Вы находитесь на странице: 1из 9

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

Power System Adequacy and Security Calculations Using Monte Carlo Simulation incorporating Intelligent System Methodology
Chanan Singh, Fellow, IEEE, Xiaochuan Luo, and Hyungchul Kim
Abstract--Monte Carlo Simulation has been extensively used in reliability evaluation of electric power systems. One of the issues with this approach has been the computational time for convergence of indices being estimated, especially when the approach to deal with this problem has been some version of variance reduction techniques. Recently some publications have proposed use of intelligent systems techniques such as selforganizing maps and linear vector quantization to tackle this problem. This paper will provide a perspective on this hybrid approach using Monte Carlo Simulation and intelligent system methods. The philosophy of this hybridization as well some
systems are highly reliable. Perhaps the most commonly used

residence times. Some research has also been done combining analytical methods and simulation methods for power system reliability studies [11-13]. Also more recently some methods of intelligent search of state space using genetic algorithms have also been shown to have promise [19- 20].

The main advantages of Monte Carlo simulation include: 1) the ability to model complex systems in more detail and accuracy than is possible in analytical methods; 2) the required number of samples for a given accuracy depends on the variance which tends to decrease with the system size; 3) the method can not only calculate the expected value of reliability indices but also their distributions. The primary results will be discussed. disadvantage of Monte Carlo simulation is the lengthy Index Terms--Reliability, Monte Carlo Simulation, Self computational time to achieve satisfactory statistical convergence of reliability index values. One of the time Organizing Maps, linear vector quantization, neural nets. consuming procedures in Monte Carlo simulation is the I. INTRODUCTION characterization of the sampled states. Additionally many The basic intent of power system reliability analysis is to similar states may be repeatedly sampled and their determine some probabilistic measure of the undesirable characteristics determined. These two factors make Monte events. Several methods of power system reliability analysis Carlo simulation time consuming. Two methods employing Self-Organizing Maps and Monte have been proposed as can be seen from a selected list of references [1-13]. These methods fall into two broad Carlo simulation are proposed to evaluate reliability of power categories: analytical and Monte Carlo simulation methods. systems. SOM has the ability to project input data from an nThere are three stages inherent in any reliability method - dimensional space to a lower dimensional (usually twostate selection, state evaluation and index calculation. The dimensional) space while maintaining the original topological analytical techniques and Monte Carlo techniques differ relationships. By combining SOM and Monte Carlo mostly in the process of state selection as the number of simulation, we can greatly reduce the computing burden of possible states is extremely large for most practical Monte Carlo simulation. applications. The analytical techniques use some device to h I. MONTE CARLO SIMULATION circumvent the problem of straightforward enumeration such as state merging, truncation, implicit enumeration and Several reliability indices have been proposed in the sequential model building. The Monte Carlo techniques literature. Perhaps Loss of load probability (LOLP) is the most accomplish this by sampling states proportional to the widely known power system reliability index. This section probability of their occurrence. Analytical techniques gives various steps of Monte Carlo for the LOLP index but the represent the system by mathematical models and compute same process can be used for other indices. Typical steps for reliability indices using mathematical solutions. Monte Carlo LOLP calculation using simple Monte Carlo simulation can be simulation methods, which treat the problem as a series of described as follows. experiments, estimate reliability indices by simulating the actual process using probability distributions of state Step 1: Select the seed for the random number generator. Set the maximum iteration number and let the initial iteration number k =1; Stp2Salehestmsaernd ly(odev, C. Singh is with the Department of Electrical Engineering, Texas A&M University, College Station, TX, USAStp2Sapetestmsaernd ly(odev, X.Luo is with the ISO New England Inc, Holyoke,MA 01140. generation status and line status) and perform a flow
H. Kim is with Korea Rail Road Research Institute, Seoul, Korea.

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

calculation to classify it as loss-of-load or otherwise.

