3928 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO.

5, MAY 2019

PV Module Fault Diagnosis Based on

Microconverters and Day-Ahead Forecast
Sonia Leva , Senior Member, IEEE, Marco Mussetta , Senior Member, IEEE, and Emanuele Ogliari

Abstract—The employment of solar microconverter al- degradation of performance [5] may require a statistical analysis
lows a more detailed monitoring of the photovoltaic (PV) of the real-time data for supervision and monitoring [6].
output power at the single module level; thus, machine Therefore, distributed electronics is also helpful, at module
learning techniques are capable to track the peculiarities
of modules in the PV plants, such as regular shadings. In and submodule level, in monitoring the PV production and for
this way, it is possible to compare in real time the day-ahead diagnostic purposes. In fact, the availability of microconverters
forecast power with the actual one in order to better evaluate directly connected to several PV modules, usually with the same
faults or anomalous trends that might have occurred in the characteristics, would also allow an easy but detailed compari-
PV plant. This paper presents a method for an effective fault
son among their performance and working status.
diagnosis; this method is based on the day-ahead forecast
of the output power from an existing PV module, linked to Furthermore, a suitable power forecast, performed in a prog-
a microconverter, and on the outcome of the neighbor PV nostic way, can be employed for the identification and predic-
modules. Finally, this paper also proposes the analysis of tive maintenance of an equipment that will no longer fulfill its
the most common error definitions with new mathematical intended function. Such prediction is usually executed starting
formulations, by comparing their effectiveness and immedi-
from the health state of the component (in particular the PV mod-
ate comprehension, in view of increasing power forecasting
accuracy and performing both real-time and offline analysis ule) and taking into account its past history and future operation.
of PV modules performance and possible faults. Typically, these methods are largely classified as model-based,
data-driven [7], and hybrid [8], respectively.
Index Terms—Day-ahead forecast, fault diagnosis, mi-
croinverter, photovoltaic (PV) system. In particular, model-based methods adopt mathematical rela-
tionships of the degradation process in order to forecast degra-
I. INTRODUCTION dation state [9]. As a matter of fact, data-driven methods are
HE study of the effective configuration of photovoltaic often used when an explicit model is not available, but there
T (PV) systems and their optimization has continuously at-
tracted scientific and industrial research in the last decade.
are enough historical data. These are statistical-based models,
learning trends from the amount of historical data [10].
Moreover, since renewable energy sources (RES) are inter- Among data-driven models can be included the following:
mittent and variable, the availability of undisrupted operation is Autoregressive moving average techniques [11], relevance vec-
extremely important, especially in view of operation and main- tor machines [12], and machine learning techniques, such as
tenance (O & M), among many purposes. artificial neural networks (ANNs), which were often used in the
Generally speaking, distributed module-converter layouts prognostic field [8], [12], [13]. In particular, feedforward ANNs
lead to a higher energy yield by diminishing the effect of mis- have been used for the prediction of rotating machineries [14]
matching and partial shading: Recently, maximum power point and Lithium-ion batteries remaining useful life (RUL) [11], and
tracking (MPPT) efficiency was increased at the module level by echo state network have been used for the fuel cells failure diag-
using dc–dc power optimizers [1] and submodule PV systems nostic [15]. Finally, hybrid approaches combine physics-based
such as dc–ac microinverters connecting a single PV module [2] models of the degradation process with the use of historical data
to the electrical grid or in island mode [3]. collected from degrading components [10].
However, the study of PV systems, along with their spread, Moreover, most of the available bibliographic sources for
has been facing also the problem of plants, which were out prognostics deal with different methods, mainly focusing on
of service due to single components failure [4], and the overall predicting and estimating the RUL of a specific system or com-
ponent. However, the number of problems and possible faults
of PV systems may depend also on different plant layouts and
Manuscript received March 9, 2018; revised June 13, 2018, September
2, 2018, and October 16, 2018; accepted October 22, 2018. Date of typology of installation, i.e., building integrated photovoltaics
publication November 7, 2018; date of current version December 28, [1], [16] or large PV plants [17]–[20].
2018. (Corresponding author: Marco Mussetta.) Given that the forecast horizon together with the temporal
The authors are with the Department of Energy, Politecnico di Milano,
20156 Milano, Italy (e-mail:, sonia.leva@polimi.it; marco.mussetta@ resolution (time sampling) strongly affect the prediction accu-
polimi.it; emanuelegiovanni.ogliari@polimi.it). racy, forecasting models with prognostics purposes should take
Color versions of one or more of the figures in this paper are available into account the different sources of uncertainty affecting pre-
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIE.2018.2879284 dictions [21], such as the following:

