You are on page 1of 7

Decision Tree-Based Fault Detection and Classification in Solar Photovoltaic Arrays

Ye Zhao, Ling Yang, Brad Lehman*


Department of Electrical and Computer Engineering Northeastern University Boston, MA, US zhao.ye@husky.neu.edu
AbstractBecause of the non-linear output characteristics of PV arrays, a variety of faults may be difficult to detect by conventional protection devices. To detect and classify these unnoticed faults, a fault detection and classification method has been proposed based on decision trees (DT). Readily available measurements in existing PV systems, such as PV array voltage, current, operating temperature and irradiance, are used as "attributes" in the training and test set. In experimental results, the trained DT models have shown high accuracy of fault detection and fault classification on the test set.

Jean-Franois de Palma, Jerry Mosesian, Robert Lyons


MERSEN USA Newburyport-MA, LLC Newburyport, MA, US

Time domain reflectometry (TDR) is proposed for detection and location of open-circuit faults and increased series resistance faults in PV strings [16]. At PV-string level, PV string monitoring has been proposed for real-time fault detection in [17]. However, none of the literature has used the decision-tree based model or similar data mining techniques in fault detection and classification. A decision-tree (DT) based model is an effective supervised technique to implement the classification methods in high-dimensional data [18]. It has been developed in power systems for fault detection, security assessment, and system control [19-21]. As for renewable system, such as solar photovoltaic (PV) systems, this paper proposes a DT based fault detection and classification method. Depending on the size of the model, the proposed DT model shows high accuracy on fault detection (up to 99.98%) and fault classification (up to 99.8%) on the test data. This paper develops a DT model at PV-array level and demonstrates its feasibility on PV experiments in real working conditions. Specifically, taking measurements, such as voltage and current of the PV array, along with weather conditions, the model can detect the fault and classify the specific fault type. In the proposed system, necessary data for the development of a DT model is created and recorded for the experimental PV system in both normal and fault conditions. The collected data consists of commonly available measurements in many PV systems, such as PV array voltage, current, operating temperature, and solar irradiance. Using WEKA software [22], the collected and pre-processed training set are randomly chosen from experimental data, and are used to build the DT model. After that, the trained DT model is tested on unseen real data for model evaluation. In summary, this paper presents the following research contributions: For the first time, a DT model is developed for fault detection and classification in solar PV arrays. The DT model has several advantages, such as fast steps

I.

INTRODUCTION

Fault detection in solar photovoltaic (PV) arrays is a fundamental task to increase reliability, efficiency and safety in PV systems. Without proper fault detection, uncleared faults in PV arrays not only causes power losses, but also might lead to safety issues and fire hazards [1]. Conventional fault detection and protection methods usually add overcurrent protection devices (OCPDs, such as fuses) in series with PV components [2-3]. However, it has been shown that certain faults in PV arrays may not be cleared by OCPDs, such as line-line faults, open-circuit faults, and partial shadings, due to the current-limiting nature, non-linear output characteristics of PV arrays, and even maximum power point tracker (MPPT) [4-9]. Furthermore, this paper shows that under different environmental conditions (irradiance and temperature), the faulted PV array could have the same operating voltage and current as the normal PV array. This would bring more difficulties to fault detection and classification in solar PV installations. Several fault detection and classification models for PV modules/arrays have been studied in the literature [8-17]. PV fault detection models based on long-term energy yield and power losses have been proposed in [8-12]. An extension diagnosis method based on the extended correlation function and the matter-element model is proposed to identify specific fault types of a PV system [13]. The study in [14] uses the discrepancy between simulated and real I-V curve of PV systems to detect and identify the faults. To prevent PV components from fire hazards, DC arc detection and protection methods for PV arrays has been studied in [15].
*The author gratefully acknowledges the support through grants by Mersen USA and the National Science Foundation (under grant 0901439).

