ai automates substation control

6
January 2002 41 ISSN 0895-0156/02/$17.00©2002 IEEE E lectric substations are facilities in charge of the voltage transformation to provide safe and effective energy to the consumers. This energy supply has to be carried out with sufficient quality and should guaran- tee the equipment security. The associated cost to ensure quality and security during the supply in substa- tions is high. Automatic mechanisms are generally used in greater or lesser scale, although they mostly operate according to an individual control and protection logic related with the equipment itself and not with the topol- ogy of the whole substation in a given moment. The automation of electric substations is an area under constant development. Nevertheless, the control of a substation is a very complex task due to the great number of related problems and, therefore, the decision variables that can influence the substation performance. Under such circumstances, the use of learning control systems can be very useful. Many papers on applications of artificial intelligence (AI) techniques to power systems have been published in the last years. The difficulties associated with the application of this technique include: Selection of the magnitudes to be controlled Definition and implementation of the soft tech- niques Elaboration of a programming tool to execute the control operations Selection, acquisition and installation of the mea- surement and control equipment Interface with this equipment and Application of the controlling technique in existent substations. Our research has focused on the first three points, and the interest of the present work is to expose the obtained results and to present them for discussion. The objective is to show that it is possible to control the sta- tus of circuit breakers (CB) in a substation making use of a knowledge base that relates some of the operation magnitudes, mixing status variables with time variables and fuzzy sets. Even when all the magnitudes to be controlled cannot be included in the analysis (mostly due to the great num- ber of measurements and status variables of the substa- tion and, therefore, to the rules that would be required by the controller), it is possible to control the desired status while supervising some important magnitudes as the voltage, power factor, and harmonic distortion, as well as the present status. M. Ayala A. is with the Electric Test and Research Center Havana, Cuba. G. Botura Jr. and O.A. Maldonado A. are with the University of the Whole State of São Paulo (UNESP), Guaratinguetá Campus, São Paulo, Brazil. ©EYEWIRE Melvin Ayala S., Galdenoro Botura Jr., Oscar A. Maldonado A. Controlling a substation by a fuzzy controller speeds up the response time and diminishes the possibility of risks normally related to human operations

Upload: oa

Post on 22-Sep-2016

220 views

Category:

Documents


4 download

TRANSCRIPT

Page 1: AI automates substation control

January 2002 41ISSN 0895-0156/02/$17.00©2002 IEEE

Electric substations are facilities in charge of thevoltage transformation to provide safe and effectiveenergy to the consumers. This energy supply has to

be carried out with sufficient quality and should guaran-tee the equipment security. The associated cost toensure quality and security during the supply in substa-tions is high. Automatic mechanisms are generally usedin greater or lesser scale, although they mostly operateaccording to an individual control and protection logicrelated with the equipment itself and not with the topol-ogy of the whole substation in a given moment.

The automation of electric substations is an areaunder constant development. Nevertheless, the controlof a substation is a very complex task due to the greatnumber of related problems and, therefore, the decisionvariables that can influence the substation performance.Under such circumstances, the use of learning controlsystems can be very useful.

Many papers on applications of artificial intelligence(AI) techniques to power systems have been publishedin the last years. The difficulties associated with theapplication of this technique include:

� Selection of the magnitudes to be controlled� Definition and implementation of the soft tech-

niques� Elaboration of a programming tool to execute the

control operations� Selection, acquisition and installation of the mea-

surement and control equipment� Interface with this equipment and� Application of the controlling technique in existent

substations.Our research has focused on the first three points,

and the interest of the present work is to expose theobtained results and to present them for discussion. Theobjective is to show that it is possible to control the sta-tus of circuit breakers (CB) in a substation making use ofa knowledge base that relates some of the operationmagnitudes, mixing status variables with time variablesand fuzzy sets.

Even when all the magnitudes to be controlled cannotbe included in the analysis (mostly due to the great num-ber of measurements and status variables of the substa-tion and, therefore, to the rules that would be requiredby the controller), it is possible to control the desiredstatus while supervising some important magnitudes asthe voltage, power factor, and harmonic distortion, aswell as the present status.

M. Ayala A. is with the Electric Test and Research Center Havana,Cuba. G. Botura Jr. and O.A. Maldonado A. are with the Universityof the Whole State of São Paulo (UNESP), Guaratinguetá Campus,São Paulo, Brazil.

©E

YE

WIR

E

Melvin Ayala S., Galdenoro Botura Jr., Oscar A. Maldonado A.

