Using a neural network to predict the dynamic frequency response of a ...

Using a Neural Network to Predict the Dynamic Frequency Response of a Power System to an Under-Frequency Load Shedding Scenario Matthew A. Mitchell Student Member INESC Porto Faculty of Engineering University of Porto Polto. Portugal

J.N.Fidalgo

J.A. PeGas Lopes Senior Member m c Porto Faculty of Engineering University of Port0 Porto, Portugal

Abstract - This paper proposes a method to quickly and accurately predict the dynamic response of a power system during an under-frequency load shedding scenario. Emergency actions in a power system due to loss of generation typically calls for under-frequency load shedding measures to avoid potential collapse due to the lack of time in which to correct the imbalance via other means. Due to the slow and repetitious use of dynamic simulators the need for a fast and accurate procedure is evident when calculating optimal load-shedding strategies. A neural network INN) seems to be an ideal solution for a quick and accurate way to replace standard dynamic simulations. The steps taken to produce a viable NN and corresp~ndingresults will be discussed.

James D.McCalley Senior Member

INESC Porto Faculty of Engineering University of Porto Porto,Portugal

Dept. of Electrical and Computer Engineering Iowa State University Ames, IA.USA

The dynamic analysis of an electrical network after a disturbance is a very time demanding task. Thus in recent years a large effort has been made toward developing fast approaches to deal with the prediction of system dynamic behavior, using namely neural networks [I], decision trees [2], and regression trees [3]. In this paper, a NN is proposed to predict the systems dynamic frequency response through the motivation of trying to identify load shedding strategies that will lead to minimum amounts of load to be shed. In section I1 the current under-frequency load shedding strategies will be discussed. In section III the steps taken to create the proposed neural network will be addressed, followed by numerical analysis and conclusions in sections IV and V respectively.

Keywords - Neural network, under-frequency load shedding,

dynamic system response. I. INTRODUCTION Dynamic security is a crucial issue with respect to the operation of power systems, especially isolated systems due to their natural weaknesses. Following a system disturbance, a fast assessment of the systems robustness in reference to dynamic behavior and the identification of effective emergency control measures that prevent system collapse are most important. Throughout the world there are numerous medium size isolated networks, especially on islands, but also in some continental regions where, due to economical reasons, it is not feasible to have interconnections. In some of these systemq, the contribution of wind power is playing a more significant role as it provides a cheaper and independent way of producing electricity. However. in these cases, the vulnerability of the system may increase if large penetrations of wind power are foreseen.

11. STATE OF THE ART

A. Load Shedding

As already mentioned, when a power system disruption creates a large generation load imbalance, resulting in a frequency decline, emergency action such as underfrequency load shedding may be needed. If system frequency reaches a given threshold, even for a short amount of time, power stations may trip off resulting in further load imbalance which may lead to a global system collapse. When there is a rapid decline in frequency, simple governor response may be neither sufficient nor quick enough to stop the frequency excursion before it reaches the protection threshold of frequency relays in other power plants. Thus, there is a need for a complementary emergency action in order to assure that the declining frequency is stopped before reaching this threshold. Although load shedding is usually effective, problems can arise due to ineffective shedding, which can lead to system collapse anyway [4]. Therefore the response of the power system to a frequency decline needs to be understood in order to judge how these influences will effect the steps taken in preparing an efficient and swift load shedding scheme [4,5,6]. Typically load shedding schemes are developed following the same basic design, in which under-frequency relays,

In any isolated sysiem, the loss of a production facility will lead to a severe disturbance that must be effectively compensated using the spinning reserve. Assuming enough spinning reserve is available, generator frequency regulators must respond fast and effectively, otherwise emergency measures should be taken, like load shedding, to avoid a frequency system collapse.