0278-0046 © 2018 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.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
LEVA et al.: PV MODULE FAULT DIAGNOSIS BASED ON MICROCONVERTERS AND DAY-AHEAD FORECAST 3929

1) randomness in the equipment future degradation path,

due to the intrinsic stochastic nature of the degradation
process and the unknown future operation and environ-
mental conditions;
2) inaccuracy of the forecasting model;
3) measurement noise or offsets;
4) reliability of the monitoring system (data transmission).
This paper aims to present a method for PV system monitoring
at the module-level, by adopting the day-ahead output power
forecast of existing PV modules, connected to a microconverter,
for an effective fault diagnosis purpose. Additionally, this paper
proposes the analysis of the most common error definitions
employed in PV forecasting, by comparing their effectiveness
with new mathematical formulations of performance indexes
aimed to provide an immediate comprehension of the faulty
status at the PV module level.
This paper is structured as follows. First of all, in Section II,
the daily diagnostic dashboard is presented, where a new strat-
egy for the real time data analysis coupled with the day ahead
power forecast is showed with maintenance purposes. Then, in
Section III, the day-ahead PV output power forecasting method
by means of ANN is presented. In Section IV, a list of error def-
initions usually adopted in PV forecasting is presented and new
ones are proposed in view of maintenance purposes. Section V
presents the real case study considered for validation. Finally,
results of the proposed method are shown in Section VI , and
comments and conclusions are drawn in Section VII.

II. DAILY DIAGNOSTIC DASHBOARD

The methodology that has been applied for monitoring and
diagnostic purposes entails two distinct steps concerning the PV
Fig. 1. Flowchart for the “faulty samples” detection and the diagnostic
module output: dashboard composition.
1) real-time monitoring analysis;
2) offline analysis of the recorded data.
In particular, the first step of this diagnostic method leads to i
related sample standard deviation Sp,t (as defined in
the “daily diagnostic dashboard” composition, which provides
Section III, 3).
an immediate information about the status of the monitored
2) The measured power Pmi ,t of each ith PV mod-
PV modules by comparing in real time the actual power with
ule/converter is recorded. This is the mean power mea-
the forecast and the actual trends of the neighbor PV mod-
sured in the tth time sample (here a time resolution of
ules. Instead, the offline analysis is performed at the end of the
1 min is considered, but for the sake of generality, a
day providing useful indicators both in determining a faulty or
different sampling can be considered by averaging the
anomalous trend and for the employment of reliable data in
measured values over the specific time interval).
the continuous training of the forecasting model, as shown in
3) The actual power Pmi ,t is compared to the correspondent
Section VI. i
forecast Pp,t for each ith PV module/converter. When
a module is not working properly, usually its measured
A. Real-Time Monitoring Analysis
power falls significantly below the forecast, and this may
The real-time monitoring analysis follows the flowchart re- indicate anomalies. Therefore, if the following condition
ported in Fig. 1, and it is based, first, on the comparison between is verified for a single PV module/converter:
the measured power and its forecast, and second, between the
measured power of similar modules, named as “neighbors.” In Pmi ,t < Pp,t
i i
− Sp,t (1)
particular, a different power prediction is computed for every PV
module/converter. Estimation of the power forecast is presented a “1st level” alert is set on the status of the relevant
in detail in Section III. module.
The implemented procedure includes the following steps. 4) If the 1st level alert condition is verified, the follow-
i
1) The predicted power Pp,t in the tth time sample is ing check is performed between the given diagnostic in-
calculated for each ith PV module/converter, with the dicator ε calculated in the time interval Δt of the ith

