2016 Joint International Conference on Artificial Intelligence and Computer Engineering (AICE 2016) and International Conference on Network and Communication Security (NCS 2016)
ISBN: 978-1-60595-362-5
Fault Diagnostic Method for Micro-grid Based on Wavelet SOM Neural
Network and Multi Agent System
Qiu LU
1,a, Ye-Yin ZHONG
2,b,*1Logistics Engineering College, Shanghai Maritime University, Shanghai China
2Electrical and Electronic Engineering College, Shanghai Institute of Technology, Shanghai China
a[email protected], b[email protected]
*Corresponding author
Keywords: Wavelet Singular Entropy, SOM Neural Network, Micro Grid, Topology Structure, Fault
Diagnosis.
Abstract. Fault diagnosis using traditional neural network method requires training samples to train the neural network. The micro grid system has the characteristics of flexible operation mode and variety of topology structure, the fault diagnosis method of it by using neural network has the problem of poor adaptability and requires a large number of training samples. A fault diagnosis method combining multi agent system with wavelet som neural network is proposed. The wavelet som neural network can judge the reason of fault and the multi agent system judge the location of the fault. A micro grid simulation system is established based on PSCAD. The simulation results prove the feasibility of the fault diagnosis method based on multi agent system and wavelet som neural network. the wavelet som neural network only need typical fault training samples, this method will not affected by fault location, fault time and other factors, it has good adaptability to the change of the topology structure of micro grid system.
Introduction
Micro grid is a small distribution power system which distributed generation, energy storage devices, load and other equipment are integrated together. It is an effective method for the distributed power access to the power grid. It has the characteristics of flexible operation mode and various topological structure. At present, the research of the micro grid system is mainly focused on the control and energy management [1-4]. Research on the fault diagnosis of Micro Grid system is less.
Zhang Han propose a method by using the improved ant colony algorithm and the fuzzy Petri net in the fault diagnosis of wind power generation simulation system of micro grid [5]. Wang Wenrui bulid up a analytical model for the fault diagnosis of micro grid though the diagnosis rules [6]. Li Mengnan design different functions of Agent and MAS architecture for the micro grid‘s topology identification [7]. Though the action information of the circuit breaker and protection, a fault diagnosis model based on petri net and advanced petri net is proposed respectively [8-9]. The fault diagnosis of micro grid multi pulse rectifier thyristor circuit is studied by paper [10]. Fault diagnosis methods of micro grid power infrastructure are summarized by paper [11].
This paper propose a fault diagnosis method of micro grid system which combine wavelet som neural network with multi agent system, the simulation results prove the feasibility of it. This method will not affected by fault location, fault time and other factors, it has good adaptability to the change of the topology structure of Micro Grid.
Wavelet Singular Entropy
stationary signals or the singularity of the signals. It’s often used in the fault conditions which has transient signal [12-13].
The matrix feature can get by the eigenvalue decomposi but the matrix must be a square matrix. In reality, most of the matrix is not a square matrix, the singular value decomposition is a method that can adequate for any matrix. If A is a M´N matrix, U is a M´M square matrix, åis a M´N matrix, T
V is a N´N matrix, then the eigenvalue decomposi of any M´N matrix A can be expressed as:
T V U
A (1)
The coefficient matrix get by the signal wavelet transform can be decomposed into a matrix DM´N in M band. The matrix DM´N can reflect the basic feature of the original coefficient matrix. At this time there must be matrixs UM´l,Rl l´ and Vl N´ .
N T M N
M U R V
D 1 11 1 (2)
The main diagonal elements (r ii =1, 2, , ) l in the Diagonal matrixRl l´ , is the singular value of
the matrix DM´N , if the matrix DM´N reflect the time frequency information of transient signal, then the matrix ri can show the basic modal characteristics of the matrix DM´N . In order to describe the frequency components and distribution characteristics of signal, the wavelet singular entropy is defined as follows:
I
i i SE P W
1 (3)
I
i i k I
i i k
i r r r r
P
1 1
log (4)
The wavelet singular entropy can reflect the uncertainty of the analyzed signal’s energy distribution of characteristic mode and will change with the change of the fault. When fault happens, the uncertainty of the fault phase is larger than that of the non fault phase, so the wavelet singular entropy of the fault phase is larger than that of the non fault phase, it is feasible to use the wavelet singular entropy for fault diagnosis.