{I sampled state is loss-of-load

Qgi -(Qi-L
(1)

CUT,* p ) V,VjY.sin(O. +,j-,5)=O


(7)
<prnax prif g<p min m0nax Qgi. <Qg, .QgS
S

Zn

Xi=
| Ootherwise

(8)

Step3:Calculate LOLP, variance of the estimated LOLP and the coefficient of variation.

max

9 (9) (10)

LOLP = k

(2)
(3) where:
L ,

V(LOLP) = (Z_X2 - LOLP2)


k k

0<L I-~~~~~~~~~~~~~ _CUTj < ~~Pli ~ ~

Vin<V<vmax

(11)

(12)

V(LALP) LOLP 7()


Step 4: Check whether the coefficient of variation cG is less than a specified threshold 6 . If c7 < 6 or k > Kmax stop; otherwise, k=k+1, go to step 2. From the steps of the straight Monte Carlo simulation, we can make two observations: 1.For each sampled state, determination needs to be made whether it is a loss of load state or not. This typically needs a flow calculation to be made. Depending on the method of flow calculation, this step could impose a significant computational burden. 2. Because of the random sampling, many similar states are sampled in simulation and their characteristics determined repeatedly. Therefore, the straight Monte Carlo simulation could be very time consuming.
III. Loss OF LOAD IDENTIFICATION

CUT,:

The amount of real power load curtailed at bus

Magnitude

Vi, 6i: Pgj, Qgi:

PI', Ql:
Sj1

and angle of iji element of the admittance matrix Y Magnitude and angle of voltage at bus i Real power and reactive power output of generation at bus i Real power and reactive power load at bus i

Apparent power flow in transmission line between bus i and j pm11npmax Minimum and maximum real power output Of g gi generator i Q9f Q max: Minimum and maximum reactive power output of generatori Vmin, Vima: Minimum and maximum of voltage magnitude

As described in section I, in Monte Carlo simulation for power system reliability calculation, we need to classify a sampled system state as either loss-of-load or not. Various network flow methods have been used to determine whether a system state constitutes loss-of-load or not. Typically for multi-area reliability studies, linearized network flow model or DC load flow are used. In the composite system reliability studies, either DC or AC load flow are used for flow calculations. In this paper, optimal power flow (OPF) is used to minimize load shedding at each bus subject to a set of constraints. The result of the objective function minimization can then be used to determine if a system state is loss-of-load or not. The problem can be stated as follows [15]. Minimize:
N

S, s
N:

max

at bus i Maximum allowed apparent power flow in

transmission line between bus i and j Number of buses

If the objective function in the above formulation is zero, the power system state does not result in loss-of-load. If the objective function is greater than zero, then the system state does constitute loss-of-load. Furthermore, the formulation above not only characterizes a system state as loss-of-load or not, but also gives the minimum amount of load which must be shed at each bus if the system state results in loss-of-load. This information can be used to determine EUE(Expected Unserved Energy) index.
IV. INTELLIGENT SYSTEM METHODOLOGIES

Subject to: Pgj - (P1, - L CUT )- E/ VjVjYjj cos(Ojj + dI- 3, ) 0

j=4

Z L CUTj

(5)

jeN

(6)

The intelligent system methodology that will be used here is a unsupervised neural network model called the SelfOrganizing Map(SOM) [14] that usually implements a nonlinear "projection" of input signals from a highdimensional space to a lower-dimensional space. The mapping often automatically finds and displays characteristic features from the raw data. A SOM is formed of neurons on a 2dimensional grid. All units in the input array are fully connected to the neurons in the feature map. Each neuron i is represented by an n-dimensional weight vector wi =[ wi0 wit

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

the map. That is, the network will map an input vector to the ... wi(n-1)] while the input vector to the SOM is represented by x = [xO xl ... xn- 1 ] where n is the dimension of the input nearest neuron. vector. The map's topology can be a rectangular lattice or a Parameter Choices for SOM hexagonal lattice. See Fig. 1 for the structure of the SOM. There are several options in training the Self-Organizing

Competitive (Kohonenj Layer

Map and in choosing parameters used during the training


process. These options are: 1) Map dimension (x-dimension and y-dimension) 2) Map topology type (rectangular or hexagonal lattice)

3) Type of learning rate oc(t) and its initial value:

Linear type: o(t) =o(0) * (1.0- l ), where rlen is rlen the running length of the training process. Inverse-time type: a(t)=a(0) *C , where C is
Fig. 1. Structure of Self-Organizing Map