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
3930 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 5, MAY 2019

TABLE I
CONFUSION MATRIX OF THE POSSIBLE ALERTS COMBINATION

anomalous module and the kth neighbors:

εiΔ t > εkΔ t , ∀k = i (2) Fig. 2. PHANN forecasting method.

where ε is any of the diagnostic indicators defined in

Section IV, and the here considered Δt is 15 min; if this c) the remaining combinations will result in a “soft
last condition is verified a “2nd level” alert is set on the fault” condition: the module has no need for main-
status of the related microconverter. tenance yet, however recorded data are not reliable
The status of all the PV modules is determined on the basis of for future uses in model training, as detailed later
the flowchart that has been explained, and a different color on in Section VI-C.
the daily diagnostics dashboard is given according to the level The faults description is provided in Section V. An example
of the alert. If the 1st level alert is reached the background of the of the real-time and offline analyses will be presented later in
relevant microinverter graph on the dashboard is yellow colored, Section VI.
and the color is kept even if the anomalous trend is subsided. It is worth mentioning that the selected thresholds have been
In the case of a faulty situation ascribed to the PV module, defined and tested considering the behavior of the modules in
a 2nd level alert is given, and the background of the related the whole year 2017, with particular reference to those days
microinverter graph is red-colored on the diagnostic dashboard. presenting anomalies, as described later in Sections V and VI.

B. Offline Analysis III. FORECASTING METHOD

The offline analysis is performed at the end of the day on The day-ahead forecast of the PV output power can be per-
the basis of the diagnostics indicators presented in detail in formed adopting several methods such as physical, statistical,
Section IV. It follows these steps: or hybrid ones. Although it can be achieved in many ways, hy-
1) daily diagnostic indicators calculation for each module, brid methods proved their effectiveness because they merge in
as presented in Section IV; a unique method the upsides of the others [22].
2) mean values μi and standard deviation σi calculation of Statistical methods, i.e., based on ANN, rely on the historical
the diagnostic indicators for all the PV modules, exclud- data measurements (weather conditions and produced power)
ing the ith one, and setting of the related μi + σi and for building up a suitable model of the plant. These have shown
μi + 3σi thresholds; the main issue of committing higher percentage errors in specific
3) if the diagnostic indicators of the ith module are falling hours of the day, in particular early in the morning and in the
below the μi + σi threshold, then the module/converter sunset hours [23].
is considered “healthy”; Moreover, the so called clear sky solar radiation model
4) if the diagnostic indicators of the ith module/converter (CSRM) [24] is a physical model that computes the solar irra-
are falling between the μi + σi and μi + 3σi thresholds, diance available without clouds for any tilted surface, whatever
then the ith module is given a 3rd level alert; oriented, at a given location.
5) if the diagnostic indicators of the ith module are greater As shown in [23], when CSRM is included as an additional
than the μi + 3σi threshold, then the module will receive input of the ANN, it significantly reduces the error by defining
the 4th level alert; the maximum hourly quantity of the solar irradiance and the
6) matching between the alerts given by the diagnostic indi- expected daily hours of sunlight available for any specific lo-
cators (3rd and 4th level alert) and the generated real-time cation. This hybrid model, here adopted, is the physical hybrid
alerts (1st and 2nd level alert), according to the confusion artificial neural network (PHANN).
table reported in Table I:
a) the combination of a 1st level alert and 3rd level
A. PHANN Forecasting Method
alert determines a “no fault” condition for the given
module/converter; As shown in Fig. 2, PHANN forecasting method adopts ANN
b) the combination of a 2nd level alert and 4th level ability to learn from historical data the existing relationships
alert determines the “hard fault” condition for the among the weather forecast and the measured output power
given module, requiring specific maintenance; of the PV plant. Besides, in order to enhance the forecasting