0-7803-6420-11001$10.00 (c) 2000 IEEE

Member

346

Table 1 Frequency Variable

Knowledge Acquired determines amount of load needed to be shed in order to, minimize Initial frequency

determines the first relay’s frequency setting and time delays, in order to insure a quick response

will be dose to nominal

I

Maxjrnum Positive Devlation

df,dt

Residua Frequency

I

helps minimize over-shedding thus avoiding the activation of any over-frequency protection devices, thus contributing more losses to the disturbance assesses the sevefity and location of the disturbance determines the amount of load to be shed, In order to insure that the resldual frequency Is above the deslred threshold

The procedure to be followed in this case involves four main steps. 1. Identification of input / output relevant variables 2. Data set generation 3. Design of the NN 4. Performance evaluation of the neural nets

In reference to the output variables, and as mentioned

Identification of input / outputs

before, the interest in knowing the maximum negative and positive frequency deviation. the final system resting frequency, and the slope of the first negative swing lead to the selection of these 4 variables as outputs of the NN.

The identification of the variables that are going to characterize a given operating scenario is an important step for a successful application of these techniques. Sometimes a pre-processing stage is needed to select the most relevant variables to be used as inputs of a NN. In this work, we decided by just selecting a set of meaningful variables considering the following issues: availability from SCADA, direct physical meaning towards the phenomenon, and potential controllability. Having this in mind, in this research the following variables have been used as inputs of the NN. 0 Actual real power generation; 0 Available real power (effective spinning reserve); Active load generation level prior to disturbance; Amount of active load being shed; Percentage of exponential type Loads being shed.

b

B.

Generation of the Data Set

The replication of a given power systems response through any machine learning technique, like a NN, can only be accurate if the data used to train these structures describes with enough coverage and quality the different operating conditions. Optimally. this data set would include all possible system scenarios, however this would require unrealistic hours of computational time. Therefore the objective of the data generation stage is to capture the breadth of the system operating conditions and behavior, while limiting computational and engineering efforts. This data set includes the data used for training a NN and the test data for evaluation purposes.

Supplying the NN with actual and available real power enables it to make the needed correlation between the dynamic frequency response and system generation and effective spinning reserve levels.

Due to the potentially numerous system disturbances that could lead to a load shedding situation it is unfeasible to create a NN that could predict the system.. dynanuc behavior for all of them. Therefore, we decided to develop the approach so that a single predicting structure will be used to deal with a single pre-specified disturbance.

With respect to system loading levels and having in mind the characteristics of isolated systems (single dynamic area), individual loading levels were not needed. However, since a complex load model [ 161 is being used within this work and each load type has its own profound influence on the systems frequency response it is critical that each load type is used separately as an input to the NN, so it can produce more accurate conclusions. Thus load level inputs to the NN can be broken down into total real ZIP (constant impedance, constant current, comtant power types), exponential 1, and exponential 2 load types.

0-7803-6420-1/00/$10.00 (c) 2000 IEEE

I

determines if any other system IntewenUon will be needed to recapture nominal frequency

Finally concerning load shedding variables. the total amount of real load, percentage of exponential types 1, and 2 that are being shed were used. These variables provide the NN with valuable information, such that it can make the needed assessment with respect to how much the generation-load imbalance has been corrected and the influence each load type has on the resulting frequency response. Therefore. for the case under study, there are a total of 21 inputs to the NN.

The next sections provide a description of the procedure adopted to deal with each one of these steps. A.

I

~~

determine the time delays needed, or what delays to expect from df/& sensflive relays

Having in mind the particularity of the network under analyses, namely that the wind parks in the network constitute its weak point. the following disturbance was considered: a short circuit in the lines connecting to the wind park, producing afterwards the disconnection of this production facility, and hence creating a generation-load imhalairce.

347

assuming instead a worst case scenario. Therefore, actual scenarios frequently lead to excess load shedding. Utilities would like to minimize the amount of load to be shed, therefore the load shedding scheme should change according to the system operating conditions. This means that investigation is needed in the field of defining remote settings of U F L S relays.

at the feeder level, are set to operate at different triggering frequencies and time delays, in order to arrest falling system frequency [4,7,8,9,10,11]. Analyzing the operating conditions and disturbances that may be encountered throughout system is the fvst step in developing successful load shedding strategies. After a thorough understanding of the system is attained, the amount of system overload for which the load shedding scheme will be designed needs to be selected. Next an acceptable value for the post-shedding residual (resting) frequency threshold of the system needs to be established. This value is of great importance since it not only affects the amount of shedding to plan for but it also influences the potential for further intervention, by manual or automatic system devices.