Input Layer

V. TRAINING ALGORITHM FOR SOM The SOM training algorithm is implemented in the following steps: Step 1 Initialize the weight vectors wIj at t=0 with small random values for j 0,1,2, ... m-1, i 0,1,2, ... n1, where m is the number of neurons in the map and n is the dimension of input vector. Also set the initial radius of the neighborhood. Step 2 Present the input vector x to the map. Step 3 Compute Euclidean distance dj between the input vector x and each neuron j in the map using
= =

in the program implemented. 100 4) Neighborhood type (bubble or Gaussian) There are two possible choices for the neighborhood kernel h C(t): a Gaussian function or a "bubble" function. The Gaussian function is:

defined as C =

rlen

hcj2
a

where r

(17) is the location of neuron j on the map grid,


(t)

||c-r||

In-i

dJ = E(XI (t) -w ji (t)

(13)

where x (t) is the input vector at time t and w p (t) is the weight vector from input component i to neuron j at the time t. Step 4 Select the winner neuron c (best matching neuron) with minimum distance. c = arg min { dj } (14) J Step 5 Weights are updated for the winner neuron c and all neurons in the neighborhood kernel defined by h Ci (t) (15) wji (t + 1) = wji (t) + a(t) * hc1(t) * (xi (t) -wi (t)) for jehCi(t) w (t + 1) = w (t) (16)
.

decreasing function of time. Effective choices for these options and their parameters have so far been determined experimentally. For a chosen map dimension, different learning processes can be defined starting with different initial weight vectors, different topologies and different learning parameters. It is obvious that some optimal map for the same input data may exist and the best map is the one that yields the smallest average quantization error. The average quantization error is defined as the mean of ||x - Wc | over all training states where x is
the input training vector and

is the location of the winner neuron c and o(t) is monotonically decreasing function of time. The "bubble" function takes the form as follows. Let Nc (t) denote a neighborhood set of array points around node c at time t, then hcj (t) = 1 if j E Nc (t), otherwise hcj (t)= 0. NC (t) is also a monotonically

rc

to x.

wc is the nearest weight vector

for jX hi (t) Here a(t) is the "learning rate", which is a


monotnicaly dereasng fuctio of tme. Step 6 Increase the time tt1~~ Step 7 Go to step 2. After training, the weight vectors are self-organized and represent prototypes of the input vectors. Input vectors which are similar in their original space share similarity in

ji

j(

The program of SOM training is implemented in two phases. Phase I is the ordering phase during which the weight vectors of the map neurons are ordered. Phase II is the procedure whereby the weight vectors are fine-tuned. In phase I, a larger neighborhood radius and initial learning rate are used so that the weight vectors can be quickly ordered. In hase II, a smaller neighborhood radius and initial learning rate are used so that the weight vectors can be fine-tuned.

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

VI. MCS-SOM ALGORITHMS The objective of this paper is to show how Monte Carlo Simulation (MCS) can be made more efficient using SOM. We will first discuss the input training features for SOM. Then the approaches to marriage between SOM and MCS will be discussed. Input Training Features For the problem of loss-of-load state identification, a power system state can be characterized by load conditions, network topology and availability status of generators. Because the probability of outage of a transmission line is very small, we assume here that the lines are fully available all the time. The input vector corresponding to a system state is: Xi [P1,Qi, . ,Pin,Qin,PGii, ..., PGim] (18) Pk: Real power load of bus k for state i Qik :Reactive power load of bus k for state i PGim: Available real power generation of bus m for state i n : The number of load buses m: The number of generation buses
=

Marriage between MCS and SOM

It may be recollected from section II that the two drawbacks of MCS are the excessive time taken by state characterization and the sampling and characterization of similar states repeatedly. This section describes two methods to overcome these drawbacks.

In Method B, states sampled by MCS are clustered before determining their load loss characteristics. Monte Carlo simulation is used first to accumulate states, then SOM is used to cluster these states and OPF is used for analysis. Thus Method B overcomes the second disadvantage of the straight Monte Carlo simulation. The overall procedure of method B consists of the following steps. Step 1: Perform Monte Carlo sampling of the system state space to get N samples. These samples are taken as training vectors to SOM. N is decided by experience, for example, N=10,000. Step 2: Put the training vectors generated in step 1 into SelfOrganizing Map and train it. After training, the weight vector of each neuron represents a kind of power system state and maintains the original data's topological relationships. Also some neurons do not map any of the training vectors, and some neurons map one or more training vectors. Step 3: Perform OPF calculation for each neuron which has mapping of the training vectors. Determine the loadloss status of the neuron by using its weight vector as inputs to the OPF program. Label the neuron as "1" if it is loss-of-load and "0" if it is not loss-ofload. Step 4: Count the number of sampled states ni which are mapped to each loss-of-load neuron i in the SelfOrganizing Map. Step 5: Calculate the estimated LOLP value. LOLP