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
LEVA et al.: PV MODULE FAULT DIAGNOSIS BASED ON MICROCONVERTERS AND DAY-AHEAD FORECAST 3931

TABLE II
PARAMETERS EMPLOYED IN THE ANALYSIS

Fig. 3. Ensemble forecast P p , t of a single module output power with

the relative standard deviation S p , t in a generic given day.

B. Ensemble Forecast
It has been proved in literature [27] that similar network
abilities, the deterministic solar irradiance under clear sky con-
models can produce slightly different results even if trained on
ditions (CSRM) is provided as an additional input.
the same dataset. This is due to the stochastic nature of ANN.
Table II shows the complete list of the parameters, provided
Therefore, it was demonstrated that it is possible to reduce the
by the weather forecast service, which are employed as the input
forecasting error, by averaging the single output xj of several
neurons together with the CSRM. In the end of this table the
parallel ANNs [28].
dc output power, which is the forecast provided by the output
Thus, the above described process for PHANN training can
neuron of the PHANN, is also reported.
be repeated on a number NE of parallel networks (trials) in
Learning phase is a very important step in ANN-based meth-
order to produce an average forecast x for each time sample,
ods and many issues are related to it, as reported in [25] and
which is the mean value of all xj samples.
[26]. In fact, PHANN method needs training with the super-
Moreover, due to the stochastic base of the PHANN, also the
vised learning. The supervised learning undergoes the matching i
sample standard deviation Sp,t of all the trials can be computed
between the historical input (weather parameters forecast and
(in each time sample t and for each module i) as follows:
CSRM) to the actual output (measured dc output power). Avail-

able historical data (days of measurements) are divided into the NE 2
following two subdatasets, with different aims: i j =1 (xj − x)
Sp,t = . (3)
1) training set: used by ANN training procedure to learn NE − 1
relationships between input and output; Equation (3) has the expression of the sample standard devi-
2) validation set: used by ANN training procedure to check ation, as the mean of the population of all the trials is not known
if the model of the relationships is right. a priori.
The amount of data that have to be included in the differ- In this paper, the daily time series of x represents the PV out-
ent subsets have to pass through a sensitivity analysis, which put power profile Pp , and it is referenced as the daily “ensemble
has been previously performed by the authors, as detailed in forecast.” In Fig. 3, the violet area Pm ,t is the actual power of the
[26]. Therefore, assuming a database of historical measure- PV module measured minutely by the microinverter “A-02”; on
ments continuously updated day by day, 90% of the available the left top corner of the picture, real-time diagnostic indicators
days are picked randomly to be used during training and the are reported, as detailed in Section IV.
remaining 10% are for the validation set. This approach allows
a continuous improvement of the prediction accuracy, as shown
in [25]. IV. DIAGNOSTIC INDICATORS
Moreover, since the real power measured in the previous days Because of the aforementioned intents, a steadily comparison
is used to train the ANN, the forecasting model can be employed between the actual energy and the expected one is extremely
either for the whole PV plant or for each single module, as in the important in order to continuously check the system operation.
case described in this paper, where the use of microconverters Hence, some statistical indicators, gathering the main features
provides the availability of detailed power measurements from of the expected output power, provide an useful support for a
each single module. reliable comparison with the actual power. As it is more likely