In order to deal with the minimization of the load curtailment, a one step shedding scheme produces a better frequency response than a stepping scheme, due to its one time bulk shed versus cascade shedding [13]. The identification of the new relay settings must be done in a fast way, thus a new tool able to assess and predict system behavior in these conditions must replace the conventional dynamic evaluation.

Since system load and generation are constantly changing to meet new consumer demands, the use of stepping schemes and different relay settings to stop declining frequency is best-suited [ 121.

In the work reported in this paper we are describing a new approach to deal with the remote setting of UFLS relays, to be applied in isolated systems that have large shares of wind power production. In this approach two frequency settings are considered, such that the first setting is optimized and remotely defined and the second is kept pre-defined and is to be used as final resource to save the system if the first measures fail.

In this work we are assuming that the operation of the system is based on AGC that is in charge of defining globally the final generation set points. The required amount of load shedding, in case of enough spinning reserve in the system, is given by (1).

111.

RESPONSE

Prs = AP, - KAfss

Here

The need to determine and evaluate the dynamic frequency response of a system during an underfrequency load shedding scenarios is of the utmost importance. Establishing an effective under-frequency load shedding scheme typically hinges on knowing what to expect from a given disturbance, especially in regards to the maximum negative and positive deviations, the residual frequency, along with the slope of the first negative swing. Prediction of these values can be done using machine learning techniques as described in [ 141.

C P U , when positive, is the required amount of

load shed, dp, is the generation lost, K is the systems stiffness, and Afss is the maximum tolerable steady state frequency after shedding. In situations when there is not enough spinning reserve the amount of load to be shed is given by (2).

Obtaining knowledge about the system behavior after a shedding scheme, following a given disturbance will help in deciding between potential load shedding schemes. In regards to specific frequency information. a summary of the knowledge that can be acquired is described in table 1. In situations where many evaluations of the dynamic behavior of a system are needed, like in this case of planning a load shedding scheme, the time taken by dynamic simulators to perform complete frequency evaluations is overwhelming. Therefore, machine learning techniques are again well suited to deal with this problem. NNs are particularly appropriate to do so, due to their ability to map a nondescriptive function (like the one associated with the system dynamic behavior) and to provide fast response with sufficient accuracy. Their success when dealing with similar problem.., for example the prediction of the operation of under frequency load shedding relays [IS]is well known.

Where SR is the amount of spinning reserve left [7]. The decision about the number of shedding steps and their settings to deal with an emergency situation takes long man hours of off-line planning and testing, analyzing system response obtained through set-by-step simulation, to come up with a well suited comprehensive plan. With this type of strategy, once fixed the settings are kept until a new assessment is made and it is determined that a different and “bette? scheme is required in order to fulfill new system security and reliability needs. With the onset of a more competitive power market and the need for better system reliability, reliance on these pre-determined load shedding schemes may not be satisfactory. Namely, step type schemes do not consider current system conditions or new topologies, due to growth in system size, nor do they optimize the load shedding selection,

0-7803-6420-1/00/$10.00( c ) 2000 IEEE

PREDICTING THE DYNAMIC FREQUENCY

348

Wind Gen 1 c 1 3 iOUVA

()T

10 kV 6

These results are evidence that the NN has the capability to predict the dynamic frequency response of the power system during an under-frequency load shedding scenario.

c\, 5Mvar

A comparison was made with a dynamic simulator, ETMSP, in which a loss of 7.9 MW of wind power production was simulated. The numerical results obtained from the dynamic simulator and the NN can be seen graphically in figure 2 and numerically in table 3. As was expected, the NN provides results that match quite well in comparison with the ones obtained from full dynamic simulation given through ETMSP.

+

6.6kV

1

-Shcdl.R

I

.OF+-

@

::s

(MVA

b 8 a

p

All conventional generators are synchronous machines and are mathematically described using a second order model;

48

Mu.Rm Max. Neg.

30

0.236

0.298

.

Case Il

0.062

1.203

1.192

0.011

2.185

2.147

0.038

2.055

2.045

0.009

dVdl(Hdx)

1.357

1.375

0.018

1.957

1.929

0.028

,,-yFw,

49.434

49.468

0.034

49.991

49.999

0.008

nrui.l,i,.nm-,

V.

