P O W E R
S Y S T E M S
AI Application Areas in Power Systems
Iraj Dabbaghchi, American Electric Power Richard D. Christie, Gary W. Rosenwald, and Chen-Ching Liu, University of Washington
HE ELECTRIC POWER INDUSTRY IS continuously searching for ways to improve the efficiency and reliability with which it supplies energy. Although the fundamental technologies of power generation, transmission, and distribution change quite slowly, the power industry has been quick to explore new technologies that might assist its search and to wholeheartedly adopt those that show benefits. This general tendency has held true to form for the various artificial intelligence technologies. General planners, expert systems, artificial neural networks, inductive learning, fuzzy logic, genetic algorithms—researchers have applied almost every form of AI tool in at least prototype form to one or more problem areas in the power industry, and new practical applications of AI appear with increasing frequency. In some cases, AI tools augment or replace existing techniques. In others, AI tools enable solutions to problems previously addressed only by natural intelligence, creating new applications for computers. The dual questions of which problems to attack with AI techniques and which AI techniques to use for a particular problem are abiding ones in the power industry, as they are in many other industries. Problemcentered approaches (problems seeking solutions), tool-centered approaches (solutions
AI TECHNIQUES PLAY A PROMINENT ROLE IN POWER SYSTEM MANAGEMENT AND CONTROL. THIS ARTICLE DESCRIBES THE
POWER SYSTEM PROBLEMS THAT ARE LIKELY CANDIDATES FOR
AI TECHNIQUES, AND THE TECHNIQUES THAT HAVE BEEN USED ON THOSE PROBLEMS.
seeking problems), and random-matching approaches (justified only by results) have all been employed in the power industry. In this introduction to IEEE Expert’s special track on AI in power systems, we’ll describe the power system problems that are likely candidates for AI techniques, and the techniques that have been used on those problems. These techniques promise to play an even more prominent role in power system management and control.
From generation to distribution
The power industry largely involves utilities—public or privately owned corporations that are regulated monopolies with the exclusive right to sell electric power in a desig0885-9000/97/$10.00 © 1997 IEEE
nated service territory. Deregulation legislation, which is creating competitive markets for the generation of electrical energy, is changing this structure. New organizational forms are appearing around the world. The changes are only beginning, and the eventual destination remains a subject of intense debate (see the interview with power systems expert Sarosh Talukdar on page 78 for a discussion of these changes). The administrative aspects of a utility—that is, billing, accounting, finance, personnel management, and so on—are not much different from those of other corporations. However, the technical aspects of electric utilities places them in a special category. The typical utility’s electric power system consists of generation, transmission, and distribution systems (see Figure 1). Generators produce electric power. The transmission
system moves it in bulk, often over large distances, and the distribution system delivers it to individual customers. A wide variety of generation exists: large nuclear or coal-fired power plants, small combustion turbines burning natural gas, hydroelectric generators, windmills, and so on. The typical utility has a broad mix of types. Generation-related problems fall easily into those related to a single generator or plant and those requiring coordination among more than one plant or generator. A utility might operate hundreds of generators. Generator fuel and capital costs are major economic factors. The transmission system consists of highvoltage (69–800 kilovolts) transmission lines connected in a network (see Figure 2). The connection points are substations. Some substations contain transformers that step voltages up or down. All have circuit breakers and switches that permit alteration of the transmission configuration. The circuit breakers also perform a protective function, opening to isolate short circuits that might occur on any portion of the system. The typical utility owns and controls several hundred transmission lines and substations, a small part of a larger, interconnected transmission system that can be quite large and can cover a substantial land area. (For example, three such large, interconnected systems are in the US and Canada). Most transmission lines are AC, and differ only quantitatively, but some major transmission is High-Voltage DC (HVDC), which is qualitatively different in operation. The distribution system consists of feeders, lower-voltage (1.1–69 kV) power lines that carry power from the substations to individual loads. Poletop or padmount transformers step the voltage down from the distribution level to the household level, and smaller circuit breakers and fuses operate to interrupt short circuits. Distribution systems usually operate radially—that is, only one path exists for the power to flow from the substation to the load, although networks are used in some urban areas where load density and demand for reliability are high. A utility has thousands of feeders, with hundreds of thousands of customers and distribution components. Even this simplified sketch of a utility illustrates several important characteristics of the problems that appear. First, problems can appear as either component or system problems. Second, many system problems exhibit combinatorial complexity. ApplicaJANUARY–FEBRUARY 1997
Figure 1. The basic structure of an electric power system.
tions that address system-level problems must be able to deal with the number of components in the typical power system, either directly or by reducing the problem size to a manageable level without losing too much accuracy in the result. Third, special cases abound—HVDC transmission, for example.