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
3932 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 5, MAY 2019

to have a not null difference between the expected power and where the numerator is the same as WMAE, while the denomi-
the actual one, a simple formulation of the error committed in nator is the sum of the maximum between the forecast and the
the tth sample of time is easy to find. Things become harder measured power.
in finding a usable error definition, which could be broadly Additionally, a new diagnostics indicator is introduced here.
used not only for diagnostic purposes but also which could It draws from the performance ratio PR% coefficient as it is
give an effective assessment on the accuracy of the forecast at a expressed in the IEC 61724 norm [33]
glance. These indicators assess the forecasts accuracy providing N P m , t
a mathematical relation between the expected value and the t=1 Pn
PR% = N G POA, t
· 100 (10)
actual value in the same sample of time. t=1 G STC
The error et made in the tth sample of time is the starting N
definition given as the difference between the values of the Pm ,t GSTC
= Nt=1 · · 100. (11)
power measured Pm ,t and the forecast Pp,t [29] t=1 GPOA,t
Pn

et = Pm ,t − Pp,t (W). (4) This indicator gives a good evaluation of the exploited energy
from the PV system, by comparing it with the solar irradiance
From the expression of the absolute value |et |, most com- measured. Instead of the actual irradiance on the plane of array
monly used error definitions can be inferred [30], [31], such as (GPOA ), which is not always an available parameter, we adopted
the following. the following objective mean absolute error (OMAE):
1) The well-known mean absolute percentage error N
|et | GSTC
(MAPE), i.e., OMAE = N t=1 · · 100 (12)
cs
t=1 GPOA,t
Pn
N  
1   et 
MAPE = · 100 (5) where
N t=1  Pm ,t 
1) GcsPOA,t is the solar irradiance on the plane of the array
where N is the number of considered time samples. In given by the clear sky solar irradiance model (CSRM) as
this analysis, it is calculated for a single day. it is described in [24];
2) The normalized mean absolute error (NMAE), i.e, 2) GSTC is the solar irradiance at standard test conditions
(equals 1000 W/m2 ).
N
|et | From (12), it is therefore possible to rewrite the OMAE
NMAE = t=1 · 100 (6)
N · Pn highlighting the existing relationship with the former indica-
tor NMAE
where the percentage of the absolute error is referred to
GSTC
the rated power Pn of the PV module, instead of the OMAE = NMAE · N · N. (13)
cs
hourly measured power Pm ,t . t=1 GPOA,t
3) The weighted mean absolute error (WMAE) that is based As it can be seen in (13), the relationship between NMAE and
on total energy production the new diagnostic indicator OMAE is set. Finally, it is notice-
N able how EMAE and OMAE are both limited between 0 and
|et |
W M AE = Nt=1 · 100. (7) 100%, providing an immediate indication on the magnitude of
t=1 Pm ,t the daily diagnostic error, which is more likely related to a fault
4) The normalized root mean square error (nRMSE) is occurrence.
based on the maximum power output max(Pm ,t )
N V. CASE STUDY
2
t=1 |e t |
N
In view of the validation of the proposed procedure, reported
nRMSE = · 100. (8) in Section VI, here, we will use experimental data collected in
max(Pm ,t )
the year 2017 at Solar Tech Lab, Milano, Italy. The whole PV
However, the daily evaluation indexes, expressed in (6), (7), plant is constituted by 21 silicon modules with different features;
and (8), could significantly differ, with WMAE and nRSME of- all the PV modules are facing South (0◦ azimuth) and lay on
ten showing values above 100%; thus, being not able to provide fixed structures that are tilted at an angle of 30◦ (see Fig. 4).
a complete information “at a glimpse” on the accuracy of the The MPPT is guaranteed by the micro dc–ac solar converters,
prediction. which are installed under each module, as described in [26]. The
Starting from these assumptions, and in view of a more useful electrical parameters of the PV modules are collected remotely
summary evaluation, additional performance indexes are pro- by the monitoring system; thus, the maximum dc power values
posed, aimed to provide a value between 0 and 100% of the are recorded minutely.
forecast accuracy [32]. Therefore, the envelope-weighted mean For example, if we consider Fig. 5, it is possible to notice
absolute error (EMAE) is defined as that the PV module connected to the microinverter “i85-87”
N had some problems (the dc output power from the PV module
t=1 |et |
EMAE = N · 100 (9) is constantly equal to 0 W). This faulty behavior can be either
t=1 max(P m ,t , Pp,t ) related to a fault that occurred in the data transmitter or to the

