IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO.

8, AUGUST 2019 1731

Data-Driven Detection of Hot Spots in

Photovoltaic Energy Systems
Hongtian Chen , Member, IEEE, Hui Yi , Bin Jiang , Senior Member, IEEE, Kai Zhang , and Zhiwen Chen

Abstract—Hot spots are common abnormalities in photovoltaic accelerate the cell degradation, and even cause fire of PV
(PV) energy systems. Their presence can potentially cause dam- energy systems [8].
age to PV modules, such as performance degradation or even Although bypass diodes are usually used for mitigating
unexpected fire to PV energy systems. By sufficiently mining the
information hidden in the test data collected from PV modules, these unfavorable situations [4], hot spots and concomi-
this paper develops a space-to-space projection method, which at tant negative influences still exist and have serious impact
its core is a linear approach via preserving the locally geometri- on series-connected PV cells as well as the whole energy
cal structure with respect to time series. Based on the nonlinear systems [9]. In fact, hot spots are regarded as one of the
model of PV modules established via the proposed projection, main defects of the PV modules, successful detection of which
data-driven detection of hot spots in PV energy systems can be
directly achieved with three key advantages: 1) its implementa- can improve safety greatly [10]. Specifically, when a large
tion does not depend on any mathematical model or physical amount of PV modules are used in new energy systems, real-
knowledge of PV energy systems; 2) it is of high-computational time, accurate, and automatic detection of hot spots is a real
efficiency especially in the online detection phase; and 3) it can challenge in both theoretical and practical domains [11], [12].
capture the dynamic characteristic because the local structure of Nowadays, two general methods are usually adopted to
samplings regarding time is given sufficient consideration. The
effectiveness and feasibility of the proposed approach are first detect hot spots in PV modules [9], [12]–[14]: 1) image anal-
presented by theoretical analysis and, then, convictively demon- ysis [15] and 2) the use of electrical characterizations [8]. The
strated via 15 sets of hot spot experiments on practical PV implementation of the former depends on analyzing the images
modules. acquired by thermal cameras or videos [16] with the aid of,
Index Terms—Data-driven detection, hot spots, photovoltaic for example, unmanned aerial vehicles [17]. It is obvious that
(PV) modules. this method is contactless and not invasive. However, it has
its inherent shortcomings, such as being immature, expen-
sive, and time consuming [16]. These weaknesses prevent the
I. I NTRODUCTION implementation of real-time detection of hot spots if large
ITH the rapid development of new energy systems, amounts of PV modules are installed in the grid-connected
W the safety and reliability issues are receiving grow-
ing concerns [1]–[7]. In the grid-connected photovoltaic (PV)
energy systems.
Based on electrical characterizations, the latter used for
energy systems, hot spots are the well-known abnormali- detection of hot spots proves more flexible than the former in
ties unexpected, which will reduce the power performance, real-time cases [14]. As described in [9], high-frequency and
low-frequency measurements are the key information used for
Manuscript received October 14, 2018; accepted January 18, 2019. Date detecting hot spots. This is because the capacitance installed
of publication February 26, 2019; date of current version July 17, 2019. will be changed when hot spots appear in PV modules. More
This work was supported in part by the National Natural Science Foundation
of China under Grant 61490703, Grant 61573180, and Grant 61503181, in specifically, abnormal changes in an impedance caused by hot
part by the Jiangsu Innovation Program for Graduate Education under Grant spots can be monitored based on low-frequency measurements
KYLX16− 0378, and in part by the Priority Academic Program Development in the dc impedance region and high-frequency measurements
of Jiangsu Higher Education Institutions. This paper was recommended by
Associate Editor D. Yue. (Corresponding author: Bin Jiang.) in the capacitive region. It is of high-computational efficiency
H. Chen and B. Jiang are with the Division of Automation Engineering, at the cost of additional circuitry to generate the low-frequency
College of Automation Engineering, Nanjing University of Aeronautics signals [18].
and Astronautics, Nanjing 211106, China, and also with the Jiangsu
Key Laboratory of Internet of Things and Control Technologies, Nanjing Recently, several data-driven approaches were also intro-
University of Aeronautics and Astronautics, Nanjing 211106, China (e-mail: duced to detect hot spots in PV modules via analyzing
[email protected]; [email protected]). electrical characterizations [19]. For example, in [20], based
H. Yi is with the College of Electrical Engineering and Control Science,
Nanjing Tech University, Nanjing 211800, China (e-mail: [email protected]). on the profile of I–V curves for PV modules, the emerging
K. Zhang is with the School of Automation and Electrical Engineering, kernel-based extreme learning machine is used for identify-
University of Science and Technology Beijing, Beijing 100083, China (e-mail: ing hot spots caused by abnormal shapes with the aid of its
[email protected]).
Z. Chen is with the School of Information Science and classification ability. By this way, direct detection of hot spots
Engineering, Central South University, Changsha 410083, China (e-mail: can be achieved without any additional sensors or circuitries.
[email protected]). However, it is a fact that the accuracy of such intelligent
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. detection methods depend on the completeness of training
Digital Object Identifier 10.1109/TSMC.2019.2896922 samples [15], [21], [22], i.e., I–V curves of all fault types are
2168-2216 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1732 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 8, AUGUST 2019

Fig. 2. Hot spots in PV modules.