CONCLUSIONS

This paper described the application of a NN to make a fast prediction of the system behavior for a load imbalance disturbance followed by a load shedding control action. The excellent results obtained in an isolated system show the applicability of this method for the purpose of evaluating the quality of different load shedding schemes for a pre-specified disturbance. Due to the fast prediction of system behavior, the dispatch center can select the feeders to be disconnected in emergency conditions (meaning the amount of load to disconnect). This can be done by adapting the settings of the feeder UFLS relays periodically, according to the range of operating condition foreseen within a day time horizon.

Test

0.037 0.041 0.0R9 0.101 0.0475 0.0503 0.1130 0.1235

0-7803-6420-1/00/$10.00(c) 2000 IEEE

I

DcvLabn(Hr)

RMS Train

2S

Neural Network vs. Dynamic Simulator

b

Neural Network Performance Test

20

I5

Table 3

Table 2

Train

1u

Figure 2. System Frequency Response

After the design of the NN (adopting the architecture described before), a performance evaluation stage was implemented. These results can be seen in table 2. As it can be seen from this table, the NN's training performance was quite satisfactory and provides a good predictability with respect to the test set.

Train Test Train Test

5

lime (sec)

Using the data set generation procedure described in section 111, 14,046 operating points have been obtained which provides an excellent picture of the system behavior.

Max'Pos' Deviation

4 0

All applicable UFLS relays were set to trigger at a frequency of 49Hz with a 3cycle time delay.

Abs. Deviation

49

41.5

The system analysis is being performed after governor response, but before new AGC determined set points can be established All system loads were modeled using the composed model referenced previously;

R M S Error

49.5

4R.S

In all synchronous generators, automatic voltage regulators and governors are adopted using IEEE models; It was assumed that an AGC is used for establishing all generators governor final production set points:

Mean

f

51

Figure 1. Terceira Network

Variable

NN Point&

9

,

51.5

Hydro Gen 1

Frequency

MW

349

EPRI’s Output Analysis Program (OAP) was used for numerical analysis and creation of the NN target variables.

The procedure used to develop the data set is laid out as follows. 1. Generate system operating scenarios; 2. For each scenario, perform a load-flow simulation to guarantee operating point feasibility and get a complete characterization of that operating point

C.

The first step in the design of a NN is to determine an architecture that will yield good results. The idea is to use the simplest architecture while maximizing performance. Usually, NN architecture is determined based on subjective assessment on the part of the engineer.

(OP); 3. 4. 5.

6.

Define shedding scenarios, for each OP, considering that the previous disturbance took place; Perform dynamic analysis; Retain frequency behavior parameters; Save variables characterizing the OP and frequency behavior parameters.

Within this work it was concluded after a few trials that an architecture of 2 hidden layers, the first with 14 nodes and the second with 10 nodes, was best suited for this application.

Each of these steps will be discussed in further detail in the paragraphs that follow.

An Adaptive Back Propagation technique was used to train the NN [20]. This consists of the same routine as typical back propagation with the exception that instead of one learning rate for all the NN nodes, a learning rate was assigned to each of the nodes in order to speed up convergence. The activation function used within this work was a hyperbolic tangent function and the inputs were normalized to have a mean of zero and a variance of one.

A structured Monte Carlo sampling technique was used to generate the system operating scenarios [17,18] with an adaptation for the purpose of localizing these scenarios in order to limits its size, yet maintain its effectiveness. A load curve was established in order to decrease the need for scenarios outside of the established curve and instead focus the generation of scenarios to areas most probable to be realized in actual system conditions, i.e. the load curve.

D.

After obtaining a complete set of scenarios covering the probable range of operating conditions, a loadflow simulation is performed on each to ensure feasibility and calculate system losses. This step was achieved using EPRI’s IPFLOW software.

+ Ploss *buffer

If the training set provides good results, in terms of accuracy, and the test set does not, this generally indicates over-fitting in the learning stage and/or that the current NN structure is too complex and needs to be simplified. On the other hand if the training set and test set provide comparable results, but not satisfactory ones regarding the user this generally implies that a more complex structure should be tried. Once desirable results are attained from both the training and test data, a comparative evaluation can be made with the dynamic simulator.