Opportunities for AI in power systems
AI tools have been useful for solving power system problems when there is a good
match between the problem characteristics and those of the AI tool. For instance, one aspect of power system problems that the previous description omits, but which is highly significant for AI applications, is the nonlinear behavior of the various components and of the entire system. The nonlinearities are of three basic types: near linear, continuous nonlinear, and discrete. In normal operation, many problems can be treated as near-linear, and many numerical applications take advantage of this to minimize solution times. As the power system becomes more stressed—by larger loads and power
Figure 2. The American Electric Power network (courtesy of AEP).
transfers, for example—the nonlinearities become too large to ignore, and control limits start to appear. In either the near-linear or nonlinear case, the power system routinely encounters planned and unplanned discrete changes caused by switching operations, either automatic or manual. On one hand, nonlinear component behavior means that even an application that addresses one component at a time will be nontrivial. On the other hand, nonlinear problems are precisely the type that AI tools can address. AI tools can also be used in a wide variety of power system areas to train personnel to solve the associated problems. Experience in solving these problems and the use of good heuristics often contribute to the quality of the solution and the speed at which it is obtained. By participating in simulations that use AI tools to apply archived knowledge, power system personnel can gain experience and can learn from the knowledge of others. To investigate which characteristics of power system problems are suited to AI tools, we can categorize the problems by time frame into real-time control, operations, operations planning, and planning. Real-time control. This category deals with those processes occurring too fast for human intervention or those subject to local, autonomous control. Often both characteristics apply. Local controls distributed in the power system must often be coordinated so that many autonomous controls work together in response to a system condition. Real-time control involves both discrete and continuous control systems. Many real-time controls are quite simple in their individual operation, but their coordinated effect on the power system and their interactions through the power system can become quite complex. Protection, a discrete control, is perhaps the most ubiquitous of the real-time controls. Protection must correctly classify the power system state as normal or faulted. It must also isolate the faulted portion of the power network in tens of milliseconds to prevent equipment damage and to continue delivery of power in the rest of the system. Protection controls are usually coordinated by controlling their relative operating times. Heuristics and logic about the effects of device operations are the bases for protection settings. For generators and the transmission system, protection is implemented with relays, devices that sense voltage or current, recognize fault conditions, and signal circuit breakers to
open. Distribution systems use relays as well as fuses for protection, sometimes in coordination with automatic switches that open or close when they sense specified conditions. Geographically separated pairs of relays sometimes communicate, but with only a few status bits. Load shedding is another form of discrete control. The loss of a power system component, such as a generator or transmission line, can instantaneously change the amount of load the system can supply. When loads exceed this supply capacity, the load must quickly be reduced to avoid widespread blackouts. Such excessive loads are indicated by dropping frequency or voltage. A load-
THE EMS, WITH ITS AUTOMATIC DATA COLLECTION, COMPUTATIONAL FACILITIES, INTERFACE TO
HUMANS IN THE OPERATIONS CONTROL LOOP, AND COMPLEX REAL-TIME TASKS, IS AN ATTRACTIVE ENVIRONMENT FOR AI TOOLS.
shedding relay senses either frequency or voltage, and de-energizes its assigned distribution feeder to reduce load when it classifies the system state as one of generation/load imbalance. Relay operating times are coordinated to drop the right amount of load at the right time, based on analytical and heuristic knowledge about load demand and system behavior. A transformer has two coils, called windings, each with a specific number of turns. The turns ratio is the ratio of the turns of the high-voltage winding to those of the lowvoltage winding, and is the same as the voltage ratio. In AC systems, power is present as a complex number with real and imaginary parts. The real part is the real power, measured in watts; the imaginary part is the reactive power, measured in VARs (VoltAmperes-Reactive). On-load tap changers are local controllers that change transformer turns ratios by small,
discrete increments to maintain the voltage on one side of the transformer at a nearly constant value while the voltage on the other side changes, or sometimes to control realpower flows. The turns ratio can normally be varied in 32 increments between ±10% of the nominal turns ratio (the ratio that produces the transformer’s rated voltage) over a time frame of minutes. Capacitor switches sometimes have local controllers that switch based on local voltage and time of day. Continuous changes in reactance can be supplied by Static VAR Compensators (SVCs) that switch thyristors every cycle. (Reactance is the imaginary part of an AC circuit’s impedance. Thyristors are semiconductor devices that can be rapidly switched on and off.) HVDC terminals also control thyristor switching to vary the real power transferred through the HVDC line. Generator exciters, which control the reactive power and terminal voltage of generating units, are important real-time continuous control systems that play a major role in maintaining the power system’s stability. Portions of the exciter control system that deal explicitly with stability are called power system stabilizers. The power system’s rich diversity of distributed real-time control functions provides many opportunities for AI applications. These opportunities occur both directly in the controllers to classify monitored conditions and execute desired logic or heuristics and indirectly in solving problems complicated by the interaction of real-time controls at the system or component level. Operations. This category includes realtime human decision making in time frames ranging from a few minutes to several hours. Most utilities have a control center with an energy management system, a large processcontrol computer system. Measurements from all over the power system are telemetered to the EMS, and control signals are sent through the EMS to power system components. These functions are called Supervisory Control and Data Acquisition (Scada). Operators, often called dispatchers, view the power system through the EMS, issue manual controls, and establish setpoints and operating modes for centralized real-time control applications. Other applications in the EMS assist the dispatchers by analyzing or optimizing the power system in various ways. Dispatchers also direct mobile field crews and power plant operators and comIEEE EXPERT