Fig. 1. Schematic of PV energy systems.

required to construct an accurate classifier. Because of the dif-

ficult acquisition of these complete samplings under various
abnormal situations, practical detection of hot spots will be
less reliable.
Motivated by the observations above, this paper puts its
emphasis on proposing a new data-driven scheme to accu-
rately, in real time, detect hot spots in PV modules. The key
contributions together with advantages of this solution are (a) (b)
summarized as follows. Fig. 3. Mechanism difference for PV modules under normal and hot spot
1) By considering the nonlinear relationship among past, situations. (a) Normal PV module. (b) PV module with hot spots.
temporal, and future samplings, the modeling of com-
plex PV modules can be achieved via the proposed
space-to-space projection which is a linear realization. electricity [23]. In general, a PV system consists of PV arrays
2) A trace operator-based test statistic is specially designed connected in parallel, and a PV array is connected by PV mod-
to evaluate the operational condition of PV modules, ules in series. For a PV module, its basic component is the
which is suitable for assessing the similarity between PV cell. As presented by the schematic in Fig. 1, PV array
two matrices linking to the space-to-space projection. collects the power generated by all PV cells and sends it to
3) The proposed scheme is of high-computational effi- the inverter for further utilization.
ciency on the one hand, and can be directly achieved
without any mathematical model or physical knowledge B. Hot Spots and Their Influences
of PV energy systems on the other hand.
The remainder of this paper is organized as follows. Hot spots should never be overlooked in PV energy systems
Section II describes the PV modules, the influence caused because they cause a dramatic temperature increase as well as
by hot spots, and the following objectives of this paper. a consequent reduction of the provided power [24]. As shown
Section III focuses on the proposed space-to-space projec- in Fig. 2, due to the existence of hot spots, some cells in the
tion, and based on which the method used for detection of hot PV array yield significantly higher temperatures than others.
spots is then formulated. Through multiple hot spot experi- More seriously, local overheating will increase the risk of fire
ments, Section IV illustrates the effectiveness of the designed if no remedial actions are taken.
method and gives some discussions. Along with the conclud- In fact, the essence of the aforementioned abnormality is the
ing remarks, Section V summarizes this paper and finally partial or full shading on PV cells which lead to a mismatch
presents the further investigation. in the radiation [4]. Fig. 3 describes the different mechanisms
for PV modules working under normal and hot spot situations,
where IPV is the module current, UPV is the module voltage,
II. BACKGROUND AND P ROBLEM F ORMULATIONS
IA and IB are the respective currents in string A and B. For
In this section, PV energy systems are first described, fol- the normal PV modules without shading, one has that IA =
lowed by the negative influences of hot spots. Based on IB < IPV . If there exists shading on PV cells, such as the case
the characteristics of collected signals from PV modules, the presented in Fig. 3(b), IA < IB and IA < IPV .
objective of this paper is then formulated. Based on the above analysis, the difference in the oper-
ational mechanisms between normal and hot spot situations
A. Photovoltaic Cells and PV Energy Systems will arise. Taking Fig. 3(b) as an example, the affected cells
Regarded as the most fundamental part of PV energy marked in black in string A operate at reverse-bias and thus
systems, a PV energy system converts sunlight into dissipate power instead of delivering power. This phenomenon
CHEN et al.: DATA-DRIVEN DETECTION OF HOT SPOTS IN PV ENERGY SYSTEMS 1733