Method A Method A can more efficiently determine the load loss characteristic of the sampled state. In method A, the SOM is trained to recognize the loss-of-load states. Once this training is complete, the SOM is used along with the Monte Carlo simulation to estimate the reliability of the system. In this version of MCS-SOM the state evaluation is done by the trained SOM rather than OPF calculation. Thus this method overcomes the first disadvantage of the straight Monte Carlo simulation. The overall procedure of method A consists of the following steps. Step 1: Prepare the training patterns for SOM. Training patterns are obtained by OPF calculations which characterizes each training pattern as loss of load or otherwise. Step 2: Carry out SOM training with the prepared training patterns. patterns. or no-lossStep 3: Label a neuron in the map as loss-of-load ^ 1 - to the 1 majority * * label 1 ae 1 voting OfI of-load according
the

Yn Z
1
N

9)

VII. MSC-SOM FOR SECURITY ANALYSIS


The MCS-SOM approach is applied to security analysis in This reference considers both steady steady state as well as transient stability. So load curtailment, voltage stability and transient stability are considered. So far as the use of SOM is concerned, the approach of Ref 19 is slight modification of the approach described in section IV. Here again, the application of SOM -Monte Carlo simulation is divided into two steps. The first step is the implementation of SOM including SOM learning and OPF/Stability calculation. The next step is using SOM in the Monte Carlo simulation. A. The implementation of SOMfor security assessment

RefT21].

tainindpater ped,ttateurotn states match the weighting vectors of The isppedto trainedMonte 4: AftertheinSOM the SOM networks network iS trained, Monte Ccomponent Carlo Step Sp:the states corresponding to weighting vectors of neurons only are
neurons.

input sets are obtained by random sampling for Hee,input obtained selt rem by bysampln of contin obtained ency selection. System states samples of component states constitute the input sets. Vectors of

~ ~ ~ ~ Here,

sam prcedue a s in sameprocdure II.B exep tha stt clsiicto is by SO inta ofOF h ls (ls-fla or nols-o-od of eac state iS determined by the label Of the nearest
seto

simuatinfolow th simultionfollos

performed~ ~~ ~~~r

sampledr

neuron in thlmp

taken as states for the calculation of OPF/Stability calculations. OPF/Stability calculations include optimal power flow (OPF), transient stability analysis and voltage stability anal sis to determine system security. Details of these yyy considerations are given in the Appendix The OPF/Stability calculations are the most time-consuming step. These calculations are made on its states of mapped neurons. The

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

mapped neuron is labeled as secure or insecure after OPF/Stability calculations. B. Security assessment using by SOM simulation results The SGM has been trained in step A ,we can get the results of security of system by on-line use of SGM because the mapped neurons are labeled as secure or insecure. Note that OPF/Stability calculation is not needed anymore. When a system is insecure, the total of duration and frequency in emergency state is updated. If not, the result of system security of expected additional outages is checked by on-line
use SOM. of

A. Input selection The total load is fixed at the peak load of 2850MW. The input features for the SOM network consist of the generating lines itiue are assumed available at all unit statuses only (the nt tms.Teeae3 t1 ue,s h

VIII. STUDY RESULTS The IEEE Reliability Test System (RTS)[1 8] was chosen to test the proposed method. It has 24 buses (1O generation buses and 17 load buses), 38 lines and 32 units. The system annual peak load is 285OMW. The total installed generating capacity is 3405MW and the generation data are given in Table 1. The one line diagram of the system is shown in Fig. 2. Two sets of studies have been performed, one considering only the peak load and the other with multiple load level.

B. SOM training A total of 500 different training patterns were used to train the SOM network. These training patterns were selected in such a fashion that they were non-repetitive but most samples were from the high probability region of the state space. This was achieved by varying the availability status of units through a preliminary Monte Carlo experiment and evaluation by OPF. Table 2 shows the characteristics of the SOM training. TABLE II CHARACTERISTICS OF SOM (METHOD A, PEAK LOAD LEVEL) Input dimension Number of training patterns Kohonen layer (x*y)
2000

input vector is: [PGI, PG2 .