(3)

Iv. NUMERICAL RESULTS In this work, the Terceira Island system (Azores Portugal) has been selected for testing the approach developed, considering the large amount of wind power production foreseen and the need for effective emergency control measures. Its single line diagram is presented in figure 1.

Here, Shed m a , is the maximum allowable load to be shed, Ploss, is the amount of generation lost and BufJer, is the percentage of extra load to be considered for shedding. No load slzeddinr! is also an acceptable load shedding scheme, in order to establish a base frequency response for a given operating condition. such that planners can see the benefit of shedding. EPRI’s, Extended Transient-Mid-Term Stability Program (ETMSP) was used to perform the dynamic simulation. 0-7803-6420-1/00/$10.00 (c) 2000 IEEE

Evaluation of NN performance

In order to get an idea for what kind of performance is to be expected from a NN architecture, a preliminary evaluation is needed of its capabilities. The “training set” data, typically composed of % of the OPs from the overall data set, is used to teach the NN and give a relative inclination as to its suitability for that particular application. The “test set,” which is comprised of the remaining data, is used to evaluate the prediction capabilities and generalization performance of the structures.

Since the scope of this work is to take into consideration the effects load shedding has on the power systems dynamic performance, it is crucial that all potential load shedding schemes are analyzed via a dynamic simulation and included in the data set. Therefore for one given OP a number of different potential load shedding schemes were developed, depending on the amount of generation that was lost. However, due to the overwhelming shedding scheme potential a few additional steps were taken to minimize the number of needed schemes. First, although load shedding is usually performed at the feeder level, an aggregation of feeder loads into one busload was performed, using the load modeling procedure addressed in [19]. Second, a shedding constraint to avoid scenarios that cause too much load shedding was applied, as seen in (3). Shed max = Ploss

Design of NN

An island model allows for a compact, simple, and precise system analysis. The following is a list of assumption and device models u e d in the dynamic simulation of Terceira network. 350

[I61 lEEE Task Force Report, ‘Toad representation for Dynamic Perfonnmce.” IEEE Tran, on Power Systems, Vol. 7, No. 1, Feb.1992.

A NN designed in this way is presently being used to identify optimal shedding levels, using a Genetic Algorithm approach. Results obtained from the application of this philosophy are also quite promising. VI.

[ 171 McCalley, J.D., et al.. “On-Line Visualization of Transmission

System Operating Constraints using Intelligent Infomtion Processing.” Final report of contract No. 219-2-384-94 between Pacific Gas & Electric and Iowa State University. May 1997.

REFERENCES

J. Pqas Lopes. J. N. Fidalgo, V. Miranda, N. Hatziargyriou. “Neural networks used for on-line dynamic security assessment of isolated power systems with n large penetration from wind production- A real case study“, hoc. of Rough Sets and Soft Computing Conference’W, San Josk, USA, November 1994.

[I81 Van Acker. Vincent, e t al. “Data Generation using Automated Security Assessment for Neural Network Training,” presented ai North American Power Symposium, L;lramie. WY., Oct. 97’ [19] Hajagos, Les M., Danai, Behnam. “Loboratory Measurements and Models of Modern L d s and Weir Effect on Voltage Stability Studies,” in IEEE Trans. Power Systems. Vol. 13, No. 2, May 1998. pg. 584-592.

N. Hatziargyriou, S. Papathanassiou, M. Pqmdopoulos, “Decision Trees for Fast Security Assessment of Autonomous Power Systems with large Penetration from Renewables”. IEEE Transactions on Energy Conversion, Vol. 10, Nr. 2, June 1995.

[20] Silva. EM.. Almeida, LB., “AcceleratioOn Techniques for the Backpropagation Algorithm,” In Neural Networks. L.B. Almeida and CJ.Wellekens (Eds.). Springer-Verlag, 1990.

J. A. P q a s Lopes, M. H. Vasconcelos, ‘‘On Line Dynamic Security Assessment Based on Kernel Regression Trees”, P m . IEEE winter power meeting 2000, Singapore. January 2OOO.

ACKNOWLEDGMENTS

Clark, H.K. “Time Simulation in Load Shedding System Design,” Power Technologies, Inc.