will lead to that abnormally high temperatures only located at where PCA is a special case of  given in (2), and X
locality. is the normalization of X. It can be observed from the
objective function described in (3) that, PCA try to find a low-
C. Design Issues dimension representation where the global variability, such as
the multivariable correlation structure [26] of X collected from
Despite fruitful research on detection of abnormalities, PV modules is taken into consideration. More details can be
direct applications are limited because of the special character- referred to in [27]–[29].
istics of signals measured from PV modules. More specifically, However, the performance of PCA used for feature extrac-
main challenges behind detection of hot spots include the tion and detection of hot spots may be greatly influenced by
following. the high dynamic of PV modules. Besides, the local structure
1) Establishing the mathematical model of PV modules of the collected data set X is not well suited for the description
could not be done without painstaking efforts, or even of PV modules whose processes consist of nonlinear relations
its accurate model may be not obtained due to internal among past, temporal, and future observations. In order to cap-
and external uncertainties. ture the local neighborhood structure on the high-dimension
2) There is no sufficient sample size in a test process to sup- X, NPE will be introduced as the basic knowledge of this
port data-based modeling, especially to describe highly application.
nonlinear dynamics of PV modules. NPE is one of the most popular manifold learning tech-
Therefore, to achieve effective detection of hot spots in PV niques [30]. Different from the locally linear embedding
modules, the following objectives should be achieved in this developed in [31], it can find a set of linear projections
paper. where the local neighborhood structure of high-dimensional
1) Propose a new projection method specially used for manifold can be redescribed via a low-dimensional submani-
modeling the dynamic and nonlinear PV modules by fold in Euclidean space. A brief summary about algorithmic
the use of small-size samples. procedures of NPE is presented as follows.
2) Design an evaluation function, based on the model estab- 1) Constructing an adjacency graph where either K-nearest
lished, which is sensitive to abnormalities caused by hot neighbors or  neighborhood can be used to select
spots in PV modules. neighbors and hence to construct an adjacency graph.
3) Evaluate the feasibility by extensive experiments. 2) Computing the weights where for the ith sampling, a set
of weights on its edges will be obtained via minimizing
III. P ROPOSED M ETHODOLOGY the embedding cost function.
In this section, we first describe PV modules by the use 3) Computing embedding projections where the embedding
of the signal-based form and briefly review neighborhood projections can be formulated by solving the generalized
preserving embedding (NPE) technique. Then, a new space- eigenvector problem.
to-space method used for the detection of hot spots in PV
modules is then discussed in detail. B. Data-Driven Modeling via Nonlinear
Feature Extraction
A. Signal-Based Description of PV Modules 1) Local Information Regarding Time: As explained in
Section II, X collected from PV modules is highly dynamic,
Define N observations from m sensors equipped in PV mod-
which results in obvious serial correlations among past, tem-
ules be X = [x1 , . . . , xi , . . . , xN ]T , where X ∈ RN×m . In
poral, and future samplings. For variable i, we first define a
accordance with the experimental platform of PV modules in +
:= {i − K, . . . , i − 1, i + 1, . . . , i + K}, where
set such that Ni,K
Section IV, xi can be stacked and described by
i and K are positive integers satisfying that i > K. Then, the
xi = [s1 (i) s2 (i) · · · s8 (i)]T (1) locality preserving criterion which adequately considers the
relation among past, temporal, and future samplings can be
where s1 to s8 are defined in Table I in Section IV-A. formulated as follows:
⎛ ⎞
To extract the nonlinear features hidden in X, the essence N 
2K
behind which is finding a projection  as follows: j = min ⎝xi − Wij xj ⎠ (4)
i j=1
Y = (X) (2)
where the weight matrix Wij takes the following form:
where Y is a compact representation of X. Without loss of  2K +
generality, we can define that Y = [y1 , . . . , yN ]T ∈ RN×l with Wij = j=1 Wij = 1 for j ∈ Ni,K (5)
l ≤ m. 0 otherwise.
Generally, the choice of  depends on the objective function Assume that there exists a linear projection P, (2) can be
related to why  is used. For example, principal compo- rewritten as
nent analysis-based techniques are used for the dimensional
reduction via retaining the maximal variance of X [25] Y = XP. (6)

T If each yi can be reconstructed with the same local structure

JPCA = max TPCA X XPCA (3) as xi , based on its neighbors along the time sequence and
1734 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 8, AUGUST 2019