P10]

(20)

M5i

AtMAw 14
17

G:
am7

Q
MA 22. MS 23

21

Topology Neighborhood type Learning rate type Iteration number for phase I

rectangular bubble linear function


0.8

10 500 20*20

Initial neighborhood radius for phase I


Initial learning rate for phase I Iteration number for phase II

15
3

23

Bus I

1S

'S e
f2t WS

Initial neighborhood radius for phase I Initial learning rate for phase II 0.03

20000

I24 ^ WS

aw X38

After training, the map was calibrated and labeled according to the samples in the input data file. The best matching neuron in the map corresponding to each data vector is found. The neurons were then labeled as loss-of-load or noloss-of-load according to the majority of labels "hitting" a particular map neuron. The neurons that got no "hits" were left unlabeled. Using this procedure 143 neurons were labeled kVi as loss-of-load or no-loss-of-load.

C. Monte Carlo simulation After the map was labeled, Monte Carlo simulation was performed to estimate the loss of load probability (LOLP). For Fig. 2. One line diagram of IEEE RTS each sampled system state, SOM, instead of OPF, was used to TABLE I GENERATING UNIT DATA characterize it as loss-of-load or not. The label of the nearest neuron to each sampled system state was the estimate of loadUnit size (MW) Unit Type -Unit FOR loss status. Ten thousand states were sampled in the Oil 0.02 12(5) simulation. There were 8759 no-loss-of-load states and 1241 Oil 0.10 20(4) loss-of-load states characterized by SOM. Thus the estimated 0.01 Hydro 50(6) Coal 0.02 LOLP is 0.1241. 76(4) Oil 0.04 Monte Carlo simulation combined with OPF was 100(3) Coal 0.04 performed to obtain the benchmark value of LOLP at peak 155(4) load level for use in assessing the accuracy of SOM. For the Oil 0.05 197(3) 10000 system states sampled above, there were 8774 no-lossNuClear 0.12 (*) nuber fuits FOR- frce outge ateof-load states and 1226 loss-of-load states characterized by GPF. The computed benchmark value of LGLP is thus Method A 1) Reliability Analysis at Peak Load Level 0.1226. Among the 8774 no-loss-of-load state classified by GPF, 8720 states were classified as no-loss-of-load correctly by SOM in method A, resulting in a classification accuracy of 99.38%0. Among the 1226 loss-of-load states, 1187 states were

4030(2)

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

classified as loss-of-load correctly by SOM, giving a classification accuracy of 96.82%. D. Computing time It required 5 seconds for the phase I of SOM training (global ordering) and 31 seconds for the phase II of SOM training (fine-tuning). For the characterization of all the 10000 sampled states, the computing time was 3 seconds. Compared to the straight Monte Carlo simulation, which needs to perform 10000 OPFs, the computing time is greatly reduced. The program was implemented in C language and run on a Sun Solaris 2.5. 2) Reliability Analysis with Multiple Load Levels A. Input selection The load profile for IEEE RTS is represented by 15 load levels [17] as shown in Table 3. In this study system, loads are distributed at 17 buses and units are distributed at 10 buses, so the input features for the SOM network consist of load levels and generating unit statuses as follows. X = [PI, Q1, ... P17 , Q17 , PGI,..., PGIo ] (21)
TABLE III 15-STEP LOAD DATA

Initial neighborhood radius for phase I Initial learning rate for phase II

3 0.03

C. Monte Carlo simulation Monte Carlo simulation was performed after the map was labeled. SOM was used to characterize each sampled state as loss-of-load or not. Twenty thousand states were sampled in the simulation. There were 19870 no-loss-of-load states and 130 loss-of-load states as characterized by SOM. Thus the estimated LOLP is 0.0065. Monte Carlo simulation combined with OPF was performed to obtain the benchmark value of LOLP for multiple load levels. For the 20000 states sampled above, there were 19940 no-loss-of-load states and 60 loss-of-load states as characterized by OPF. Therefore the computed benchmark value of LOLP is 0.0030. Among the 19940 no-loss-of-load states classified by OPF, 19866 states were classified correctly by SOM in method A, resulting in a classification accuracy of 99.63%. Among the 60 loss-of-load states, 56 states were classified correctly by SOM, giving a classification accuracy of 93.33%.
D. Computing time

Load level(MW) Probability Load Level(MW) Probability


2850.0 0.08333

2365.5

0.0416
7
7

It required 5 seconds for the phase I and 36 seconds for the phase II of SOM training. For the classification of the total 20000 sampled states, the computing time was 14
seconds.