978-1-4577-1216-6/12/$26.00 2012 IEEE

93

of training and classification, explicit interpretation, and easy implementation as a software/algorithm. The proposed DT model is able to identify faults in PV systems in real-time, with high prediction accuracy. Depending on the size of the DT model, the fault detection accuracy varies from 93.56% to 99.98%; the fault classification accuracy ranges from 85.43% to 99.8%. DECISION TREE-BASED MODEL IN SOLAR PV ARRAYS

A line-line fault is an accidental short-circuit connection between two points of different potential in PV arrays. Mismatch faults occur when the electrical parameters of module(s) is significantly changed from those of the remaining modules. Mismatch fault could be temporary, such as partial shading on PV modules. Also, it could be permanent, such as open circuit in PV modules/strings, degradation, or defective modules.

II.

A. Typical Grid-Connedted PV Systems A typical grid-connected PV system shown in Fig. 1 is the research target of this paper. It consists of several major components, including solar PV arrays, centralized inverter with MPPT algorithm, electrical connection wirings, and protection devices, such as overcurrent protection devices (OCPDs) and ground fault protection devices (GFPDs). Note that the PV system in the research is a grounded system, which has a system grounding point Gsys according to National Electric Code (NEC) in the US [23]. The PV array typically contains mn PV modules connected electrically in series and parallel configuration. This array configuration is, nowadays, most common in PV technologies [24]. There are n number of PV strings in parallel. Each PV string consists of m number of modules in series. B. Faults in Solar PV Arrays Typical faults in PV arrays consist of ground faults, lineline faults, and mismatch faults among PV modules [25-26]. Among these faults, line-line faults and mismatch faults are studied in this paper, since they are more difficult to detect by conventional protection devices than ground faults [27]. A ground fault is an accidental electrical short circuit involving ground and one or more normally designated current-carrying conductors.

It is necessary to mention that the PV array will work at a new system maximum power point (MPP) after the fault, as long as the array voltage can sustain the inverters operating [4-7]. For this reason, the data used for DT model is recorded only at the maximum power point (MPP) of the PV array. Furthermore, it is considered that the PV array is the only source of fault current, since most PV inverters contain transformers that could provide good galvanic isolation between PV arrays and utility grids [28]. C. The Process to Build a DT Model The DT model will be built according to four key steps in the process shown in Fig. 2. The first step is data acquisition, which obtains the training and test set from experiments. The second step is to pre-process the experimental data, including data cleaning, sampling, creating new attributes and attribute selection. The third step is to train the DT model by using 66% of randomly chosen pre-processed data. The last step is using remainder of pre-processed data to test the model.
Data acquisition

Data pre-processing

+
String 1 String 2 String n-1

Ipv
String n

GFPD

Centralized inverter
Utility grid

Training the model

(Open circuit)

Mismatch faults

Testing the model

Gsys

Figure 2. The process to develop a DT model

Ig
(Partial shading)

Vpv
PV module

Ipv
Isc-array I1

+ *

Normal PV array at STC Faulted PV array at STC MPPs of normal PV array MPPs of faulted PV array

Ground faults

Rf

Line-line fault

Series fuse for overcurrent protection

Area2 Area3

Area1 + + + + ** * * * + ++++ * * ** * * + ++++++++ + ** + ++ * * ** * * * +++ * * * * * * * * + + +++ + + * * * * * * * + ++ + +++ + + + * * ** * * * * * * * * + + +++ ++ ++++++ +

Vpv
Area4 V2 V1 Voc-F Voc-N

Figure 1. Typical faults in solar PV arrays

Figure 3. I-V curves of normal and faulted PV arrays

94

Internal node

Vpv < V1?


Y N
Normal
(Area1)

Root node

III.