according to the similar rule given in (5), (4) changes into of samplings regarding time series. As shown in (12), the
⎛ ⎞ space-to-space projection from Xi,s to Yi,s is described in an
N 2K
explicit form, by which some decisive impacts on the detec-
j = min ⎝yi − Wij yj ⎠
tion performance of hot spots in PV modules are emphasized
i j=1
in the following remark.
= min P X IN×N − W T (IN×N − W)XP
T T
(7) Remark 2: Compared with (6), the projection given in (12)
with the constrain that has two inherent advantages in detecting hot spots: 1) it can
partially weaken the negative influence caused by random
PT X T XP = Il×l (8) uncertainties in PV modules and 2) it is more credible in
where W is the weight matrix on the basis of Wij , whose evaluating the reconstruction error from the space viewpoint.
formulation can be found in [31]. By introducing a Lagrange
multiplier λ, the optimization problem given in (7) and (8) is C. Data-Driven Detection of Hot Spots
equal to the following formulation: Define that the data set Xf collected from PV modules with
P X IN×N − W
T T T
(IN×N − W)XP hot spots is
− λ P X XP − Il×l = 0.
T T
(9) Xf = X + F (15)
It is obvious that the solution of j , i.e., the expected projection where F is the abnormal matrix caused by hot spots and F ∈
matrix P, is reduced to solve the generalized eigen-problem RN×m . Similar to (12), the new low space can be obtained via
as follows:
f f
Yi,s = Xi,x P. (16)
X T IN×N − W T (IN×N − W)XP = λX T XP. (10)
Recalling (14), one has
Specifically, to minimize the reconstruction error in the
obtained model Y, l eigenvectors corresponding to the first f −1
Yi,s = Yi,s + Fi,s P PT P
l smallest eigenvalues are chosen to formulate P such that f −1 T
P ∈ Rm×l . =⇒ Ei,s = Ei,s + Fi,s − Fi,s P PT P P
Remark 1: Note that although data-driven modeling is =
f
Ei,s + Ei,s (17)
achieved via the linear operator given in (6), the projection
where Ei,s takes the form Ei,s = Fi,s − Fi,s P(PT P)−1 PT . Here,
matrix P sufficiently considers locality information hidden in f f
observations from PV modules. Therefore, the new compact f
Ei,s is a reconstruction-error matrix mainly affected by hot
model Y that incorporates the serial correlation is helpful for spots. It means that the detection of hot spots in PV modules
describing nonlinear dynamics of PV modules. can be transformed into the problem of how to monitor the
2) Space-to-Space Projection: Because of the strict align- f
changes reflected in Ei,s .
ment of samplings in test processes on PV modules, it makes
Remark 3: With the aid of locality structure preserved, Ei,s
the space-to-space projection according to time series possible.
is independent of changes in operating points. It is said that,
Introducing the notations
⎡ T ⎤ ⎡ T ⎤ different from the multivariate analysis techniques [32]–[34],
xi−s+1 yi−s+1 such as the principal component analysis, the proposed scheme
⎢ xT ⎥ ⎢ yT ⎥ is able to deal with the signals with various mean information.
⎢ i−s+2 ⎥ ⎢ i−s+2 ⎥
Xi,s = ⎢ . ⎥, Yi,s = ⎢ . ⎥ (11) Therefore, the proposed method shows more effectiveness than
⎣ .. ⎦ ⎣ .. ⎦
the multivariate analysis methods for the PV modules whose
xiT yTi signals are of obvious nonlinearity.
leads to the following space-to-space projection: Therefore, we can define an evaluation function used for
the detection of abnormalities caused by hot spots as follows:
Yi,s = Xi,s P (12)
where Xi,s ∈ Rs×m , Yi,s ∈ Rs×l , and P ∈ Rm×l , and s is an J(i) = tr Ei,s
T
Ei,s (18)
integer. based on which, its threshold Jth can be chosen as
Consider that
Jth = max J(i) (19)
Xi,s = Yi,s PT + Ei,s (13)
where tr(·) means the trace operator, and Jth provides a reliable
it yields
bound of uncertainties in PV modules. If there is a hot spot
−1
Yi,s = Xi,s P PT P appearing in PV modules, (18) becomes
−1 T   T 
=⇒ Ei,s = Xi,s − Yi,s PT P P (14) f f
J(i) = tr Ei,s
T
Ei,s + Ei,s Ei,s
where Ei,s is the reconstruction-error matrix which is only   
f T f
related to the system uncertainties. = tr Ei,s Ei,s + tr Ei,s Ei,s
T
The underlying idea of space-to-space projection described
in (12) using (5) lies in the sufficient use of alignment =⇒ J f (i) > Jth . (20)
CHEN et al.: DATA-DRIVEN DETECTION OF HOT SPOTS IN PV ENERGY SYSTEMS 1735

Algorithm 1 Off-Line Modeling

1: Read the normal data set X.
2: Set K and then form the weight matrix Wij according
to (5).
3: Calculate the projection matrix P via solving (10).
4: Obtain the matrix Ei,s based on (14).
5: Determine the threshold Jth by (18) and (19).

Algorithm 2 Online Detection of Hot Spots

1: Read the real-time sample xon and form it as Xon,s (a) (b) (c)
according to (11).
Fig. 4. Hot spots in PV modules from first to third cases. (a) Case 1.
2: Calculate the reconstruction-error matrix Eon,s for Xon,s .
(b) Case 2. (c) Case 3.
3: Obtain the online J via (18).
4: Make a decision whether there exists hot spots or not
according to (21), and go back to Step 1.

TABLE I
M AIN PARAMETERS OF P HYSICAL PV M ODULES

(a) (b) (c)

Fig. 5. Hot spots in PV modules from fourth to sixth cases. (a) Case 4.
(b) Case 5. (c) Case 6.