SOM Neural Network
Figure 1. Som neural network model.
The som neural network can show the results of classification in the competitive layer by self-organizing learning without the help of teacher. Through repeated learning of input mode, the spatial distribution density of connection weights and the probability distribution of input mode are beginning to converge.
The Fault Diagnosis Method Based on Som Neural Network and Multi Agent System
A. Multi Agent System
A Multi Agent System is composed of many agent units which are distributed in different Locations, these
agent units can be implemented by different software and hardware platforms. Each agent unit can solve the limited problem independently. By interacting with other agent it can solve complex problems [14‐17].
B.The fault diagnosis model of Multi Agent System
This paper design a fault diagnosis model based on multi agent system for the topology structure identification of dynamic network system. The model includes three parts: Monitor Agent, ID Agent and Aco as shown in Figure 2:
ij N N
A = a ´ æ ö ç ÷ ç ÷ è ø
Figure 2. Multi agent system fault diagnosis model.
Monitor Agent:
Monitor the number of nodes in the environment, record the number of current running nodes Max(i) i= 1, 2， ， N, then activates the ID Agents with the same number of current running nodes, each activated ID Agent corresponds to a running node. When the number of nodes in the environment changes, according to the value of Max i（ ） corresponding changes in the number of ID Agents.
[image:3.612.230.379.446.610.2]Though identifying all the neighbors agent of each ID agent, transfer the neighbor agent matrix to Aco.
Aco:
It can get the final topology identification matrix, it is the topological structure of the current system, it can be used to determine whether and where the topology changes.
C.Neighbor Agent search algorithm
Assumption 1: The loss of the numerical package and the interference is not consider. Assumption 2: Each ID Agent has at least one neighbor Agent
Assumption 3: The value of the element in neighbor agent matrix can only be 0 or 1, 0 is not a neighbor Agent, 1 is a neighbor Agent, the element value of the ID Agent self is 0.
Step1: According to the number of running nodes N which are monitored by Monitor Agent, the corresponding ID Agents are established, then number them. Each ID agent has a neighbor agent matrix with random initial value.
Step2: All ID Agent according to the serial number in ascending order
Step3: The ID agent with the smallest number as the starting ID Agent, In turn to send a value of 1 package to the remaining Agent ID address, this package can only be received by ID Agent once.
Step4: If the target ID Agent ID receive a value of 0, then return to starting Agent ID 1, the starting ID Agent will represent the target ID Agent in the neighbor matrix elements set 1, else set 0.
Step5: Repeat step4 until all values in the starting ID Agent neighbor matrix are determined, then transfer neighbor matrix to Aco
Step6: Repeat steps3 to step5 until all ID Agent transfer their neighbor Agent matrix to Aco. Step7: Finish
D. Micro grid system fault diagnosis method based on som neural network and multi agent system Step1: The three phase voltage of static switch, three-phase current of public bus, zero-sequence current of normal operation and fault of micro grid system in different operation modes are collected and stored.
Step2: The collected signal is decomposed by wavelet transform, then the wavelet singular entropy of each signal is calculated. Sum the single signal wavelet singular entropy in the same running state then save it.
Step3: The som neural network is trained by the saved data of Step 2.
Step4: According to the set time interval, the real-time acquisition of microgrid system’s three phase voltage of static switch, three-phase current of public bus, zero-sequence current, the data after wavelet processing is an input feature vector of the trained som neural network. Judge the reason of fault according to the output state of som neural network at this time.
Step5: The topological structure of micro grid system is monitored by multi agent system constantly
Figure 3. Fault diagnosis structure based on mas and wavelet som.