DATA ACQUISITION AND PRE-PROCESSING OF EXPERIMENTAL RESULTS

A. Data Acquisition 1) Experimental setup: A small-scale grid-connected PV system has been set up to create and record faults under real working conditions. The schematic diagram and the photo are shown in Fig. 6 and Fig. 7, respectively. The parameters of PV components are summerized in Table I. Four types of faults mentioned previously have been created: 1) solid line-line fault (LL) between the middle of String 1 and negative bus bar; 2) the same solid line-line fault with fault impedance Rf=20 ohms (LL-20); 3) Open-circuit faults on String 2 (OPEN); 4) Partial shading faults on the PV modules (SHADE). When fault occurs in experiments, the PV array may have some transients and it could operate off its nominal MPP. But after a few seconds, the MPPT algorithm will make the PV array operate at a new MPP, which is called post-fault steady state. In other words, the faulted PV system may still work at its MPPs, under changing and various faults. 2) Data acquisition in experiments: Seven parameters (or called attributes in the DT model) in the PV array are recorded under both normal and fault conditions. They are time, ambient and PV modules operating temperature, array current (Ipv), reference current (Isc-ref), array voltage (Vpv) and reference voltage (Voc-ref). The experiments are carried on two clear days in Boston, MA. Under changing irradiance, 28 fault cases of aforementioned faults are recorded repetitively at 20~40HZ sampling frequency. Each fault case lasts for 5~15 minutes.
+
3)
... Ig=0
Isc-ref
A

Ipv < I1?


Y N N
Normal
(Area4) (Area2)

Vpv < V2?


Y
(Area3)

Fault

Fault

Leaf node

Figure 4. The DT model for fault detection

For example, in Fig. 3, faults (e.g. line-line faults) in PV arrays usually cause changed current vs. voltage (I-V) curves and reduced maximum power points (MPPs). Fault MPPs may vary from normal MPPs in the I-V curves. A DT is a flow-chart-like tree structure used to detect the faulted MPPs from normal ones (in Fig. 4). A DT consists of a root and internal nodes that are labeled with attribute test conditions. Starting from the root node, every instance (consisting of Vpv and Ipv at each MPP) will be split by the test at internal nodes and proceed down to the terminal nodes Normal or Fault (called class-label attributes, or leaf node in the DT model). For illustration purposes only, the simple model in Fig. 4 only uses two attributes (Vpv and Ipv) to detect the fault. However in the real world, normal and fault MPPs may overlap in I-V curves. Therefore, high-dimensional attributes (e.g. irradiance, temperature, and output power) should be considered. Once the DT model is built and tested, it can operate online for fault monitoring, as shown in Fig. 5. The DT model could be either programmed as if-then statements in a separate microcontroller or integrated with the PV inverter for real-time fault detection and classification.
Start Data acquisition and pre-processing Fault detection by DT model N Fault occurs? Y Fault classification by DT model Alarm

Ipv

GFPD
V

Grid-connected inverter

Utility grid

4)
F Rf
. . .

Vpv

Short-circuit current reference


V V oc-ref

1) and 2)

Open-circuit voltage reference

Figure 6. Schematic diagram of the experimental PV system

Figure 5. The flowchart of proposed fault detection and classification model

Figure 7. Photo of the experimental PV system

95

TABLE I. Equipment PV module Entire PV array Grid-connected inverter

EXPERIMENTAL PV COMPONENTS Parameters


Type Detailed parameters

TABLE II.
Attribute name

ORIGINAL AND NEW ATTRIBUTES


Unit Description

Power Film (amorphous silicon) 2 8 modules configuration Enphase microinverter M190

At STC: Voc=18V, Isc=0.9A, Vmpp=14V, Impp=0.75A, Pmpp=10.5W At STC: Psys=168 W, Vsys mpp=31 V, Isys mpp=6 A Max. output power 190W, min. start voltage: 28V; MPPT voltage range: 22 ~ 40V

1 2 3 4 5 6 7 8 9 10 11 12 13 14

time Temp_amb Temp_op Ipv Isc-ref Voc-ref Vpv Ppv Inorm Vnorm FillFactor Irradiance Detection Classification

seconds C C A A V V W 0~1 0~1 0~1 W/m


2

Time Ambient temperature Operating temperature Array current Short-circuit current reference Open-circuit voltage reference Array voltage Array power = Ipv* Vpv Ipv/(8* Isc-ref) Vpv/(2* Voc-ref) Vnorm* Inorm Isc-ref / Isc * 1000 NORMAL, FAULT NORMAL, LL, LL-20, OPEN, SHADE