2821.5 2736.0

0.04167

2109.0

0.0416
7 0.0416

0.12500

2080.5

0.0416

Method B

2707.5 2679.0

0.16667

1909.5

1) Peak Load Level


E. SOM training

0.04167

1795.5

0.0833

2650.5
2593.5

0.04167
0.04167

1710.0
1681.5

0.0833

0.0833 3

2451.0

0.04167

B. SOM training Similar to the procedure in selecting training patterns for the peak-load-level case, a total of 1200 different training patterns were generated by varying the load levels and the availability status of generators. They were evaluated by OPF and used to train the SOM network. Table 4 shows the SOM training parameters. Again, the map was calibrated and labeled by the majority voting of training vectors. There were 217 neurons labeled as loss-of-load or not.
TABLE IV CHARACTERISTICS OF SOM (METHOD A, MULTIPLE LOAD

total of 10,000 training vectors generated from Monte Carlo sampling were used to train the SOM network. The training parameters of SOM are listed in Table 5.
TABLE V CHARACTERISTICS OF SOM (METHOD B, PEAK LOAD LEVEL)

The total load is fixed at the peak load of 2850MW. The input features for the SOM network are the same as (20). A

* .

Input dimension Number of training patterns Kohonen layer (x*y) Topology Neighborhood type type Neighborhood Learning rate type Iteration number for phase I Initial neighborhood radius for phase I

~~~~~bubble le

10 10000 30*30 rectangular 20


3000

LEVELS)

Number of training patterns

Input dimension

44

Initial learning rate for phase I 0.9 Iteration number for phase II 30000 Initial neighborhood radius for phase I 3 Initial learning rate for phase II 0.03

Kohonen layer (x*y) Topology Neighborhood typ

Learning rae type

~~

20*20 rectangular bubl liner function


2000

1200

F. LOLP calculation

Initial neighborhood radius for phase I15 Initial learning rate for phase I 0.8
Iteration number for phase II

Iteration number for phase I

20000

weight vectors as inputs to GPF. After the map was labeled, the total number of samples mapped to the loss-of-load neurons was counted. Among the 10000 samples, there were

of the training vectors. These neurons were hlabeled ~ ~ ad mapping as loss-of-load or no-loss-of-load by using their

After SGM training, there were 368 neurons in the map that

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

1187 samples mapped into loss-of-load neurons. Thus the estimated LOLP value is 0.1187. As shown before in straight Monte Carlo simulation, there were 8774 no-loss-of-load states and 1226 loss-of-load states of the total 10000 samples and the benchmark value of LOLP at peak load level is 0.1226. Among the 8774 no-loss-of-load states classified by OPF, 8763 states were classified correctly by SOM in method B, resulting in a classification accuracy of 99.87%. Among the 1226 loss-of-load states classified by OPF, 1176 states were classified correctly by SOM, giving a classification accuracy of 95.920.
G. Computing time The computation time required for method B consists of two major components. The first is the time required for training the SOM. The second is using OPF to label the neurons as loss-of-load or not. For the peak load condition, it required 6 seconds for the phase I of SOM training and 55 seconds for the phase II of SOM training. Also because there were 368 neurons that had mapping of the training vectors, 368 OPFs were performed to label these neurons after the SOM training. Compared to the straight Monte Carlo simulation that needs 10000 OPFs, the computing time is greatly reduced. Method B was also implemented in the C language and run on a Sun Solaris 2.5.

correctly by SOM, resulting in a classification accuracy of 99.96%. Among the 60 loss-of-load states classified by OPF, 40 states were classified correctly by SOM, giving a classification accuracy of 66.67%.
C. Computing time It required 25 seconds for the phase I of SOM training and 183 seconds for the phase II of SOM training. Also because there were 574 neurons that had mapping of the training vectors, 574 OPFs were performed to label these neurons after the SOM training. Compared to the straight Monte Carlo simulation that needs 20000 OPFs, the computing time is also greatly reduced.