AI in Power Systems in the next IEEE Expert:
municate with dispatchers in other utility control centers. The principal operating functions are monitoring the power system’s health and diagnosing the cause or location of abnormal conditions through the telemetered values from the Scada system. These values are saved for historical analysis and are automatically checked against limits. Limit violations generate alarms to notify the dispatchers of abnormal conditions. Normally, few alarms appear, but when something goes wrong, the effects can be widespread, and hundreds or thousands of alarms can be generated in a few seconds. Using the monitored data to diagnose a problem often requires applying logic and heuristics about system relations and the operation of system components. Automatic generation control is a centralized real-time control function that matches generation to load and controls the energy interchange between interconnected utilities. It does this by controlling the real power outputs of operating generators, issuing a new control setting every two to four seconds. Dispatchers apply their experience to changing conditions and might change the modes and setpoints from minute to minute, but the real-time control loop is automatic. Short-term economic optimization seeks to meet the load demand at minimum fuel cost with existing on-line equipment. This is principally accomplished by economic dispatch, the numerical optimization of generator output levels. Additional savings come from sales or purchases of energy (called interchange) negotiated between the dispatchers of interconnected utilities, using economic-evaluation tools. A third source of savings is demand side management, which turns some loads off during high-cost time periods and on during lower-cost times. Experience and heuristic search techniques can aid the search for the economic optimum. Dispatchers must maintain the power system’s security—that is, its ability to continue to deliver electric energy to all or most of its customers despite component failures. The power system is designed for good fault tolerance, but its response to any particular component failure will vary depending on system conditions. One goal of security assessment is to classify the system’s ability to withstand a failure, based on the power system’s current operating state. The accepted strategy for security assessment is to simulate the system’s static and transient
response to selected possible failures. Selecting the relatively short list of failures from all potential failures is a difficult problem that usually relies on the application of experience. Dispatchers operate various controls to maintain or improve security. When done under time pressure to correct problems, this is called emergency control. When done in normal operation, optimization methods can help maintain security at minimal cost. When the supply of power to loads is interrupted, dispatchers take restorative actions to return power to as much of the system as possible. When the effect of the interruption is localized, dispatchers might reconfigure the system to pick up the dropped loads. Determining switching sequences and system configurations that do not violate constraints requires the application of switching logic. This process benefits from heuristics that limit the search space. Occasionally, power systems experience a blackout, a widespread loss of loads. Restoration of a power system after such an outage is a complex task that relies on generic plans prepared beforehand and modified for the specific situation. Dispatchers use the plan to select and coordinate switching actions and the restart of generators. The EMS, with its automatic data collection, computational facilities, interface to humans in the operations control loop, and complex real-time tasks, is an attractive environment for AI tools. They can augment existing applications or provide new applications beyond the scope of traditional numerical control, analysis, and optimization applications. Operations planning. This category deals with operating strategies for time frames ranging from a day to a year into the future. Operations planning occurs in interactive environments using either the EMS or offline processing. The three principal focuses are economics, security, and maintenance scheduling. All require a forecast of the loads on the power system. Short-term forecasting predicts loads for up to two weeks, using historical data. Longer-term load forecasts depend more on economic forecasts. The unit-commitment problem involves scheduling generator start-ups and shutdowns to meet predicted loads at minimum cost. Generator start-up costs and minimum up-and-down times make this a complex problem that can benefit from heuristic search techniques. Hydroelectric generator
“Knowledge-Based Assistance for Contract Compliance,” by David G. Leahy, Jonathan G. Wallace, Maurice D. Mulvenna, and John G. Hughes
scheduling is an extended form of unit commitment made even more complicated by the interrelationships of hydroelectric generators located on the same river system and by the annual rainfall cycle, which extends the scheduling horizon. The economic evaluation of long-term energy-purchase agreements uses similar techniques. Operations planning uses the same security-assessment techniques that operations use, but with more human interaction and analysis and fewer approximations. These techniques examine the security impacts of proposed energy purchases and component maintenance schedules. The time frames range from the next day to six months into the future. Maintenance schedules are optimized in terms of their impact on operating costs, particularly for generators. The complex problems in load forecasting, the heuristics used in large nonlinear optimization problems, and the knowledgeintensive tasks associated with operations planning are all areas of potential AI application. Planning. The primary concern in power system planning is decision making for capital projects with lifetimes measured in tens of years. Each alternative is evaluated for its security and economic impacts on the existing system. Although driven by technical issues such as load growth and generation or transmission capacity, major planning decisions dealing with new generation or new transmission lines tend to be dominated by political and financial considerations. The analysis process is often highly specialized for individual utilities and even for individual decisions. The large amount of capital involved can justify extensive analysis using all available tools. Heuristics often guide the technical analysis. Empirical rules often help identify critical outages that might cause line overloads or voltage violations. Experience is also important in deter61
Expert systems Artificial neural networks Fuzzy sets Heuristic search 0 10
Real-time Operations Operations Planning control planning
20 30 40 Number of applications 50 60
Figure 3. Intelligent system methods at ISAP94.