Note that Isc-ref and Voc-ref are measured to incorporate the information of incident irradiance and temperature on the whole PV array, since the reference modules have the same location and identical electrical parameters as other modules. Furthermore, the transients during the fault are not included in the dataset. In other words, the data for the DT models are only collected at pre-fault MPPs and post-fault MPPs. B. Data Pre-processing 1) Data cleaning and sampling: To improve the accuracy and decrease the training cost of the DT model, outlier and missing values in the data are removed. Also, original experimental data are randomly sampled down to 764,529 instances, which is approximately 80% of its original size. The Ipv-Vpv plot of experimental results is given in Fig. 8. Notice that it is difficult to identify NORMAL, LL, LL-20 and SHADE merely by the two-dimension (I-V) dataset. Therefore, more attributes will be used to detect and classify the faults. 2) New attributes construction: Five new attributes are constructed to help the development of the DT model. The original and new attributes are summarized in Table II. The last two classes are called class-label attributes: Detection and Classification. They are created off-line for training and testing the DT model. The class Detection includes NORMAL and FAULT. The class Classification consists of NORMAL, LL, LL-20, OPEN, and SHADE.

Nominal Nominal

3) Attribute selection: To expedite the training process, it is necessary to remove the relatively less correlated and redundant attributes. Then, the attributes are selected only if they have low inter-correction but high correlation with the class-label attributes Detection and Classification. In WEKA software, information gain is used to evaluate the attributes. Information (entropy) needed to classify data Dj into class Ci, i=1,,m: InfoDj = m pi log 2 (pi ) i=1 (1)

where pi is the probability of class Ci in Dj.

At a non-leaf node, the information (entropy) needed after using attribute A to split D into v number of partitions D1,, Dv is defined as: InfoA (D) = v j=1
D j D

Information gain by splitting at attribute A:

Five attributes with high information gain ranking are chosen to develop the DT model: Inorm, FillFactor, Vpv, Voc-ref, and Ppv. It is interesting to see that irradiance and temperature attributes are ruled out, since they have been implicitly included in the selected attributes already. Notice that the new-attribute 3D plot of Inorm, Vpv and FillFactor (in Fig. 9) gives better fault classification than original attributes in Fig. 8. However, to further classify the faults, high-dimensional attributes are necessary.

GainA (D) = Info(D) InfoA (D)

Info (Dj )

(2)

(3)

Figure 8. Vpv vs. Ipv under various conditions

96

TABLE III. NORMAL 467,494 61.15%

ENTIRE DATASET FOR THE DT MODEL Class-label attribute LL LL-20 OPEN 65,647 8.59% 28,719 3.76% 115,791 15.15% SHADE 86,878 11.36%

# of instances Percent

The developed DT models can be easily programmed in a microcontroller for real-time fault monitoring. For example, if a new set of measurements are: Voc_ref=17V, Inorm=0.6, FillFactor=0.32, Ppv=60W and Vpv=25V, then according to the DT model, the PV array will be classified as a SHADE fault. By increasing the size, the DT model will become more complex and more accurate (up to 99.8%). The test accuracy of different sizes of DTs have been summarized in Table IV and Table V, where the number of leaf nodes for the DT model is defined as the external nodes that will denote a class prediction; and the size of the tree is defined as all nodes included in the tree structure. At leaf-nodes, the first and the second number in the parentheses represent the correct and error predictions. C. Discussion To train and test the DT model, 764,529 instances are collected and pre-processed from 28 various fault cases under changing irradiance. Both high accuracy at detection and classification of specific pre-defined faults are achieved. Depending on the size of the model, the proposed DT model shows high accuracy on fault detection (up to 99.98%) and fault classification (up to 99.8%) on the test data.
Voc-ref <= 17.645 | Ppv <= 99.384: FAULT (226119.0/7911.0) | Ppv > 99.384: NORMAL (20007.0/7932.0) Voc-ref > 17.645 | Inorm <= 0.57744 | | Voc-ref <= 17.969: FAULT (61090.0/7835.0) | | Voc-ref > 17.969: NORMAL (20512.0/2479.0) | Inorm > 0.57744 | | Vpv <= 25.07: FAULT (21312.0/8359.0) | | Vpv > 25.07: NORMAL (415489.0/2208.0)