Controlling a substation by afuzzy controller speeds up theresponse time and diminishes the possibility of risks normallyrelated to human operations

Page 2: AI automates substation control

42 IEEE Computer Applications in Power

Plant DescriptionThe system under study represents a test substationwith two 30 kVA three-phase transformers, two CBs, twoswitches, three current transformers, and two potentialtransformers. It also contains an auto transformer (toregulate the input voltage) as well as an impedance tosimulate the existence of a transmission line. The inputvoltage and the output voltage are the same (220 V).This characteristic was selected in order to analyze theoperation of the controller in a laboratory scale in a sec-ond stage of the development of the present work.

Therefore, the first transformer increases the voltage toa value of 13.2 kV, while the second lowers it again to 220V. A fixed filter, an automatic filter for the control of thepower factor and the regulation of the voltage, and threefeeding lines with diverse types of loads of differentnature (including nonlinear loads) are connectedthrough CBs to the output bar. Figure 1 shows the pro-posed outline.

Since the control elements in substations are the CBs

and switches, the goal is to allow the control of the fiveselected CBs (in the output bar) according to some con-figurations and measurements of the observation vari-ables. Initially, the computer-aided system will try tocontrol the plant and will send alarm signals when it can-not find a solution, waiting for the human intervention.So it will learn how the human operator reacts to theinputs and will generate the corresponding behaviorrules. In this way, the system will replace gradually thehuman operator.

Controller Design

Definition of the Input and Output VariablesThere is a great number of variables that can be chosento control a substation. Nevertheless, a limited numberof variables was selected for this study.

The following input variables have been defined:� Vout: Voltage at output bus, phase A (V)� PF: Power factor at output bus, phase A� THDv(%): Total voltage harmonic distortion at out-

put bus (%)� Tv,t (s): Amount of time the voltage Vout is in range

[114.3 V; 119.5 V] (tolerance zone)� Tv,nt (s): Amount of time the voltage Vout is below

114.3 V (alert zone).In selecting the variables, several aspects were kept in

mind. For example, the voltage influences in the connec-tion and disconnection of loads when its value leavessome ranges during certain time. These ranges are repre-sented in Figure 2.

A decrease of the voltage 7.5% below the nominalvalue is allowed during a certain time (for example, 10min), while from 10% on, the maximal allowed time ismuch smaller (for example, 20 s). In case these limits are

Figure 1. Schema of the experimental substation under study

Vn = 220 V - 20 kA

AT

0 - 220 V

FixedFilter

Vn = 220 V

ControlledFilter

Load 1(var)

Load 2(var)

Load 3(var)

30 kVA

13.2 kV

30 kVA

PT

CT

ZTL

Figure 2. Allowed times for voltage ranges

Page 3: AI automates substation control

exceeded, loads will be disconnected in an establishedpreference order. The reconnection of loads will occurwhen the voltage arrives at values above the toleranceinterval, i.e., the normal interval.

On the other hand, the power factor and the total volt-age harmonic distortion influence the connection anddisconnection of capacitors and filters. These variablescan be read by means of sensors and/or transducers,including signal conditional accessories and directed tothe data acquisition card (DAC) by means of analoglines. In this example, the currents and voltages in eachphase are not kept in mind due to the great number ofrules that would be required by the controller.

The CBs were defined as status variables. The switcheswere not included in this study, because their status isonly important for switching and not for control purposes.

The following status variables were defined:� Df: Status of the CB connecting the fixed filter� Dc: Status of the CB connecting the controlled filter� Dl1, Dl2, Dl3: Status of the CBs connecting the load

feeders 1, 2, and 3.Each combination of status variables defines a topolo-

gy and is an input for the controller. Possible values foreach status variable are 0 (open) and 1 (closed).

Thus, what is intended to control is the moment whenthe filters and the loads should be connected or discon-nected by means of signals that are sent to the CBs. Thecontrolled filter, once connected, will maintain its func-tionality as an automatic filter in dependence of the pre-sent harmonic distortion over the time. In case of adisconnection of some filter due to overcurrents, thecontroller can activate connection rules after some timethat can be freely defined before the controller starts.

The actions to carry out as a response to distur-bances in the measurements (values out of normalranges) are dependent not only on the characteristics ofthe disturbances, but also on the present topology of thesubstation, i.e., on the values of the present status vari-ables. For example, the connection of the controlled fil-ter can only occur when the fixed filter is connected.Similarly, the disconnection of the fixed filter can onlyoccur after the controlled filter has been disconnected.

The status variables are input and also output vari-ables. Each CB will maintain its standard protection func-tion against overcurrents. Since signals to these devicesare sent by means of additional relays to activate theCBs, they can be reassembled to have the ability to becontrolled in parallel.