mining remedial actions and suggesting improved alternatives. Smaller planning problems involve more purely technical design decisions, but are still large, complex optimization problems. These problems range from determining protective relay settings to deciding where to dig the ditches for underground distribution cables. Such problems are often routine and repetitive and are often solved heuristically. Interest is growing in applying more optimal solution methods to a wide range of “small” design problems in the distribution, transmission, and even generation systems. Planning also occurs to develop procedures for shorter time frames. For example, power system blackouts occur occasionally, and generic plans are developed off line for restoring the power system. These restoration plans require coordinating the generation, transmission, and distribution systems in abnormal operation, to identify available generation capacity, transmission paths, and loads to pick up, all subject to numerous constraints. The preparation of a restoration plan involves heuristic search, rules, and prioritizing. Generic tasks exist among restoration plans from different utilities. Software tools based on these tasks can help utilities develop their specific restoration plans.
cation of AI to power systems, it does represent the focus of current research. It also provides insight on the types of applications and the extent of their development, from ideas to systems in practical use. Figure 3 shows the number and time frame of applications presented at ISAP94 in the four categories. Applications that use techniques from more than one category (for example, fuzzy sets and ANNs) are counted once in each category. The statistics in Figure 3 reflect the length of time over which each category has been applied to power system problems, as well as its suitability for power system applications and, to some extent, the level of development of each category. Expert system applications are notably more advanced than the other categories, and ANN applications are more numerous than fuzzy sets or heuristic search. Expert systems. These systems are computer programs that possess expertise in a given area. This expert knowledge is normally stored separately from the procedural part of the program and might be stored in one of many forms, including rules, decision trees, models, and frames. Many areas of applications in electric power systems match the expert systems abilities—for example, decision making, archiving knowledge, and solving problems by reasoning, heuristics, or judgment. Expert systems are particularly useful for these problems when a large amount of data must be processed in a short time period. Expert systems appeal to many power system operators and engineers. These systems can consistently apply knowledge gained from years of human experience to simplify complex power system problems and make high-level operating, planning, and design decisions. Automated symbolic reasoning permits automated solutions for problems that have required many hours of routine intellectual work, and that have not yielded to numerical approaches.