Figure 9. 3D plot of Inorm, Vpv and FillFactor under various conditions

IV.

TRAINING AND TESTING THE DECISION-TREE MODEL

A. Training the Model WEKA is used to train the DT model according to the training set that we have collected and selected previously [22]. The algorithm for DT induction is shown in Fig. 10, in which the Create_node() function grows the DT by adding a new node; the Classify() function determines the class label to a leaf node; the Attribute_selection_method() function finds the best splitting criterion based on information gain; the Stop_condition() function terminates the growth of DT if stopping conditions are met. Applying the training set in the DT algorithm, the DT models of fault detection and classification are developed. Among 764,529 instances in all dataset, 66% of them are randomly chosen as the training set. The remainder is used for test set. Note that the performance baseline of model accuracy (majority vote is NORMAL) for fault detection and fault classification is approximately 61.15% (see Table III). B. Testing the Model At first, a simple and small-size DT model has been trained and tested. The DT model for fault detection and fault classification are shown in Fig. 11 and Fig. 12. The detection accuracy on test data is 93.56% and classification accuracy is 85.43%.
Generate_decision_tree(D, A) if Stop_condition(D, A) = true then leaf = Create_node(); leaf.label = Classify(D) return leaf; else root = Create_node() root.split = Attribute_selection_method(D, A) let Dj (j=1,,v) be the possible outcome of root.split for each Dj, do child = Generate_decision_tree(Dj, A) attach child to root; end for end if return root Figure 10. The DT induction algorithm

Figure 11. A simple fault detection model


Inorm <= 0.57629 | FillFactor <= 0.31584: LL (65653.0/9.0) | FillFactor > 0.31584 | | Ppv <= 56.926: NORMAL (20006.0/1020.0) | | Ppv > 56.926 | | | Voc-ref <= 17.818 | | | | Ppv <= 77.507 | | | | | Vpv <= 24.872: SHADE (20002.0/4430.0) | | | | | Vpv > 24.872: OPEN (62443.0/17166.0) | | | | Ppv > 77.507: OPEN (78487.0/13272.0) | | | Voc-ref > 17.818: NORMAL (20230.0/9816.0) Inorm > 0.57629 | Vpv <= 25.314 | | Voc-ref <= 17.739 | | | Voc-ref <= 17.091: SHADE (23107.0/8169.0) | | | Voc-ref > 17.091: LL-20 (27024.0/9805.0) | | Voc-ref > 17.739: NORMAL (20117.0/3495.0) | Vpv > 25.314: NORMAL (427460.0/8627.0)

Figure 12. A simple fault classification model

97

Even for the simple DT models, the fault detection and classification accuracy can reach 93.56% and 85.43%, respectively. The classification accuracy for each class-label attribute is: NORMAL 99.48%, LL 99.99%, LL-20 0%, OPEN 89.78% and SHADE 21.3%. Notice that LL-20 is the most difficult fault to classify. At the most complex model, the fault classification accuracy can achieve 99.8%. In this case, the classification accuracy for LL-20 can reach 99.7%. However, the size of the DT model will be 1637 nodes with 819 leaf nodes, which might be too large for practical application. Therefore, a proper size DT model should be chosen by considering the tradeoff between accuracy and model complexity. Limitations may exist in the DT model for solar PV arrays: 1) Training cost may be high. The experimental data are generated in the real world. This could draw peoples concern of cost and safety; 2) The DT model may not perform correctly at unknown data remarkably different from the training set. Since the operation of PV arrays greatly depends on environmental conditions, and a huge number of various faults conditions may occur in the PV array, it seems difficult to obtain the sufficient training and test dataset that can cover all possible fault scenarios.
TABLE IV.
Size of the tree

V.