Remark 4: Different from the common evaluation func-

tions adopted in [27] and [35]–[37], the trace operator used
in (18) provides an effective solution to find a boundary in
(a) (b) (c)
high-dimensional space with respect to the uncertainties.
Therefore, based on the above discussions, for a new sam- Fig. 6. Hot spots in PV modules from seventh to ninth cases. (a) Case 7.
pling xnew , it can be determined whether there are hot spots (b) Case 8. (c) Case 9.
in PV modules or not as follows:
 A. Experimental Platforms
Hot spots exist if J(i) > Jth
(21) In the energy system, a physical PV module developed
Normal if J(i) ≤ Jth .
by National Energy Solar Center in China Electric Power
Research Institute is chosen as the test platform. All exper-
D. Implementation Procedures of the Proposed Scheme
iments are conducted under the same conditions (for example,
Based on the above presentation, the implementation of the irradiance is 1000 W/m2 , temperature is 25 ◦ C, and air mass is
proposed methodology can be divided into two parts, i.e., the AM1.5). In addition, eight sensors are installed in the test plat-
off line and online parts. Detailed procedures in both parts are form, the basic information of them is listed in Table I, where
summarized in Algorithms 1 and 2, respectively. MPPT is the acronym of maximum power point tracking.
In this application, hot spots are considered via shading
IV. E XPERIMENTS PV cells using paperboards, electric wire, and colored tapes.
In this section, 15 sets of hot spots are artificially made As shown in Figs. 4–8, 15 different portions of PV cells are
on a practical PV module. Based on this platform, data sets shaded to generate different levels of hot spots.
collected from both normal and hot spot cases are used to illus-
trate the efficiency of the method we design. On the basis of B. Detection Results of Hot Spots
theoretical analysis and experimental results, some concluding In the proposed method used for data-driven model, K = 2
remarks are then drawn in Section V. and s = 10. Therefore, the time neighboring 2K is 4 and
1736 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 8, AUGUST 2019

(a) (b) (c)

Fig. 7. Hot spots in PV modules from 10th to 12th cases. (a) Case 10.
(b) Case 11. (c) Case 12.

Fig. 10. Detection results for cases 4–6.

(a) (b) (c)

Fig. 8. Hot spots in PV modules from 13th to 15th cases. (a) Case 13.
(b) Case 14. (c) Case 15.

Fig. 11. Detection results for cases 7–9.

Fig. 9. Detection results for cases 1–3.

the dimension s regarding space projection is 10. In addition,

the threshold used for detection of hot spots is set to Jth =
3.5 × 10−6 .
Fig. 12. Detection results for cases 10–12.
For 15 kinds of hot spots in PV modules, detection results
obtained with the proposed space-to-space projection method
are shown in Figs. 9–13, where black solid lines correspond to
the normal case, blue solid lines are the test statistics regarding proposed method for three cases under consideration. As
hot spot cases, and red dashed lines are thresholds used for the presented in Fig. 5, case 4 exhibits much greater shades than
detection decision. It should be noted that for emphasizing the cases 5 and 6. According to the consistency between the eval-
detection performance, the first 100 samples in Figs. 9–13 are uation function J in Fig. 10 and the shaded areas in Fig. 5,
collected under normal conditions, and the last 100 samples the hot spot in case 4 exhibits much greater temperature
are related to different hot spots. differences than those in cases 5 and 6.
Fig. 9 shows the detection results for cases 1–3 using the For cases 7–9 shaded using one, two, and three paperboards,
proposed method. It is obvious that the test statistic J is able to detection results are, respectively, given in Fig. 11. It is obvi-
detect three types of hot spots with high degree of satisfaction. ous that, shaded areas are positively related to abnormally
In cases 4–6 of hot spots, detailed detection results are local temperatures. As the shaded areas increase, the detection
depicted in Fig. 10, and also show the effectiveness of our performance will be improved.
CHEN et al.: DATA-DRIVEN DETECTION OF HOT SPOTS IN PV ENERGY SYSTEMS 1737