The status variables are read by sensors and/or trans-ducers and connected to digital lines to a DAC. The out-

puts are also sent by digital lines to the CB relays.For the definition of sets, triangular and trapezoidal

shape functions were used.The status variables were not fuzzified because their

measurements are not provided with uncertainty. Theycan only accept two values: 0 and 1.

The time the voltage is in tolerable and not tolerableranges is supervised through the event counters.

Rule SystemThe syntax of each rule can be expressed for example asfollows:

IF(V is Tolerable) and (PF is Low) and (THD is Tolerable)and (Tv,t is Acceptable) and (Tv,nt is Zero)and (Present topology is 00110)THEN(Desired topology is 10110)

The topology is expressed as a five-digit binary num-ber that refers to the five CBs in the following order:� First digit: CB of the fixed filter� Second digit: CB of the controller filter� Third digit: CB of the first load (priority one)� Fourth digit: CB of the second load (priority two)� Fifth digit: CB of the third load (priority three).

In simple words, the rule expressed above means: Ifthe present topology is 00110 (i.e., both filters are dis-connected and only load 1 and 2 are connected) and thefollowing situation is found:� Voltage is tolerable� Power factor is low� Total harmonic distortion is tolerable� Voltage has been in a tolerable zone for an accept-

able time� Voltage has not yet entered a not acceptable zonethen the desired topology is 10110, which means that wemust switch in the fixed filter.

To establish the connection and disconnection rulesof the loads, it was attributed to load 1 the biggest pref-erence and to load 3 the smallest.

The definition of the analog variables was carried outusing the following terms:� For Vout, the fuzzy sets: NT (not tolerable), T (toler-

able), N (normal)� For PF, the fuzzy stes: L (low), T (tolerable), H

(height)� For THDv, the fuzzy sets: L (low), T (tolerable), H

(height)� For Tv,t, the crisp sets: Z (zero), A (acceptable), NA

January 2002 43

The status of circuit breakers can be controlled by using a knowledgebase that relates some of the operation magnitudes, mixing status

variables with time variables and fuzzy sets

Page 4: AI automates substation control

(not acceptable)� For Tv,nt: Z (zero), A (acceptable), NA (not accept-

able).Figure 3 shows the defined sets for each input variable.In the present work, a preliminary calculation of the

maximum number of rules yields that 7,776 rules are nec-essary in order to start the controller. However, thisnumber could be reduced notably by keeping in mindthat, among the 32 available states, not all can be consid-ered as possible. This way, only 9 topologies have beenfound as possible, decreasing to 2,187 the maximum num-ber of necessary rules so that the controller can totallyreplace the human operator. Nevertheless, this numberwas reduced with the inclusion of some initial rules.

Initial Rule Base. In case there exists a knowledgebase on the plant to be controlled, some rules can beincluded as a starting point. The initial knowledge basecan be defined in such a way that diminishes the neces-sary number of rules for the controller to work properly.

This happens in most cases where some situations arenot possible. In the presented study, 144 initial ruleswere included (Figure 4), all in form of extended rules(including one term U, which indicates that the attributecan take any value).

Automatic Rule Extraction. The rule base representsthe knowledge base of the controller. The proposedapproach is able to start the operation with an empty oruncompleted rule base.

During the inferenceprocess, the member-ship degree corre-sponding to eachcolumn in the rule iscalculated. The type ofthe output of this func-tion in the universe ofdiscourse depends onthe variable type and

on the selected term.The rule extraction takes place each time the con-

troller does not find any rule in the rule base with a firedegree bigger than zero, and, as a result, an alarm requir-ing an operator action is sent.

Inference ModuleThe controller outputs are decided by searching in therule base. In this step, called inference, the fire degree ofeach rule is calculated.

Since the consequence part in each rule only dealswith status variables whose values are crisp numbers (0and 1), use of defuzzification methods is not necessary.Therefore, the controller output in each case will be theconsequent part of the rule with the biggest fire degree.

Operation Modus of the ControllerThe controller operation can be carried out in two ways.Following the typical way the controllers run, it can beplaced in operation with a completed rule base, totallyreplacing the human operator. Nevertheless, it can bestarted with an empty or incomplete rule base, whichmeans that it will activate a learning mechanism. Thisway, the controller will complete step by step the opera-tion rules and at the same time will replace the humanoperator. A representation of the controller function canbe viewed in Figure 5.

For the disconnection of loads when the voltagediminishes, the following steps can be processed as anautomatic response of some devices:� Automatic tap change in transformers� Voltage regulation in the auto transformer.