The authors would l i e to thank the Power System5 Unit of the Dept. of Electrical and Computer Engineering at Iowa State University (USA) for the funding they have contributed. They would also like to thank the financing obtained from the Portuguese research fund PRAXIS XXI, project no 2/2.11TIT/1639/95.

CIGRE Task Force 38-02-14, ‘large Frequency Disturbances: Analysis and Modeling Needs,” (To be published) Anderson , Paul M.. “Power System Protection,“ Final Draft ... (To be Published) Prasetijo. D., Lachs. W.R.. Sutanto, D.. “A New h d Shedding Scheme for Limiting Underfrequency.” LEEE Trans. Power Systems, Vol. 9, No. 3. Aug. 1994.

BIOGRAPHIES

Matthew A. Mitchell is from Marion, Iowa, U.S.A. He graduated with a B.S. in Electrical Engineering from Iowa State University in 1998. He is currently enrolled in the MSc program at the University of Porto and a researcher of the Power Systems unit at INESC Porto, Portugal.

Concordia, Charles. Fink, Lester .H.; Poullikkns. George, “Load Shedding on an Isolated System.’’ IEEE Trans. Power Systems Vol. 10, No. 3. Aug. 1995.

.

Lokay, H.E., Burtnyk. V., “Application of Underfrequency Relays for Automatic Load Shedding.” IEEE Trans. Power App. Sys., May 1968.

JoHo A. P e p s Lopes was born in Portugal. He graduated in Electrical Engineer from the Faculty of Engineering of Porto University (FEUP) in July 1981. and obtained the Ph.D. and Aggregation degrees also from FEUP in October 1988 and November 1996 respectively. Dr. P e p s Lopes is presently Associate Professor with Aggregation at FEUP and Assistant Coordinator of the Power Systems unit at INESC Porto, Portugal.

Maliszewski. R.M.. Dunlop, R.D., Wilson, G.L., “Frequency Actuated Load Shedding and Restoration Piut I- Philosophy,” IEEE Trans. Power App. Sys.. July/Aug. 1971. Horowitz. S.H.,Politis. A., Gabrielle, AE., “Frequency Actunted Lond Shedding and Restoration Piut II- Implementation.” IEEE Trans. Power App. Sys.. July/Aug. 1971. Chuvychin, V.N., Gurov. N.S.. Venkata. S.S., Brown, R.E., “An Addaptive Approach to Local Load Shedding and Spinning Reserve Control During Underfrequency Conditions,” IEEE Trans. Power Systems, Vol. 11, No. 4, Nov. 1996.

J.N. Fidalgo was born in Portugal. He graduated in Electrical Engineer from the Faculty of Engineering of Porto University (FEUP) in July 1985. and obtained the Ph.D in 1995 also from FEUP. Presently he is an Assistant Professor at FEUP and a Senior Researcher of the Power Systems unit at INESC Porto, Portugal.

[I31 P y a s Lopes, J.A., et al.. “Genetic Algorithm for Planning Optimal Load Shedding Strategies,” presented at IEEE Power Tech 99’ Conference, Budapest, Hungary, Aug. 29 Sept. 2, 1999. [I41 N. Hattzixgyriou, J. A. P e p s Lopes. E. Karapidakis. M. H. Vasconcelos, “On-Line Dynamic Security Assessment of Power System. in Large Islands with High Wind Power Penetration”. Proceedings of PSCC’99 - 13th Power Systems Computation Conference, vol. 1. Trondheim - Norway, June 1999. pp. 331-

James D.McCalley is an Associate Professor of Electrical and Computer Engineering Department at Iowa State University, where he has been employed since 1992. He worked for Pacific Gas and Electric Company from 1986 to 1990. Dr. McCalley received the B.S. (1982), M.S. (1986), and Ph.D. (1992) degrees in Electrical Engineering from Georgia Tech. He is a registered professional engineer in California and a senior member of the IEEE.

337. [I51

Kottick, Daniel. Or Ofer. “Neural-Networks for Predicting the Operation of an Under-Frequency Load Shedding System,” IEEE Trans. PWRS., Vol. 11. No. 3, Aug. 96‘

0-7803-6420-1/00/$10.00 (c) 2000 IEEE

35 1