V. C ONCLUSION
Aiming at the safety issue of new energy systems, this
paper presents a novel space-to-space projection for detec-
tion of hot spots in PV modules. The designed scheme, with
the ability to model nonlinear discharging processes of PV
modules accurately, overcomes the drawbacks of traditional
detection methods called I–V curves. Superior performances
of the proposed detection method is evaluated by 15 sets of
practical experiments on PV modules. Observed from these
experimental results, the proposed detection scheme proves
effective in the detection of all hot spots with a high degree
of satisfaction. This paper puts its focus on the preliminary
data-driven detection of hot spots in PV modules, and further
Fig. 13. Detection results for cases 13–15. work will be mainly targeted at diagnosis and propagation of
Fig. 12 presents the detection results regarding cases 10–12. hot spots to provide instructions on the maintenance of PV
It should especially be pointed out that, despite tiny shades energy systems.
caused by case 12, the proposed method is still effective.
As shown in Fig. 12, the evaluation function J exceeds the
R EFERENCES
threshold in the sampling/time internal from 101 to 125.
Fig. 13 reports the detection results for cases 13–15 per- [1] C. Dou, D. Yue, J. M. Guerrero, X. Xie, and S. Hu, “Multiagent
system-based distributed coordinated control for radial DC microgrid
formed to validate the proposed hot-spot detection scheme, considering transmission time delays,” IEEE Trans. Smart Grid, vol. 8,
where the hot spot in case 13 is blocked by a tiny and translu- no. 5, pp. 2370–2381, Sep. 2017.
cent tape. These detection results also indicate that the same [2] T. Strasser et al., “A review of architectures and concepts for intelligence
trends are expected for increasing shaded areas and that the in future electric energy systems,” IEEE Trans. Ind. Electron., vol. 62,
no. 4, pp. 2424–2438, Apr. 2015.
proposed space-to-space method works well for all hot spots [3] D. Ding, Z. Wang, Q.-L. Han, and G. Wei, “Security control for discrete-
appearing in PV modules. time stochastic nonlinear systems subject to deception attacks,” IEEE
Trans. Syst., Man, Cybern., Syst., vol. 48, no. 5, pp. 779–789, May 2018.
[4] S. Daliento, F. D. Napoli, P. Guerriero, and V. D’Alessandro, “A
C. Discussions modified bypass circuit for improved hot spot reliability of solar pan-
Following the above experimental results together with the els subject to partial shading,” Solar Energy, vol. 134, pp. 211–218,
Sep. 2016.
theoretical analysis in Section III, some concluding remarks on [5] M. Rahmani-Andebili and H. Shen, “Price-controlled energy man-
detection of hot spots using the proposed method are formu- agement of smart homes for maximizing profit of a GENCO,”
lated as follows, with an aim to provide informative guidance IEEE Trans. Syst., Man, Cybern., Syst., to be published.
doi: 10.1109/TSMC.2017.2690622.
in practical applications. [6] X. Xie, D. Yue, and S. Hu, “Fault estimation observer design of discrete-
1) There exists a positive correlation explicitly between time nonlinear systems via a joint real-time scheduling law,” IEEE Trans.
hot spots caused by the shaded areas and the evalua- Syst., Man, Cybern., Syst., vol. 47, no. 7, pp. 1451–1463, Jul. 2017.
[7] C. Dou, D. Yue, X. Li, and Y. Xue, “MAS-based management and
tion function in this paper. Therefore, prompt inspections control strategies for integrated hybrid energy system,” IEEE Trans. Ind.
(and immediate remedies) should be given for PV mod- Informat., vol. 12, no. 4, pp. 1332–1349, Aug. 2016.
ules if the higher value of the evaluation function [8] M. Dhimish, V. Holmes, B. Mehrdadi, M. Dales, and P. Mather, “Output-
power enhancement for hot spotted polycrystalline photovoltaic solar
appears. cells,” IEEE Trans. Device Mater. Rel., vol. 18, no. 1, pp. 37–45,
2) Observed from the detection results, our proposed Mar. 2018.
method is more effective in the early discharging phase [9] K. A. Kim, G.-S. Seo, B.-H. Cho, and P. T. Krein, “Photovoltaic hot-spot
than in the subsequent phase. The main reason may be detection for solar panel substrings using AC parameter characteriza-
tion,” IEEE Trans. Power Electron., vol. 31, no. 2, pp. 1121–1130,
the increasing voltages of capacitive loads of PV mod- Feb. 2016.
ules in the whole test process. Alarms appearing in the [10] M. Dhimish, V. Holmes, P. Mather, and M. Sibley, “Novel hot spot
subsequent phase usually correspond to large-area hot mitigation technique to enhance photovoltaic solar panels output power
performance,” Solar Energy Mater. Solar Cells, vol. 179, no. 1,
spots and, thus, more attention should be given. pp. 72–79, Jun. 2018.
3) A promising solution for the improvement of detectabil- [11] C. Dou, D. Yue, and J. M. Guerrero, “Multiagent system-based event-
ity in the subsequent phase is, because of the change of triggered hybrid controls for high-security hybrid energy generation
systems,” IEEE Trans. Ind. Informat., vol. 13, no. 2, pp. 584–594,
operation mechanisms when hot spots appear, to focus Apr. 2017.
on discovering the difference in the correlation among [12] S. R. Madeti and S. N. Singh, “A comprehensive study on different types
all measured signals. of faults and detection techniques for solar photovoltaic system,” Solar
Energy, vol. 158, pp. 161–185, Dec. 2017.
4) Obvious characteristics (such as increasing values in
[13] J. A. Tsanakasa, L. Ha, and C. Buerhop, “Faults and infrared thermo-
three sets of statistics in Fig. 13) are revealed in graphic diagnosis in operating c-Si photovoltaic modules: A review of
the detection results. These features displayed in the research and future challenges,” Renew. Sustain. Energy Rev., vol. 62,
detection results are of great importance if further inves- pp. 695–709, Sep. 2016.
[14] M. Bressan, A. Gutierrez, L. G. Gutierrez, and C. Alonso, “Development
tigations (such as estimation of the shaded areas in PV of a real-time hot-spot prevention using an emulator of partially shaded
modules) are needed. PV systems,” Renew. Energy, vol. 127, pp. 334–343, Nov. 2018.
1738 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 8, AUGUST 2019