Applications of AI to power systems
Various AI techniques have found numerous applications to power systems. These techniques conveniently fall into four categories: expert systems, artificial neural networks, fuzzy sets, and heuristic search. The applications presented at the 1994 International Conference on Intelligent System Application to Power Systems, held in Montpellier, France, illustrate the degree to which each category has been applied to power system problems.1 Although this conference does not reflect all developments in the appli62
However, expert systems have their weaknesses. The variability and length of expert system runtimes are sometimes concerns for real-time applications. Perhaps the biggest immediate concern about expert systems in power systems is the ability to keep the knowledge base current. Often, rules depend in some way on the power system’s configuration (for example, “If load is peak, then power flow is north to south.”) or on the utility’s operating policies (“If bus voltage is below 0.85 per unit, then shed load.”) Both are mutable. Configuration changes are slow but continuous, and policy changes can be sweeping and sudden. Utilities are not used to providing the level of software maintenance expert systems appear to require. To gain acceptance in their end use, all applications must justify their cost. Because expert systems are not good at obtaining optimal solutions, they generally justify their cost in two ways. First, they can reduce the chances of costly errors. Although costly errors (for example, blackouts) typically have a low probability of occurring, the benefits of applications that reduce these errors are considered to be high, although difficult to determine accurately. Second, they can improve the performance of other applications that do save money directly. For example, large, complex mixedinteger and nonlinear optimization problems, such as those in power engineering, settle for near-optimal solutions. These programs are often tuned using problem characteristics, to reduce runtimes. The solution’s optimality can depend on the talent of the person tuning the program. In theory, expert systems should be able to apply the experts’ knowledge to obtain results at least as near optimality as are the experts’ results, without continuous hands-on tuning. Expert systems find application in realtime control, operations, operations planning, and planning. The expert systems presented at ISAP94 use methods including rule-, model-, and case-based systems; decision trees; objectoriented programming; and distributed and parallel processing. Figure 4 shows the expert system application areas. The control applications include generator voltage regulators, power plant operations, and control of SVCs. Intelligent systems for real-time control use fuzzy logic to increase their speed of operation because of the nature of the knowledge required. Fuzzy logic also enables a robust output signal. Operations
Monitoring and diagnosis Control
control systems apply heuristics appropriate for the system’s current state to suggest control actions for the system operators. The monitoring and diagnosis applications are for operations and operations planning. Monitoring tasks take system data about components such as a generating plant, substation, transformer, or the system network. Diagnosis tasks then apply logic and heuristics to assess the system’s health and diagnose the cause of any abnormal conditions (such as a fault). For example, some of these systems monitor equipment for maintenance scheduling, diagnose faults, or process alarms from the EMS (see the related sidebar). The restoration applications use available data, logic, and heuristics to determine actions to restore power to as much of the system as possible while complying with constraints. The security applications classify the system’s security or select critical contingencies for further numerical analysis, based on extracted heuristics or learned knowledge (such as that represented in decision trees). The load-forecasting applications integrate fuzzy expert systems with ANNs. One rule-based expert system handles transmission planning. Other expert systems help with the design of transmission and distribution systems and substations, and with setting relays. Expert systems also find applications in training for operations and restoration. Artificial neural networks. ANNs are biologically inspired systems that transform a set of inputs into a set of outputs through a network of neurons, each of which generates one output as a function of its inputs. The inputs and outputs are usually normalized, and the output is a nonlinear function of the inputs that is controlled by weights on the inputs. The network learns these weights during training, which can be supervised or unsupervised. A wide variety of network connections and training techniques exist. Power system problems regarding classification or the encoding of an unspecified nonlinear function are well-suited for ANNs. ANNs can be especially useful for problems that need quick results, such as those in realtime operation, because of the ability of ANNs to quickly generate results after receiving a set of inputs. The ability to scale ANN applications to realistic dimensions for power system problems is a major issue. Component-related
Restoration Planning Security Load forecasting Other 0 3 6
Idea Prototype Field test Practical use
9 12 15 Number of applications 18 21
Figure 4. ISAP94 expert system applications.
Control Load forecasting Security Monitoring and diagnosis Other 0 3 6 9
Idea Prototype Field test Practical use
12 15 18 21 Number of applications
Figure 5. ISAP94 artificial neural network applications.
applications generally have a limited number of inputs, but realistic power systems have potentially tens of thousands of inputs at the system level. Training times are usually nonlinear with problem size, and ANN applications that work quite well for six-bus systems might be computationally infeasible when facing more realistic 1,000-bus systems. Feature-extraction mechanisms that reduce the number of inputs to a manageable value are almost mandatory for system problems. A related concern is the size of the training set. Because of the complex, nonlinear behavior of power systems, ANNs can require large training sets to obtain sufficient accuracy. The trade-off between training set size and solution accuracy is difficult to control analytically. It is quite possible to add cases to the training set that do not appreciably improve the accuracy of the ANN’s solutions. A final concern is the validity of an ANN’s training over time. Suppose the ANN attempts to predict the result of some numerical simulation of power system behavior, using a training set that assumes a certain network topology. The topology changes temporarily from hour to hour, and permanently from month to month. The accuracy
