarchy?
The foundation for a control hierarchy develops from the components and the activities of designers. A control hierarchy with a weak foundation has the propensity for cleavage between problem-solving agents and serious implications for stability. In addition to component selection and control loop design, these three broad activities are useful to improve the posture to meet the challenges of complexity:
- define the work
- divide the work
- organize to adapt
Define the work required
Understanding the work to be done is essential before it can be decided who will do it and how to do it. This activity is of primary importance because it lays the foundation for the scope and direction of subsequent activities. It also has huge consequences because it significantly molds the overall problem-solving structure and function.
A logical requirement to define the work involved in operating a system is to understand the purpose, structures, and functions of the plant. This knowledge provides a solid basis to define the work required to establish and maintain the plant's functions. Work in three broad areas is especially significant: (1) links within hierarchical layers, (2) links between hierarchical layers, and (3) links with the integrated system.
To a great extent, the material configuration, functional requirements, and natural laws determine the work and its pace. In addition to the invariant plant properties, social norms and the environment shape work. All these realities exist independently regardless of human perceptions, desires, and dislikes. They impose constraints on work and create the context for people's behavior.
Capture key plant characteristics that shape work. Minute details are often unnecessary. A general description of what is necessary is usually sufficient.
Broadly, work situations can classify into routine and unusual. Formal rules and procedures, embedded in regularity and familiarity, are generally adequate to handle routine situations. Much of the routine work is definable by proceeding from the overall purpose down to the individual functions. The approach also serves to validate the work by checking its contribution to the overall purpose.
By contrast, work in unusual situations is less predictable and difficult to formalize. Fixed rules and procedures might be inadequate to handle such situations. Additional work has to add on instantaneously in response to new demands. Situational criteria dictate the kind of work required and its pace. Typically, exploration is necessary to identify possibilities that may arise. In any case, distinguish between the work required and how a system works.
Three factors make defining work a difficult task. First, all work associated with the system may not be easily identifiable. Second, all people who have to interact with the system may not be able to state all their needs clearly. Third, all the influences on work, and the effects of work, may not be obvious.
Identify the pertinent people and get their perspectives of the work involved. Include their perceived difficulties and desires. Generally, people's interpretations of a problem have a greater influence on problem-solving responses than the problem itself. So, understand their cognitive biases. Do not ignore casual or naïve intermittent users. They can be major contributors to dangerous or costly incidents.
Mutual reliance interfaces
Next, the work identified must divide up between humans and automation systems. Put another way, functions must tag to humans and automation.
Dividing work requires decomposing along whole-part lines and assigning on the basis of relative strengths. Automation generally excels in executing preprogrammed solutions at high speed and precision. It makes decisions in a split second but without contemplation. However, when the underlying assumptions are invalid its behavior may be unexpected or outright "insane." Humans, on the other hand, can make decisions with a strategic sense. They can integrate disparate data, recognize patterns, and use experience and common sense. They can therefore improve existing solutions or generate new solutions in unstructured situations.
Division of work can meet certain objectives. One objective may be to minimize the human role to decrease the possibility of human error. Another may be to maximize the human role to increase the readiness for human intervention. In reality, neither objective is entirely applicable because of potentially serious side effects. To illustrate, the unfamiliarity or overload is likely to make humans error-prone during transient demands.
Understand the effects of automation on human performance. New encumbrances, errors, and work can arise. These in turn demand new insight, knowledge, and skill. Typically, one must understand how automation works, its failures, pathways to failures, and how to make it work. Consider ease of interaction. Complex interaction increases workload, anxiety, stress, and detracts from the work itself. Poor representation can conceal incipient problems for a long time. Automation must be convenient and reassuring. People are unlikely to sacrifice ease of use for inconvenience, even if automation can do more.
Identify the functions necessary to achieve the purpose. Allocate functions between humans and automation to enhance their integrated performance. Mutual reliance and interfaces are significant issues here. They determine the effectiveness of autonomous operation and manual control. In addition, consider reliability, safety, regulatory requirements, and cost.