[15] P. Henriquez, J. B. Alonso, M. A. Ferrer, and C. M. Travieso, “Review Hongtian Chen (M’18) received the B.S. and
of automatic fault diagnosis systems using audio and vibration signals,” M.S. degrees in electrical engineering from the
IEEE Trans. Syst., Man, Cybern., Syst., vol. 44, no. 5, pp. 642–652, School of Electrical and Automation Engineering,
May 2014. Nanjing Normal University, Nanjing, China, in 2012
[16] A. Arenella, A. Greco, A. Saggese, and M. Vento, “Real time fault and 2015, respectively. He is currently pursuing
detection in photovoltaic cells by cameras on drones,” in Proc. Int. Conf. the Ph.D. degree in control science and engineer-
Image Anal. Recognit., Montreal, QC, Canada, Jun. 2017, pp. 617–625. ing with the College of Automation Engineering,
[17] M. Dalsass et al., “Correlation between the generated string powers of Nanjing University of Aeronautics and Astronautics,
a photovoltaic: Power plant and module defects detected by aerial ther- Nanjing.
mography,” in Proc. IEEE 43rd Photovolt. Specialists Conf., Portland, He has been a Visiting Scholar with the Institute
OR, USA, Jun. 2016, pp. 3113–3118. for Automatic Control and Complex Systems,
[18] S. Weng, D. Yue, C. Dou, J. Shi, and C. Huang, “Distributed University of Duisburg-Essen, Duisburg, Germany. His current research
event-triggered cooperative control for frequency and voltage stability interests include incipient fault detection and diagnosis under multivariate
and power sharing in isolated inverter-based microgrid,” IEEE Trans. statistical frame, machine learning-based incipient fault detection and diagno-
sis, and their applications to the electrical traction system of high-speed trains
Cybern., to be published. doi: 10.1109/TCYB.2018.2803754.
and new energy systems.
[19] K.-H. Chao, C.-T. Chen, M.-H. Wang, and C.-F. Wu, “A novel fault
diagnosis method based-on modified neural networks for photovoltaic Hui Yi received the B.E. and Ph.D. degrees in
systems,” in Proc. Int. Conf. Swarm Intell., Beijing, China, Jun. 2010, control science and engineering from the College
pp. 531–539. of Automation Engineering, Nanjing University of
[20] Z. Chen et al., “Intelligent fault diagnosis of photovoltaic arrays based Aeronautics and Astronautics, Nanjing, China, in
on optimized kernel extreme learning machine and I-V characteristics,” 2005 and 2012, respectively.
Appl. Energy, vol. 204, pp. 912–931, Oct. 2017. In 2012, he joined the College of Electrical
[21] F. Sun, G.-B. Huang, Q. M. J. Wu, S. Song, and D. C. Wunsch, II, Engineering and Control Science, Nanjing Tech
“Efficient and rapid machine learning algorithms for big data and University, Nanjing, as a Lecturer, where he is cur-
dynamic varying systems,” IEEE Trans. Syst., Man, Cybern., Syst., rently an Associate Professor. His current research
vol. 47, no. 10, pp. 2625–2626, Oct. 2017. interests include fault diagnosis, health monitoring,
and artificial intelligence.
[22] C. L. P. Chen and Z. Liu, “Broad learning system: An effective and
efficient incremental learning system without the need for deep architec- Bin Jiang (M’03–SM’05) received the Ph.D. degree
ture,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 1, pp. 10–24, in automatic control from Northeastern University,
Jan. 2018. Shenyang, China, in 1995.
[23] A. Triki-Lahiania, A. B.-B. Abdelghania, and I. Slama-Belkhodjaa, He has been a Post-Doctoral Fellow, a Research
“Fault detection and monitoring systems for photovoltaic installations: A Fellow, and a Visiting Professor in Singapore,
review,” Renew. Sustain. Energy Rev., vol. 82, pp. 2680–2692, Feb. 2018. France, the United States, and Canada, respec-
[24] D. Rossi, M. Omaña, D. Giaffreda, and C. Metra, “Modeling and detec- tively. He is currently a Chair Professor of the
tion of hotspot in shaded photovoltaic cells,” IEEE Trans. Very Large Cheung Kong Scholar Program with the Ministry of
Scale Integr. (VLSI) Syst., vol. 23, no. 6, pp. 1031–1039, Jun. 2015. Education and the Dean of College of Automation
[25] Q. Jiang and X. Yan, “Parallel PCA–KPCA for nonlinear process Engineering, Nanjing University of Aeronautics and
Astronautics, Nanjing, China. He has been the
monitoring,” Control Eng. Pract., vol. 80, pp. 17–25, Nov. 2018.
Principle Investigator on several projects of the National Natural Science
[26] X. Peng, Y. Tang, W. Du, and F. Qian, “Multimode process monitoring