of ANN results might not be sensitive to such changes. Such sensitivities are difficult to predict analytically. ANNs are a clear first choice for most classification problems, and for fast approximations to the results of complex numerical analyses. So, they are suitable for many applications in real-time control, operations, and operations planning. The ANN research presented at ISAP94 uses many techniques, including adaptive weights, unsupervised learning, and clustering. Some applications combine ANNs with fuzzy sets. Figure 5 shows the ISAP94 ANN application areas. The control applications include real-time generator exciter control and intelligent relays. The monitoring and diagnosis applications include alarm processing and monitoring and fault diagnosis for networks, substations, and equipment. One ANN application that helps detect shorted turns in transformers and generators has progressed beyond prototyping to field testing. The security applications classify the system as secure or insecure, and in some cases suggest actions to increase system security based on features of the current operating state. They
assess security for operations as well as the operations planning problem of generation scheduling. ANNs are also used for shortterm and long-term load forecasting for operations planning and planning, respectively (see the related sidebar). Fuzzy set theory. By embracing partial set membership, fuzzy set theory allows a certain level of ambiguity throughout an analysis. Because this uncertainty can characterize available information and reduce problem complexity, fuzzy sets are useful in a wide range of applications. The mathematical framework provided by fuzzy sets has been successfully integrated with techniques that traditionally use classical mathematical methods. For example, fuzzy logic has been developed as an extension of multivalued logic that moves from classical logic with crisp propositions toward approximate reasoning with
imprecise propositions, using fuzzy set theory. For power systems, fuzzy sets are appropriate in many areas where the available information involves uncertainty. For example, a problem might involve logical reasoning, but applied to numerical, rather than symbolic, inputs and outputs. Fuzzy sets provide the translations from numerical to symbolic inputs, and back again for the outputs. A typical application is intelligent control where the control knowledge is specified as fuzzy rules. Combining fuzzy set theory with other analysis techniques can create difficulties. Also, the use of fuzzy sets requires the identification of proper membership functions. For power system applications, identifying these functions and keeping them up to date as the system changes can require significant effort. Fuzzy set theory is used for real-time control, operations, operations planning, and planning.
Although many opportunities exist for the application of fuzzy set theory to power systems, most fuzzy set developments reported at ISAP94 are in conjunction with expert systems and ANNs. Figure 6 shows the range of ISAP94 fuzzy set applications. The majority of the real-time fuzzy control applications implement fuzzy logic. The restoration and load-forecasting applications also use fuzzy logic. Fuzzy sets find application in the monitoring and diagnosis of faults and equipment. They also determine a fuzzy assessment of system security and represent uncertainty in generation expansion planning. Heuristic search. Genetic algorithms and simulated annealing are two forms of heuristic search, which solves optimization problems by randomly generating new solutions and retaining the better ones. The solutiongeneration process is the key to these
Three applications of AI to power systems
We’ll now show how expert systems, artificial neural networks, and genetic algorithms can support three specific areas of power systems management. Expert systems for alarm processing. System operators keep informed about the power system’s current state through an Energy Management System (EMS), which conveys information to the operators as alarms. Traditionally, each alarm has represented one event in the power system—for example, the opening of a circuit breaker. Alarms normally occur at a relatively slow rate. Casualties can cause bursts of hundreds of alarms. Scanning through these alarms to extract the most important information and construct a picture of the system’s new state can be difficult. One approach to alarm processing ranks the alarms according to their priority to help the operators identify the most important pieces of information. Other approaches try to piece together EMS alarms to synthesize higher-level information to present to the operators, while suppressing the constituent EMS alarms. Whatever the specific task of the alarm processor, the operators’ expert knowledge is useful in analyzing the alarms.1–3 From their understanding of and experience with power system operation, operators can determine what information is important and how it can help construct higher-level information. Expert systems for alarm processing store this extracted knowledge to perform the task in a way similar to that of the expert. With the expert system prioritizing, filtering, and assimilating low-level information, operators can gain high-level information more quickly and use their time for other tasks. These benefits are most important during crises, when there are many alarms and processing them is most difficult. Alarm processors must meet very high acceptance standards. The operators rely on the information they receive from alarms for making critical decisions, often under time pressure. Adding an extra layer of information processing must not corrupt or distort the transfer of information to the operators. So, the alarm processor must be accurate and consistent. With advances in power system devices and communication, even more information will be gathered and presented to operators. The increase in information will enable a more complete view of the power system’s state, but it will also