IX. CONCLUSION A methodology MCS-SOM for power system reliability evaluation is proposed in this paper for adequacy and security evaluation. The objective of proposed methodology is to make Monte Carlo simulation more efficient by utilizing the characteristics of SOM. Cases studies are presented for adequacy evaluation. we can see that the training time for the SOM is trivial, while the number of OPFs needed is much smaller than the number required for straight Monte Carlo simulation. Similar results are obtained for security analysis as described in Ref 19.
X. APPENDIX

2) Multiple Load Levels A. SOM training The input features for the SOM network are the same as (21). A total of 20,000 training vectors generated from Monte Carlo sampling were used to train the SOM network. The training parameters of SOM are listed in Table 6.
TABLE VI CHARACTERISTICS OF SOM (METHOD B, MULTIPLE LOAD LEVELS)
Number of training patterns Kohonen layer (x*y)

A. Description ofProbabilistic Criteriafor Security analysis Transient stability The power system is considered stable if the fault is cleared before the critical clearing time (CCT), that is CCT is greater than fault clearing time (CT). The probability of successfully clearing a fault (Bt,) for contingency k can be calculated from probability density function of clearing time:

Input dimension

Topology Neighborhood type Learning rate type Iteration number for phase I Initial neighborhood radius for phase I Initial learning rate for phase I Iteration number for phase II Initial neighborhood radius for phase I
Initial learningrateforphasell

44 20000 40*40

B (k)
[s

P[CT < CCT]

(A - 1)

Rectangular

Bubble Linear function 3000 20 0.8 30000


0.03

The corresponding probability is compared with minimum acceptable probability of stability. The minimum acceptable probability of stability (Bmin) which depends on the required ty , degree of system stability. If Bt, is larger than Bt,, B 'ts is equal to one. If not, B't, is equal to zero. F B >mi B' (k) (A -2) {l when B,s < Bmi 0(

tse

B. LOLP calculation Steady state analysis After the SOM training, there were 574 neurons that had a) Load curtailment mapping of the training vectors. These neurons were labeled The formulation for load curtailment using OPF is the same as as loss-of-load or no-loss-of-load by using their weight in the main body of the paper for adequacy analysis. We can vectors as inputs to GPF. Among the 20000 samples, 48 hte odsedn xsso o o samples were mapped into loss-of-load neurons. Thus the epesa , estimated LGLP is 0.0024. contingency k When there iS no load shedding for As shown before in straight Monte Carlo simulation, there contingency k, B,c is equal to one. Otherwise, B,c is equal to were 19940 no-loss-of-load states and 60 loss-of-load states Of zero. the total 20000 samples and the benchmark value of LGLP for multiple load levels is 0.0030. Among the 19940 no-loss-ofload states classified by GPF, 19932 states were classified

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

1 BIC (k) = O
b)

when w when

Xi

=0 >0

(A

3)

[1]

A voltage indicator ( L ) is expressed [22] in the following equation to assess the voltage stability of load bus i.

b) Voltage stability
Voltage

stability

[2]
[3] [4]

Mar/Apr 1975, pp.508-517.


855.

Hutchinso, London, 1977. C. K. Pang, A. J. Wood, "Multi-Area Generation System Reliability IEEE Transactions on Power Systems, Vol. PAS-94, Calculations,"

Singh, C., Billinton, R., System Reliability Modeling and Evaluation,

XI. REFERENCES

F. N. Lee, "Multi-Area Reliability - A New Approach," IEEE Transactions on Power Systems, vol. PWRS-2, no.4, Nov 1987, pp. 848-

Li =
S i+ = S

Si+
Yi
E j i
+

V.
where

(A -4)
[5]
1

* S

C. Singh, A. D. Patton, A. Lago-Gonzales, A. R. Vojdani, G. Gross, F. F. Wu, N. J. Balu, "Operating Considerations in reliability Evaluation of Interconnected Systems - an Analytical Approach," IEEE Transactions on Power Systems, vol. PWRS-2, Feb. 1988, pp123-129. A. D. Patton, S. K. Sung, "A Transmission Network Model for Multiarea Reliability Studies," IEEE Transactions on Power Systems, vol.8, no.2, May 1993, pp. 459-465.

+ V

* V

[6]
[7]

= complex power of bus i) The voltage stability indicator of the overall system is maximum value among voltage stability indicators of the load
(S
Si

0. Bertoldi, L. Salvaderi, S. Scalcino, "Monte Carlo Approach in Planning Studies: An Application to IEEE RTS," IEEE Transactions on