Foundation of China. He has authored 8 books and over 200 referred inter-
and fault detection: A sparse modeling and dictionary learning method,”
national journal and conference papers. His current research interests include
IEEE Trans. Ind. Electron., vol. 64, no. 6, pp. 4866–4875, Jun. 2017. fault diagnosis and fault tolerant control and their applications in aircrafts,
[27] H. Chen, B. Jiang, and N. Lu, “A newly robust fault detection and satellites, and high-speed trains.
diagnosis method for high-speed trains,” IEEE Trans. Intell. Transp. Dr. Jiang was a recipient of the Second Class Award of National Natural
Syst., to be published. doi: 10.1109/TITS.2018.2865410. Science of China in 2018. He currently serves as an Associate Editor
[28] Q. Jiang, X. Yan, and B. Huang, “Performance-driven distributed or Editorial Board Member for a number of journals, such as the IEEE
PCA process monitoring based on fault-relevant variable selection T RANSACTIONS ON C ONTROL S YSTEMS AND T ECHNOLOGY, International
and Bayesian inference,” IEEE Trans. Ind. Electron., vol. 63, no. 1, Journal of Control, Automation and Systems, Nonlinear Analysis: Hybrid
pp. 377–386, Jan. 2016. Systems, Neurocomputing, and Control and Decision, Systems Engineering
[29] H. Chen, B. Jiang, W. Chen, and H. Yi, “Data-driven detection and and Electronics Technologies. He is the Chair of Control Systems Chapter in
diagnosis of incipient faults in electrical drives of high-speed trains,” IEEE Nanjing Section, and a member of IFAC Technical Committee on Fault
IEEE Trans. Ind. Electron., vol. 66, no. 6, pp. 4716–4725, Jun. 2019. Detection, Supervision, and Safety of Technical Processes.
doi: 10.1109/TIE.2018.2863191 Kai Zhang received the B.S. degree in automation
[30] X. He, D. Cai, S. Yan, and H.-J. Zhang, “Neighborhood preserving from Shandong University, Jinan, China, in 2009, the
embedding,” in Proc. IEEE 10th Int. Conf. Comput. Vis., Beijing, China, M.S. degree in control science and engineering from
Oct. 2005, pp. 1208–1213. the University of Science and Technology Beijing,
[31] S. T. Roweis and L. K. Saul, “Nonlinear dimensionality reduction by Beijing, China, in 2012, and the Ph.D. degree
locally linear embedding,” Science, vol. 290, no. 5500, pp. 2323–2326, in electrical engineering and information technol-
Dec. 2000. ogy from the Institute for Automatic Control and
[32] K. Zhang, H. Hao, Z. Chen, S. X. Ding, and K. Peng, “A comparison and Complex Systems, University of Duisburg-Essen,
evaluation of key performance indicator-based multivariate statistics pro- Duisburg, Germany, in 2016.
cess monitoring approaches,” J. Process Control, vol. 33, pp. 112–126, He is currently a Post-Doctoral Fellow with the
Sep. 2015. University of Science and Technology Beijing. His
[33] S. J. Qin, “Survey on data-driven industrial process monitoring and current research interests include process monitoring and fault diagnosis for
diagnosis,” Annu. Rev. Control, vol. 36, no. 2, pp. 220–234, Dec. 2012. industrial processes.
[34] S. Yin, S. X. Ding, X. Xie, and H. Luo, “A review on basic data- Zhiwen Chen was born in Yongzhou, China. He
driven approaches for industrial process monitoring,” IEEE Trans. Ind. received the B.S. degree in electronic information
Electron., vol. 61, no. 11, pp. 6418–6428, Nov. 2014. science and technology and the M.S. degree in
[35] S. X. Ding, Data-Driven Design of Fault Diagnosis and Fault-Tolerant electronic information and technology from Central
Control Systems. London, U.K.: Springer-Verlag, 2014. South University, Changsha, China, in 2008 and
[36] H. Chen, B. Jiang, S. X. Ding, N. Lu, and W. Chen, “Probability-relevant 2012, respectively, and the Ph.D. degree in electri-
incipient fault detection and diagnosis methodology with applications cal engineering and information technology from the
to electric drive systems,” IEEE Trans. Control Syst. Technol., to be University of Duisburg-Essen, Duisburg, Germany,
published. doi: 10.1109/TCST.2018.2866976. in 2016.
[37] B. Jiang, K. Zhang, and P. Shi, “Integrated fault estimation and accom- He is currently with the School of Information
modation design for discrete-time Takagi–Sugeno fuzzy systems with Science and Engineering, Central South University
actuator faults,” IEEE Trans. Fuzzy Syst., vol. 19, no. 2, pp. 291–304, as a Lecturer. His current research interests include model-based and data
Apr. 2011. driven fault diagnosis.