Fault Diagnosis of Micro Grid System
Table 1. Initial typical sample of som neural network. stat
e
a
i ib ic io ua ub uc
1
s 4.64 4.83 5.01 0.58 6.46 6.40 6.43
2
s 4.63 4.82 4.98 0.56 6.44 6.38 6.41
3
s 4.75 4.91 5.03 0.59 6.53 6.47 6.51
4
s 3.22 3.25 3.46 1.46 4.03 4.11 4.10
5
s 3.34 3.38 3.54 1.58 4.13 4.26 4.20
6
s 3.19 3.24 3.42 1.46 3.95 3.97 3.96
7
s 4.75 3.05 3.08 7.41 7.71 4.40 4.46
8
s 4.71 3.03 3.05 6.98 7.35 4.24 4.25
9
s 4.88 3.16 3.20 7.52 7.80 4.37 4.36
10
s 6.72 7.10 3.98 6.11 7.88 7.81 2.89
11
s 6.58 6.99 3.75 6.01 7.85 7.79 2.87
12
s 6.94 7.37 4.32 6.23 8.12 8.02 3.12
1
s
~s
12: The states of micro grid system that running in different time, different running mode, has only one change in topology structure and the location of fault is not the same.1
s
~s
3: Normal operation, working load change from 60% to full.4
s
~s
6: External fault, full working load.7
s
~s
9: A phase to ground short circuit, full working load, different fault location, 3 distributed generation devices out of operation separately.10
s
~s
12: AB phase to ground short circuit, full working load, different fault location, 3 distributed generation devices out of operation separately.For the convenience of data processing, the data in Table 1 is normalized and input into som neural network as training data, adjust weights repeatedly. Because the input feature vector is 7, the input layer of the som neural network has 7 neurons, the output layer has 49
(
7 7´)
neurons.Although the samples are initially classified, when the number of training steps is 50, but each sample cannot be classified separately, so the classification is not accurate enough. When the number of training steps is 300, each sample can be classified separately. It’s no sense to increase the number of training steps now. The training results as shown in Fig5-Fig6.
- 1 0 1 2 3 4 5 6 7 - 1
0 1 2 3 4 5 6
SOM Topol ogy
Figure 5.(1) Initial output mapping.
- 1 0 1 2 3 4 5 6 7
- 1 0 1 2 3 4 5 6
1 0 0 1 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 1
0 0 0 0 1 0 0 1 0 0 0 0 0 0
0 0 1 0 0 0 1 1 0 0 0 0 1 1
[image:7.612.343.527.556.708.2] [image:7.612.107.280.559.713.2]0. 135 0. 14 0. 145 0. 15 0. 155 0. 1
0. 11 0. 12
0. 13 0. 14
0. 15 0. 16
0. 17 0. 12
0. 14 0. 16 0. 18
W( i , 1) Wei ght Vect or s
W( i , 2)
W(i
,3
[image:8.612.218.394.72.227.2])
Figure 6. Weight vector graph after training.
The output of som neural network is: Y= [42 49 48 21 12 7 1 16 3 30 43 38]
[image:8.612.167.445.354.434.2]Table 2 is the samples that fault mode and fault location is known. These samples are normalized to be used as the input feature vector of som neural network, to verify the feasibility of this fault diagnosis method.
Table 2. The known fault samples. stat
e
a
i ib ic io ua ub uc
Case1 s6 4.37 4.39 4.54 2.13 4.84 4.93 4.89
Case2 s9 4.91 3.38 3.41 7.64 7.90 4.41 4.43
Case3 s11 7.30 7.62 4.84 6.74 8.54 8.45 3.42
Case1: The fault location is Line 12 Case1: The fault location is Line 11 Case1: The fault location is Line 6 The output of som neural network is: Y= [7 3 43]
[image:8.612.117.285.555.647.2]In this paper, the multi agent system identifies topology structure by monitoring circuit breaker status. When three kinds of faults in Table 2 occurred, the change of micro grid’s topology structure as shown in Fig. 7-Fig. 10.
[image:8.612.359.530.558.650.2]Figure 9. Topology of case 2. Figure 10. Topology of case 3.