While dividing work it is not necessary to divide and obtain the smallest unit of work. Work that is too elementary can distance people from the overall purpose. So, keep work meaningful. When assigning work, it is easy to focus on selecting the best performer and overlook the final result. Therefore, consider interdependence, coordination, and communication required to jointly achieve goals. Do not assign functions exclusively to exploit technological capabilities of automation. Such an approach is likely to leave a person with disjointed functions. As a result he or she may be error-prone or underutilized.
Serious disturbances change the operation of a complex system from coordinated whole to functionally separate subsystems. Work becomes fragmented and compartmentalized. There are more functions to perform and responses to watch. Communication and coordination decline. Comprehending all operations is difficult. Furthermore, there is a sense of urgency. As a result, it is easy to mismanage these situations into serious and risky incidents or even accidents.
Organize to adapt
Organize the interface to support adaptation. People interact with automation systems through the human–machine interface. The content of the interface profoundly influences work. Among other things, it affects analysis, decisions, and actions in individual and group activities. Interface design must exploit human capabilities and compensate for human limitations. Two important principles are useful to organize the interface. One, conceal the complexity inherent in a system. Two, reveal the system at multiple levels of abstraction. To illustrate, an abstraction must emphasize or suppress details relevant to the work performed at a given level.
Represent functional relations and constraints governing system operation. Use visual forms to aid perception of system state as well as abstract concepts and relationships with little cognitive effort. Present data in the proper context. Articulate meaning and facilitate interpretation by using suitable forms of data. Provide the means to execute corrective actions with little effort. Excessive navigation consumes time, adds to memory load, and increases error possibilities. Get a good balance between cramming displays and curtailing navigation. Combine related data into a coherent picture.
Minimize the possibility of demand–resource mismatch in both routine and unusual work situations. Support the performance of routine work by cue–action relations. Support knowledge-based problem solving for unusual work. Reduce cognitive data integration as well as navigation and manipulation. Don't force people to figure out invariant relationships. Stratify control to perform perception–decision–action operations over multiple time scales. Give sufficient time to solve problems. Expose abnormalities promptly. Highlight urgency.
Information is vital to improve collective awareness and joint performance in the technical and temporal domains. Provide the information necessary to think in terms of high-level strategy and low-level tactics. An information differential in the organization hinders coordination. So, provide easy access to information. Organize information so that people can quickly obtain answers to the "what, why, who, where, when, and how" questions. After all, people respond to their environment according to the needs and constraints they see.
Organize people to adapt. A centralized, functional hierarchy has long been a staple of organizational methods. Sure, such a structure assigns responsibility, establishes accountability, and provides stability. But, this is more likely to occur in routine situations only. Move to unusual situations and the organization's effectiveness may soon decline. Redefining strategy, reassigning functions, and maintaining coordination can suffer. Worse, the organization may be unable to quickly adapt to new demands.
The efficacy of adaptation rests on flexibility and speed. People must spontaneously reorganize themselves and redistribute their work without outside direction. They must make qualitative changes in structure, roles, and behavior in response to the dynamics of a situation. Often, an organization must change from a centralized, functional, hierarchical entity to a distributed, informal, and flat entity. People must adopt different or multiple technical and social roles. And they must form responsive coalitions to temporarily replace formal groups set up previously.
So exercise great care while limiting roles a priori. Limits on roles can stifle adaptation. Similarly, indiscreet independence can stifle adaptation through uncoordinated change.
Behind the byline
Mujeeb Ahmed is a P.E. and a senior member of ISA. This article is from his presentation at the 47th Annual Power Industries (POWID) Conference and the 14th Annual Joint POWID/EPRI Controls and Instrumentation Conference.
Control system behavior is algorithmic
The Smith predictor is the most common MPC formulation for dead-time compensation.
Here is the block diagram of the original Smith predictor with a type-I disturbance—a load that applies to the system prior to entering the process block. (A type-II disturbance describes a load that applies after leaving the process block.)
InTech's resources on MPC theory and optimization are extensive. Here are six recent scientific papers on the topic.
Return to Previous Page