CONCLUSION AND FUTURE WORK

For the first time, a new fault detection and classification in solar photovoltaic (PV) systems based on decision-tree (DT) models has been proposed. The process of training the model is straightforward and easy to implement. The trained model shows good detection accuracy (up to 99.98%) and classification accuracy (up to 99.8%) that may have promising application in the real world. The future research plans to address some of the previously discussed limitations, which include model optimization, cost reduction of fault data acquisition, and integration with PV inverters. REFERENCES
[1] [2] B. Brooks. (2011) The Bakersfield Fire - A Lesson in GroundFault Protection. SolarPro. Page 62 - 70. W. Bower and J. Wiles, "Investigation of ground-fault protection devices for photovoltaic power system applications," in Photovoltaic Specialists Conference, 2000. Conference Record of the Twenty-Eighth IEEE, 2000, pp. 1378-1383. J. Wiles. (2008) To Fuse or Not to Fuse? Home Power #125, code corner. 106-108. Y. Zhao, B. Lehman, J.-F. de Palma, J. Mosesian, and R. Lyons, "Fault evolution in photovoltaic array during night-to-day transition," in Control and Modeling for Power Electronics (COMPEL), 2010 IEEE 12th Workshop on, 2010, pp. 1-6. Y. Zhao, B. Lehman, J.-F. d. Palma, J. Mosesian, and R. Lyons, "Fault Analysis in Solar PV Arrays under: Low Irradiance Conditions and Reverse Connections," in the 37th IEEE Photovoltaic Specialist Conference, Seattle, WA, 2011. Y. Zhao, B. Lehman, J.-F. d. Palma, J. Mosesian, and R. Lyons, "Challenges of Overcurrent Protection Devices in Photovoltaic Arrays Brought by Maximum Power Point Tracker," in the 37th IEEE Photovoltaic Specialist Conference, Seattle, WA, 2011. Y. Zhao, B. Lehman, J.-F. d. Palma, J. Mosesian, and R. Lyons, "Challenges to Overcurrent Protection Devices under Line-line Faults in Solar Photovoltaic Arrays," in the third IEEE Energy Conversion Congress & Exposition (ECCE), Phoenix, AZ, USA, 2011. A. Chouder and S. Silvestre, "Automatic supervision and fault detection of PV systems based on power losses analysis," Energy Conversion and Management, vol. 51, pp. 1929 - 1937, 2010. A. Drews, A. C. d. Keizer, H. G. Beyer, E. Lorenz, J. Betcke, W. G. J. H. M. v. Sark, W. Heydenreich, E. Wiemken, S. Stettler, P. Toggweiler, S. Bofinger, M. Schneider, G. Heilscher, and D. Heinemann, "Monitoring and remote failure detection of gridconnected PV systems based on satellite observations," Solar Energy - Elsevier, pp. 548 - 564, 2007. H. Haeberlin and C. Beutler, "Normalized Representation of Energy and Power for Analysis of Performance and on-line Error Detection in PV-Systems," presented at the 13th EU PV Conference on Photovoltaic Solar Energy Conversion, Nice, France, 1995. S. K. Firth, K. J. Lomas, and S. J. Rees, "A simple model of PV system performance and its use in fault detection," ELSEVIER SOLAR ENERGY, 2009. S. Vergura, G. Acciani, V. Amoruso, G. E. Patrono, and F. Vacca, "Descriptive and Inferential Statistics for Supervising and Monitoring the Operation of PV Plants," Industrial Electronics, IEEE Transactions on, vol. 56, pp. 4456-4464, 2009. K.-H. Chao, S.-H. Ho, and M.-H. Wang, "Modeling and fault diagnosis of a photovoltaic system," Electric Power Systems Research, 2008. D. Stellbogen, "Use of PV circuit simulation for fault detection in PV array fields," in Photovoltaic Specialists Conference, 1993., Conference Record of the Twenty Third IEEE, 1993, pp. 13021307.