Power Systems, vol.3, no.3, August 1988, pp. 1146-1154.

G. C. Oliveria, M V. F. Pereira, S. H. F. Cunha, "A Technique for


no.4, Nov 1989, pp. 1309-1315. Roy Billinton, Wenyuan Li, Reliability Assessment of Electric Power Systems Using Monte Carlo Methods, Plenum Press, New York, 1994. Luigi Salvaderi, Roy Billinton, "A Comparison Between Two Fundamentally Different Approaches to Composite System Reliability Evaluation," IEEE Transaction on Power Apparatus and Systems, vol.

Reducing Computational Effort in Monte Carlo Based Composite Reliability Evaluation," IEEE Transactions on Power Systems, vol.4,

buses.
Lk
=

[8]

max(Li)

(A - 5)
r

[9]

Similar to load curtailment analysis, we can express as BV


contingency k. When it is less than the threshold value (Lth) for contingency k, B, is equal to one. If not, BV is equal to
zero.

[10]

PAS-104, no. 12, Dec 1985, pp. 3486-3492. C. Singh, Z. Deng, "A New Algorithm
Evaluation--Simultaneous
Decomposition

[11] N. Gubbala, C. Singh, "A Fast and Efficient Method for Reliability

Simulation Approach," Electric Power System Research, vol. 21, 1991, pp. 129-136.

for Multi-area

Reliability

B,, (k) =

It
O

when

Lk k

when

<L Lk > Lth


th

(A - 6)

B. System operating state

The system operating state can be divided into normal, alert and emergency states [23]. Security constraints include all of transient state and steady state constraints. If the system violates any of the above conditions for contingency k, the system is in emergency state.. That is,

Evaluation of Interconnected Power Systems - Preferential Decomposition Method," IEEE Transactions on Power Systems, vol.9, no.2, May 1994, pp. 644-652. [12] M. Pereira, M. Macceira, etc, "Combining Analytical Models and Monte-Carlo Techniques in Probabilistic Power System Analysis," IEEE Transactions on Power System, vol.7, no.1, Feb. 1992, pp. 265-272. [13] C. Singh, J.Mitra' Composite System Reliability Evaluation Using State Space Pruning," IEEE Transactions on Power Systems, Vol. 12, No. 1,

pp. 471-479, Feb. 1997. [14] Teuvo Kohonen. Self-Organizing Maps. Springer-Verlag, Heidelberg,

1995. [15] X. Luo, C. Singh and A.D. Patton, "Loss-of-Load State Identification Using Self-Organizing Map," Proceedings of IEEE PES 1999 Summer Meeting, vol.2, 1999, pp. 670-675.
[16] Tharam S. Dillon, Dagmar Niebur, Neural Networks Applications in Power Systems, CRL Publishing, Ltd, London, 1996.

I normal or alert =k) normal (k){Bss alert1A 7) 0 emergency


-

[17] G. Hamound, "Probabilistic Assessment of Interconnection Assistance


no.2, May 1998, pp. 535-542.
Between Power Systems," IEEE Transaction on Power System, vol.13,

Bss B 'ts (k) * Blc (k) * B (k) Determination of the indices of normal and alert states based on system operating state is explained in the next section.
=

where

[18] IEEE Committee Report, "IEEE Reliability Test System," IEEE Transaction on Power Apparatus and Systems, vol. PAS-98, no.6, Nov/Dec 1979, pp. 2047-2054. [19] Nader Samaan, C. Singh, "Adequacy Assessment of Power System
Generation

Transactions on Power Systems, Vol 17, No 4, November 2002, pp 974-

Using

A Modified

Simple

Genetic

Algorithm",

IEEE

problem nor voltage stability problem exists in spite of a contingency, additional outage should be considered in order [22] . toditigus th te noma sate stt an in.... to a1t1ng1sn nomal na alr aert state sate as s described aesr1oel ln
[21]

C. Consideration of additional outages When there iS no contingency or when neither a system

[20] N.Samaan, C. Singh, "Genetic algorithms approach for the assessment of composite power system reliability considering multi-state [21] Hyungchul Kim, C.Singh "Probabilistic Security Analysis using SOM

981.

[23] Tomas E.Dy Liacco, "The adaptive reliability control system" IEEE transaction on power apparatus and systems Vol, PAS-86 No.5 May 1967.

System", IEEE Trans. On Power pp.346-354.

Jan27- 31, New York, 2002. P.Kessel and H.Glavitch, "Estimating the Voltage Stability of a Power

components",PMAPS ,2004. and Monte Carlo Simulation", Proceedings of Winter Power Meeting,

Delivery,

Vol

1,

No.

3,

Jul.

1986,

C Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden - June 11-15, 2006

C Copyright KTH 2006

Вам также может понравиться