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
LEVA et al.: PV MODULE FAULT DIAGNOSIS BASED ON MICROCONVERTERS AND DAY-AHEAD FORECAST 3933

the same modules. In addition, we will illustrate the acceptance

criterion of the recorded sample in view of the forecasting model
training, on the basis of the daily diagnostic indicators scored
by each microinverter PV module set.

A. Real-Time Monitoring
An example of the daily diagnostic dashboard for 13th April
2017 at 4:45 P.M. is provided in Fig. 6. It can be noticed that
PV modules connected to microinverters “C-03” and “A-05” are
colored differently because of different alerts. In the first case,
the actual power of the PV module provisionally differed from
the forecast because of a temporarily shading, but returned to
normal later. In the second case, the measured power constantly
showed an anomalous lower trend compared both to the forecast
and to the neighbor PV modules.
This analysis brought to the following four scenarios for each
Fig. 4. Solar Tech Lab picture on 13th April 2017 afternoon, with shad- module.
ings of the railings on the PV modules.
1) “Healthy” or “good day”: the online analysis gives
no alerts related to anomalous trend (see, for instance,
Fig. 3).
2) Regularly “partially shaded day,” that is, either the shad-
ing has been properly forecasted or the neighbor PV mod-
ules show the same trend. For instance, as shown in Fig. 7,
the partially shading in the morning is consolidated in the
historical data of the PV module and the related forecast
clearly shows this well-known trend. This condition still
gives no alert for anomalies.
3) Unexpected partially shaded PV module. In this case, the
power produced by the PV module diverges from the fore-
cast and its trend is not consistent with the neighbors. For
example, as shown in Fig. 8, the PV module connected
to microinverter “C-03” has been intentionally partially
shaded, starting from 11:45. As reported in Fig. 8, start-
ing from that moment the measured power differed from
the expected one sufficiently to generate a 1st level alert
in the online analysis. In addition, neighbor modules did
not behave in the same way and a 2nd level alert was gen-
erated. Finally, as the shading suddenly stopped at 12:15,
the alert remained on the daily dashboard signaling the
anomalous trend of the system. In this case, for this mod-
Fig. 5. DC output power of 21 PV modules, recorded at Solar Tech Lab ule, data were recorded as “soft fault,” as described in
during the 6th May 2017, with faulty samples for the converter i85-87. Section II and in Table I.
4) “Faulty day.” In this case, the PV module output power
has a completely anomalous trend, i.e., it differs both from
dc–ac converter. On the other hand, regular partial shadings
the related forecast and from the neighbors’ trends. For
occurring in the morning and in the afternoon can be easily
example, as shown in Fig. 9, the actual power of mod-
recognized in the different output of some PV modules.
ule “A-05” is significantly below the predicted power,
and the computed diagnostic indicators are high, this
VI. EXAMPLES AND DISCUSSION
means that a 2nd level alert is still on in the current
In Section II, a dual-step diagnostic method was presented. time when the online analysis is performed.
Here, we present and discuss the results obtained by applying This steep power loss is not associated to the regular
the proposed approach on measured data from the experimental ageing of the PV module; hence, this sudden behavior
plant described in the previous section. In particular, an example was not predicted by the forecasting method. In fact,
of the “daily diagnostic dashboard” is presented, showing real- the regular ageing usually occurs along the years; there-
time alerts for 15 specific microinverter PV-module status for fore, it can be predicted by the forecasting procedure de-
a given day; then, the related offline analysis is presented for scribed in Section III, since this is continuously trained on

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
3934 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 5, MAY 2019

Fig. 6. Daily diagnostic dashboard for real-time monitoring of 15 modules (current status on 13th April 2017, afternoon).