[3] [4]

RESULTS OF FAULT DETECTION


Training time (seconds) Detection Accuracy (%)

# of leaves

[5]

11 17 21 47 55 69 81 85 105 115 139 159

6 9 11 24 28 35 41 43 53 58 70 80

35.46 48.03 50.62 54.82 54.64 56.3 65.43 62.2 65.83 66.24 66.32 75.83

93.5639 95.8202 98.8476 99.6838 99.8369 99.9111 99.9161 99.9515 99.9608 99.9731 99.9835 99.9892

[6]

[7]

[8] [9]

TABLE V.
Size of the tree

RESULTS OF FAULT CLASSIFICATION


# of leaves Training time (seconds) Classification Accuracy (%)

19 31 59 113 133 235 311 431 667 919 1247 1637

10 16 30 57 67 118 156 216 334 460 624 819

34.94 48.61 54.48 64.14 69.91 81.33 87.6 93.74 116.37 128.9 142.79 161.68

85.4343 91.0749 93.1827 97.8033 98.5643 98.9909 99.3083 99.4991 99.6218 99.7207 99.7761 99.8

[10]

[11] [12]

[13] [14]

98

[15]

[16]

[17] [18] [19] [20] [21]

H. Haeberlin, "Arc Detector for Remote Detection of Dangerous Arcs on the DC Side of PV Plants," presented at the 22nd European Photovoltaic Solar Energy Conference, Milano, Italy, 2007. T. Takashima, J. Yamaguchi, and M. Ishida, "Fault detection by signal response in PV module strings," in Photovoltaic Specialists Conference, 2008. PVSC '08. 33rd IEEE, 2008, pp. 15. N. Takehara and K. Fukae, "Abnormality detection method, abnormality detection apparatus, and solar cell power generating system using the same," 5669987, 1997. J. Han and M. Kamber, Data Mining Concepts and Techniques, Second ed., 2006. S. Yong and S. M. Rovnyak, "Decision tree-based methodology for high impedance fault detection," Power Delivery, IEEE Transactions on, vol. 19, pp. 533-536, 2004. S. Rovnyak, S. Kretsinger, J. Thorp, and D. Brown, "Decision trees for real-time transient stability prediction," Power Systems, IEEE Transactions on, vol. 9, pp. 1417-1426, 1994. H. Mori, "State-of-the-Art Overview on Data Mining in Power Systems," in Power Systems Conference and Exposition, 2006. PSCE '06. 2006 IEEE PES, 2006, pp. 33-34.

[22] [23] [24]

[25] [26] [27] [28]

I. H. Witten and E. Frank, Data Mining: Practical Machine Learning Tools and Techniques: Morgan Kaufmann, 2000. NFPA70, "Article 690 - Solar Photovoltaic Systems of National Electrical Code," ed, 2011. J. Wiles, "Photovoltaic Power Systems And the 2005 National Electrical Code: Suggested Practices," Southwest Technology Development Institute, New Mexico State University November 26, 2008. S. E. Forman, "Performance of Experimental Terrestrial Photovoltaic Modules," Reliability, IEEE Transactions on, vol. R-31, pp. 235-245, 1982. E. L. Meyer and E. E. van Dyk, "Assessing the reliability and degradation of photovoltaic module performance parameters," Reliability, IEEE Transactions on, vol. 53, pp. 83-92, 2004. Y. Zhao, "Master of Science Thesis: Fault Analysis in Solar Photovoltaic Arrays," Electrical and Computer Engineering, Northeastern University, Boston, 2010. L. Quan and P. Wolfs, "A Review of the Single Phase Photovoltaic Module Integrated Converter Topologies With Three Different DC Link Configurations," Power Electronics, IEEE Transactions on, vol. 23, pp. 1320-1333, 2008.

99

You might also like