increase the need for alarm processing to effectively handle all the system information. Artificial neural networks for load forecasting. To ensure that load demand can be met, operations planning and planning must ensure that enough generation is available and that the transmission and distribution capacity is sufficient to deliver it. In operations planning where generation capacities are fixed but generators might be on or off line, keeping some units on line represents a fuel cost even when they are generating no power. For example, the boiler must be kept hot on large thermal plants. For planning, installed capacity represents capital costs. A key factor in these analyses is the load that must be served in the future. The quality of the load prediction can greatly influence the quality of the developed plans. Load forecasting is difficult, partly because the load is based on distributed, individual decisions by people responding to their desire for electricity. Load forecasting normally uses historical load behavior to form a prediction.4–12 Historical data shows a short-term correlation between the total load demand and such climatic information as temperature, cloudiness, and wind, and such sociological factors as the day of the week. Long-term load forecasts are usually predicted using economic forecasts. The relationships between the load and these factors are difficult to determine. Because ANNs can encode complex, nonlinear relations, researchers have used them to capture the relationships between the load and selected factors. These relationships are represented in historical data, which is used to train the network. Often, these networks use predicted data as input (such as temperature or economic indicators), which might limit the accuracy of the load predictions. Research shows that ANNs can produce load forecasts with an accuracy comparable to other (for example, statistical) methods. Genetic algorithms for generator maintenance scheduling. Generators are typically large, expensive power system components that require
searches. The new solutions should tend toward improvement, yet explore the problem space to minimize the chance of getting stuck in a local minimum. Both techniques can be applied to optimization problems with arbitrary objective functions and constraints. GAs encode solutions as bit strings and generate an initial population. Then, they use techniques observed in natural genetic reproduction (crossovers and mutations) to create a new generation of solutions. Better solutions have a higher probability of reproduction, but poor solutions have a nonzero chance, which permits escape from local optima. SA generates new solutions in a neighborhood around an existing solution. It gradually decreases the neighborhood’s size until the solution settles to a nearly optimal point. Heuristic searches raise several concerns. Defining stopping criteria is an art. Stopping too soon can result in a solution far from the
Monitoring and diagnosis Load forecasting Planning Restoration Security 0 3 6 9
Idea Prototype Field test Practical use
12 15 18 21 Number of applications
Figure 6. ISAP94 fuzzy set applications.
optimum. Not stopping in time can result in excess computation without significantly improving the solution. A solution at or near the optimum is also not guaranteed; getting caught in local optima is quite possible. A number of parameters, such as the cooling rate in SA or the ratio of each type of reproductive
technique in GAs, must be tuned to obtain good algorithm performance. The solution’s quality depends on the achieved coverage of the solution space. Many power system problems are large enough that achieving adequate coverage is difficult and computationally costly. Finally, the evaluation of a solution
regular maintenance to sustain their efficiency and expected life span. However, an unused generator is also costly, because the capital invested in that machine is sitting idle. The generator cannot be used to supply loads, so more expensive means of generation might be necessary. Determining the optimal scheduling of generator maintenance and incorporating this plan with other operational concerns such as unit commitment and security are difficult. Generator maintenance scheduling presents a large optimization problem that suffers from nonlinearity and combinatorial complexity.13 Finding a solution to this problem in a reasonable length of time involves techniques that achieve near-optimal solutions. The problem is a good candidate for heuristic search techniques such as genetic algorithms. Certain aspects of the problem have a natural GA formulation—for example, the binary state of the generators (on or off line) fits the required GA form. The objective—minimum cost—is incorporated into the fitness function for comparing the quality of the members of the population. To ensure a good solution, the parameters for the execution of the GA, such as population size, survival rates, and mutation rates, should be tailored to the problem at hand. With these parameters properly set, the evolution process can provide a near-optimal solution.
4. V. Bohm et al., “Neural Network Predicting System for Electric Load Data in West Bohemia,” Proc. ISAP94, 1994, pp. 857–863. 5. S. Canu, M. Duran, and X. Ding, “District Heating Forecast Using Artificial Neural Networks,” Proc. ISAP94, 1994, pp. 767–773. 6. P.K. Dash and A.C. Liew, “A Comparative Study of Load Forecasting Models Using Fuzzy Neural Networks,” Proc. ISAP94, 1994, pp. 865–872. 7. A. Garcia-Tejedor et al., “A Neural System for Short-Term Load Forecasting Based on Day-Type Classification,” Proc. ISAP94, 1994, pp. 353–360. 8. M. Kanda et al., “Long Term Maximum Load Forecasting Using Modified Neural Network Approach,” Proc. ISAP94, 1994, pp. 783–790. 9. S.J. Kiartzis, A.G. Bakirtzis, and V. Petridis, “Neural Networks Application to Short Term Load Forecasting,” Proc. ISAP94, 1994, pp. 339–344. 10. K. Makino et al., “Short-Term Load Forecasting Using an Artificial Neural Network of Locally Active Units,” Proc. ISAP94, 1994, pp. 849–856. 11. H. Mori and H. Kobayashi, “A Fuzzy Neural Net for Short-Term Load Forecasting,” Proc. ISAP94, 1994, pp. 775–782. 12. R. Satoh, E. Tanaka, and J. Hasegawa, “Daily Load Forecasting Using a Neural Network Combined with Regression Analysis,” Proc. ISAP94, 1994, pp. 345–352. 13. T. Sutoh, H. Suzuki, and N. Nagai, “Large-Scale Generator Maintenance Scheduling Using Simulated Evolution,” Proc. ISAP94, 1994, pp. 567–573.