The results consistent with the fault reason and fault location. Analyze another 100 datas of different locations at different times:
25 cases with the normal operation of micro grid and the working load changes, 2 cases of misjudgment;
25 cases with external fault, full working load, 1 cases of misjudgment;
25 cases with A phase to ground short circuit, full working load, different fault location, 3 distributed generation devices out of operation separately, 3 cases of misjudgment;
25 cases with AB phase to ground short circuit, full working load, different fault location, 3 distributed generation devices out of operation separately, 2 cases of misjudgment;
The rate of correct diagnosis is 92%.
Conclusions
For the fault diagnosis of micro grid system with flexible operation mode and various topology structures, a fault diagnosis method combining multi agent system with wavelet som neural network is proposed. The simulation results show that within one sampling period, when the topological structure of the micro grid system has only one change, this method has a good adaptability and will not be affect by the fault location or fault time. It can reduce the training samples of wavelet som neural network and judge the fault location and fault reason of micro grid system accurately. The situation that more than one change with topology structure in one sampling period or interference problem in topology identification are to be further investigatigated.
Acknowledgement
This research was financially supported by the National Science Foundation(61374132); Shanghai Maritime University graduate student innovation foundation(2014ycx057).
References
[1] Hatziargyriou N., Asand H., Itavani, et al, Microgrids, J. IEEE Power and Energy Magazine. 5 (2007) 78-94.
[2] Su Ling, Zhang Jianhua, Wang Li, et al, Study on some problems and technique related to microgrid, J. Power System Protection and Control. 38 (2010) 235-238.
[3] Li Y.W., Vilathgamuwa D.M., Poh C.L., Microgrid power quality enhancement using a three-phase four-wire grid-interfacing compensator industry applications, J. IEEE Transactions on Industry Applications. 41 (2005) 1707-1719.
[4] Wang Chengshan, Xiao Zhaoxia, Wang Shouxiang, Synthetical control and analysis of microgrid, J. Automation of Eleetric Power System. 32 (2008) 98-103.
[image:9.612.358.528.72.143.2][6] Wang Wenrui, Research on fault diagnostic method for Micro-grid based on analytic model, D. Shanghai institute of Technology. (2015) 9-19.
[7] Li Mengnan, An investigation on topology identification of microgrid based on mas, D. Shanghai Institute of Technology. (2015) 19-31.
[8] Wu Hongxia, Yang Guoming, Zhang Ailing,et al, Fault diagnosis of micro-grid based on petri net, C. Proceedings of the 2012 International Conference on Electrical and Electronics Engineering. (2012) 475-482.
[9] Shi Ji Ying, Sun Wen Shan, Li Chun Ling, Fault diagnosis of micro-grid based on advanced Petri net, C. 2011 3rd International Conference on Mechanical and Electronics Engineering. (2011) 3015-3018.
[10] Han Jingyong, Liu Hongda, Fault Detection Method in Micro-grid Multi-pulse Thyristor Rectifier Circuit, C. Power and Energy Engineering Conference. (2012)
[11] Hare, James, Shi Xiaofang, et al, A review of faults and fault diagnosis in micro-grids electrical energy infrastructure, C. IEEE Energy Conversion Congress and Exposition. (2014) 3325-3332. [12] He Zhengyou, Liu Zhigang, Qian qingquan, Study on wavelet entropy theory and adaptability of its application in power system, J. Power System Technology. 28 (2004) 38-43.
[13] He Zhengyou, Cai Yumei, Qian qingquan, A study of wavelet entropy theory and its application in electric power system fault detection, J. Power System Technology. 28 (2004) 17-21.
[14] Choudhury R., Paul K., Bandyopadhyay S., A multi-agent routing protocol for mobile wireless Ad hoc networks, J. Autonomous Agents and Multi-agent Systems. 8 (2004) 47-68.
[15] Hao Liang, Bong Jun Choi, Weihua Zhuang, et al, Multi agent Coordination in Microgrids via Wireless Networks, J. IEEE Wireless Communications. 6 (2012) 14-22.