If all these possibilities have been tried and the volt-age continues being low, then operator outputs areprocessed.

To avoid interference with this response, the execu-tion of controller outputs (i.e., status changes of the CBs)can be postponed each time for one sampling interval,provided the inference in the next sampling time yieldsthe necessity of status changes.

In case the system cannot find a satisfactory solutionbecause no rule could be fired, the system sends an alarm tothe operator. He will have some time to decide what actionto take using the switch components on the PC screen.

If the operator action does not arrive during this timelimit (either due to delays or to operator absence), then

44 IEEE Computer Applications in Power

Figure 3. Sets definition for the decision variables (Fuzzi sets for Vout, PF and THDv; crisp setsfor Tv,t and Tv,nt)

Figure 4. Some of the first 144 rules as they are presented tothe software

Table 1. Setup of the feed-forward ANN

Layer Number of Neurons Type of Activation Function

1 Number of input sets = 20 Lineal (input = output)

2 2 * (Number of input sets) + 1 = 41 Sigmoidal

3 Number of state variables = 5 Sigmoidal

Page 5: AI automates substation control

the controller will execute a protection rule previouslydefined, which could be, for example, the total discon-nection of the substation. After saturation of the knowl-edge base or after a certain operating time, the systemwill generate and train an artificial neural network(ANN), in order to replace the rule base.

Substitution of the Inference Engine by an Artificial Neural NetworkAccording to the type and the number of initial rulesas well as the number of existent linguistic variables,the system will calculate the maximum number of pos-sible rules.

When arriving at this number or time limit, the systemwill start a module, where a feed-forward ANN with a hid-den layer will be generated and trained with a codifiedrule base.

This way, the system will try to diminish the time thecontroller needs for the inference, avoiding searching inan extensive knowledge base. Once the network hasbeen trained, the inference will be carried out by meansof this network and not through the rule base.

Each training pattern is coded with binary digits (0, 1).In the case of the analog variables, these digits are thecodification of the sets in each rule.

The network of the present example is a feed-forwardnetwork built by 20 neurons in the input layer, 41 neu-rons in the hidden layer, and 5 neurons in the exit layer(Table 1).

The training process is carried out via a back-propa-gation algorithm. Stop criteriafor the training is the total net-work error as well as a timelimit. In fact, the inference bymeans of the rule base will fin-ish when the network is com-pletely trained.

During the inferencethrough the neural network,the analog measurements willhave to be processed initiallythrough a membership func-tions module before being pre-sented to the network. Thisstep is necessary in order tocodify the rules with member-ship degrees of each sets forall magnitudes, i.e., instead ofhaving binary digits (0, 1), theinput pattern will be vectors ofreal numbers in range [0, 1]that represent the member-ship degrees of the inputs ineach set.

The results of the outputlayer are real numbers

between zero and one; thus, the controller will roundthese results to an integer-type value in order to find theproposed status for the CBs.

Experimental ResultsTo carry out the experiment, a software for the platformWindows 9x/2000 using Delphi was elaborated. Signalgenerators for the analog input variables were used.

The experiment was started with normal values ofmeasurements starting from the status 00111. During thefirst 400 measurements, 243 actions could not be deter-mined by the controller, i.e., the answers were given byan expert, and, as a result, 243 new controller rules wereextracted.

Figure 6 shows the status behavior of the plant duringthe first 400 sam-pling times withextreme simulatedvariations in theinputs. The statusfrequency is shownin Table 2.

The followingcan be said aboutthe comparison ofthe plant behaviorwith and withoutthe controller. Thiscontrol system has

Figure 5. Information flow during the initial status of the controller (dynamic rule genera-tion after human operation combined with periodical controller outputs) and the fullautomatic stage (after learning the rule base by means of an artificial neural network)

LearningTerminated

?

Training Stop

Learning AlgorithmStand-AloneStage

New Rule Generation

Human OperatorAlarmN

N

Y

Y

SystemProtection

Control Devices

SolutionFound?