1. A.O. Ekwue et al., “System Monitoring Expert System,” Proc. Int’l Conf. Intelligent System Application to Power Systems (ISAP94), 1994, pp. 541–547. R. Khosla and T. Dillon, “Dynamic Analysis of a Real Time Object Oriented Symbolic-Connectionist System Using State Controlled Petri Nets,” Proc. ISAP94, 1994, pp. 495–502. Z.A. Vale and M.F. Fernandes, “Knowledge-Based Applications in Control Centers: Key Issues in Development and Integration,” Proc. ISAP94, 1994, pp. 549–556.
Operations planning Planning Operations Other 0 3 6 9
Idea Prototype Field test Practical use
12 15 18 21 Number of applications
Figure 7. ISAP94 genetic algorithm applications.
often requires time-consuming numerical analysis, and many solutions must be evaluated during the search. Heuristic search algorithms have potential applications in operations planning and planning, where better, or more realistic, solutions are worth the extra computational time required. Although no papers at ISAP94 dealt with simulated annealing, several described prototype systems using GAs (see Figure 7). The main application of GAs in power systems is for large, complex, nonlinear optimization problems. At ISAP94, the application of GAs ranged from operations to design, including economic dispatch with environmental constraints, unit commitment, maintenance scheduling (see the related sidebar), generation expansion, and optimal fuelcell placement on distribution lines.
Second, the complexity of operations will increase dramatically, as energy transactions under the purview of a control center increase by an order of magnitude, each requiring technical analysis, approval, oversight, and accounting. Finally, the time pressures associated with competitive markets seem likely to appear in power systems operations and operations planning. This is certainly true for generators, who will have to market their power, and probably also true for the transmission system, which will be pressured to make rapid decisions on the suitability of various sales. The need for fast answers to complicated problems with uncertain and incomplete data will undoubtedly grow as deregulation progresses. This situation presents an enormous opportunity for AI applications to solve the challenging problems looming in the future of power systems.
Iraj Dabbaghchi’s biography appears on page 57.
Richard D. Christie is an associate professor of electrical engineering at the University of Washington, working on problems in power system operations and distribution-system reliability assessment. He has a BE and an ME in electric power engineering from Rensselaer Polytechnic Institute. His PhD is from Carnegie Mellon University. Before pursuing his PhD, he worked for Leeds and Northup, a manufacturer of Energy Management Systems. He is a member of the IEEE Power Engineering and Computer Societies. Contact him at the Dept. of Electrical Engineering, Box 352500, Univ. of Washington, Seattle, WA 98195-2500; [email protected]
HE POWER INDUSTRY HAS SEEN the extensive application of AI techniques to a wide range of power system problems. Although no universal applications or spectacular successes have emerged, AI techniques are quietly finding their way into operational use. However, much room for improvement remains in power system operation and in the AI applications. Deregulation will probably be the primary force driving the long-term future of AI applications to power systems. Three principal features of deregulation seem likely to spur increased efforts to apply AI to power systems. First, uncertainty in the input data will increase. Market forecasts will be required, and there is no reason to believe that electric energy markets will exhibit any more predictability than do other markets.
Gary W. Rosenwald is a software engineer at ABB Power T&D Co. His research interests include the application of artificial intelligence, particularly for power system operation and planning. He received his BSEE and PhD from the University of Washington in 1992 and 1996. He received the University of Washington’s top undergraduate honor, the President’s Medal, in 1992 and a National Science Foundation Graduate Fellowship in 1993. He is a member of Tau Beta Pi and of the IEEE Power Engineering Society. His address is the ABB Energy Planning Center, 110 Corning Rd., Ste. 101, Cary, NC 27511; gary. [email protected]
This material is based on work supported under a National Science Foundation Graduate Research Fellowship. Chen-Ching Liu is a professor of electrical engineering at the University of Washington. His technical interests are AI applications to power systems and knowledge engineering. He chairs a subcommittee on intelligent system applications in the IEEE Power Engineering Society. He was guest editor for a special issue of the Proceedings of the IEEE on knowledge-based systems in electric power systems. He also cofounded the Intelligent System Applications to Power Systems conferences. He received his PhD in electrical engineering and computer science from the University of California, Berkeley. He is a Fellow of the IEEE. Contact him at the Dept. of Electrical Engineering, Univ. of Washington, Box 352500, Seattle, WA 98195; [email protected]
1. A. Hertz, A.T. Holen, and J.-C. Rault, eds., Proc. Int’l Conf. Intelligent System Application to Power Systems (ISAP94), Montpellier, France, 1994.