Interference

Rule Base

Initial Stage

Rule BaseCompleted orTime Criterium

Measurementsand State

SubstationInput Parameters Output Parameters

N

Y

ArtificialNeural Network

January 2002 45

Table 2. Frequency of topologies obtained during the first 400 system operations

Topology Frequency 0 (=00000) 0.019

4 (=00100) 0.009

6 (=00110) 0.102

7 (=00111) 0.046

20 (=10100) 0.93

22 (=10110) 0.056

23 (=10111) 0.009

28 (=11100) 0.491

30 (=11110) 0.157

31 (=11111) 0.019

Page 6: AI automates substation control

been designed for human operator replacement, i.e., todecide about actions beyond the conventional automaticcontrol procedures, i.e., actions for which a human oper-ator is always needed. Since the implemented controllerresponds according to a knowledge base extracted byhuman operations, the topology change will always bethe same (with or without the controller). Nevertheless,the difference and advantage due to the use of this con-troller reside in:� Speeding up response time� Avoiding operation mistakes� Gradually replacing a human operator.

The network was trained with a group of 1,377 rules. Itwas found that, for activation functions of lineal type(first layer) and sigmoidal type (hidden and last layer),an approximate time of 2 hours using a 455 MHz Pen-tium-based PC was sufficient to obtain approaches witherrors smaller than 0.1 for each output neuron. Since thefinal result of the approximation by means of the net-work is based on whole rounds to zero or one, this levelof accuracy was found acceptable.

Fast Response and Diminished RiskFor the control of a substation by means of the connec-tion of devices for improving its performance, it is neces-sary to keep in mind not only the measurements of theelectric magnitudes but also the status of some controldevices that define their topology.

Control and protection devices governed by individ-ual decision methods that allow the intervention of thesubstation operator when needed can be included in acontrol system so that their operations can be coordi-nated from an upper supervisory level. It is possible tofollow a preference criterion for the disconnection orconnection of CBs (in this case for load feeders and fil-ters) when using a controller based system.

Controllers can be used for rule extraction in a firstworking stage. Control systems that do not requireinstantaneous responses in an initial stage, can bedesigned for learning the operator actions and con-structing a decision table for total replacement of thehuman operator in the future.

Even when the number of rules for controlling the

substation is very high, it is possible to obtain theserules automatically by means of a fuzzy controller. Aftera certain operating time (which depends on the initialknowledge base as well as on the status variables andthe fuzzification of the inputs), the inference processthrough a rule base can be replaced by an approxima-tion via an ANN to diminish response time.

The number of necessary magnitudes to superviseand to control a substation can be very high. In the pre-sent research work, many magnitudes were not includedto avoid the extensive number of required rules.

Nevertheless, controlling a substation by a fuzzy con-troller has the advantage that it can speed up theresponse time and diminish the possibility of risks nor-mally related to human operations.

AcknowledgmentThe authors gratefully acknowledge the Research Foundation of TheState of São Paulo (FAPESP) for supporting this research work.

For Further ReadingA.G. Lima, et al., “Cost of power quality problems in large industrialcustomers,” PQA ’93, San Diego, Oct. 1993.

J.A. Jardini, “Automation in power plants and high voltage substa-tions,” Editorial EB/USP, 1st ed., São Paulo, Brazil, 1997.

K.S. Fu, “Learning control systems and intelligent control systems,”IEEE Trans. Automat. Cont., vol. 16, pp. 70-72.

K.Y. Lee, “Current trend and the state of the art in intelligent sys-tem applications to power systems,” 1999 International Conference onIntelligent Systems Application to Power Systems (ISAP), Rio deJaneiro, Brazil, Apr. 1999.

A. Patrikar and J. Provence, “Control of Dynamic systems usingfuzzy logic and neural networks,” Int. J. Intell. Syst., vol. 8, no. 6, pp. 727-748, 1993.

BiographiesMelvin Ayala S. is a researcher at the Polytechnic Institute of Havana,Department of Computer Sciences, Cuba. He received his Ph.D. in 1987.His research interests include software development, fuzzy logic, andartificial neural networks.

Galdenoro Botura Jr is a senior professor at the FEG, UNESP,Department of Electric Engineering, Guaratinguetá, São Paulo, Brazil.He received his Ph.D. in 1991. His research interests include electronicand telecommunications.

Oscar A. Maldonado A. is a senior professor at the FEG, UNESP,Department of Electric Engineering, Guaratinguetá, São Paulo, Brazil.He received his Ph.D. in 1991. His research interests include power sys-tems and automation.

46 IEEE Computer Applications in Power

Figure 6. Register of the first 400 system outputs (maintained or changed status) starting form the initial status00111 (=7) and the following parameter (mean -m- and variance -S-) of the signal generator (Vout(V): m=115, S=15,PF: m = 0.93, S = 0.02, THDv(%): m = 5, S = 1)

30

25

20

15

10

159 16 26 36 46 56 65 75 85 95 107120 134 148 162 175 189 203 217230 244 258

Number of Operations

State

272 286 299 313 327 341 354 368 382 396

5

0