Fig. 7. Regularly “partially shaded” PV module in the morning. Fig. 8. Unexpected partially shaded PV module.

measured historical data for any specific module. In this

case, as shown later, the complete diagnostic analysis will
values in all the indicators are “A-05” and “C-03.” The high-
provide an “hard fault” alert. Indeed, after this alert, the
est indicator is WMAE which exceeds 100%, while NMAE is
PV module was inspected and many cells showed mi-
the lowest. Instead it is noticeable that OMAE and EMAE are
crocracks [34] on the surface, which brought the output
greater than the regular ones (the indicators for the other mi-
power to decrease 20% within a few days.
croinverters are below 10% in all the cases), but they are not
exceeding 100% maximum cap even if they are related to a
B. Offline Analysis probable faulty situation.
At the end of the day, given that in the online analysis the In particular, an example of nRMSE for the offline analysis is
alerts referred to possible faults have been already set, the diag- shown in Fig. 10, highlighting the most critical anomalies with
nostic indicators listed in Section IV are collected and analyzed respect to the thresholds defined in Section II-B.
offline, as explained in Section II-B. The mean diagnostic indi- Therefore, on the basis of the offline statistical analysis, it is
cators computed for the considered day are reported as an ex- now possible to point out the 3rd- and 4th-level alerts. By means
ample in Table III. The microinverters scoring the highest daily of the alerts combination that has been previously described in

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
LEVA et al.: PV MODULE FAULT DIAGNOSIS BASED ON MICROCONVERTERS AND DAY-AHEAD FORECAST 3935

Fig. 10. nRMSE offline analysis with respect to thresholds.

Fig. 9. “Faulty day” for the PV module connected to “A-05” microin-

verter.

TABLE III
MEAN DIAGNOSTICS INDICATORS RECORDED ON 13th APRIL 2017

Fig. 11. Δ of the “faulty days” in 2017 with respect to nRMSE.

fault” alerts (including all the possible anomalies that could be

referred to the PV system, from the module to the PV converter,
before the connection to the grid).
To properly set the threshold for the offline analysis, a Δ
value can be considered to classify each diagnostic indicator for
the ith microconverter, e.g., for the nRMSE it can be defined as
nRMSEi − μi
Δi = (14)
σi
where μi and σi are the values defined in Section II-B. As re-
ported in Fig. 11, the values of Δ with respect to nRMSE are
usually defined in the range proposed in Section II-B, when a
“soft fault” was identified (σ ÷ 3σ). However, the alerts for
Section II-B and Table I, the faulty day condition for the “A- these 38 days can be further analyzed ex-post, as proposed
05” inverter is determined as an “hard fault.” Instead “C-03” above, to evaluate how many false positive the method gen-
inverter is marked as a “soft fault,” which means that it is not erated; in particular, the operator inspection confirmed that the
in fault condition yet, but it is likely to have problems in the “soft fault” alerts (marked in blue in Fig. 11) were mostly re-
future if the same behavior is confirmed in the following days. lated to maintenance and other experimental activity conducted
Thus, if no additional alerts are found later, the module can be in the laboratory, which caused isolated performance degrada-
considered healthy, while if a “soft fault” condition is found for tion similar to the one shown in Fig. 8. On the other hand, the
consecutive days, the module will need to be carefully inspected values highlighted in red in Fig. 11 represent the cases when the
by an operator in order to verify the cause of this systematic soft-fault alert was repeated on the same module for consecutive
performance degradation. For this reason, data measured from days. In this case, an inspection was necessary to investigate the
modules in “soft fault” condition will not be considered for cause and, although the power reduction was not so relevant
the training of the model for the next days, as explained in the to activate a hard-fault alert, it was found that the module pre-
following. sented issues (hot spots) that caused a permanent performance
In order to define the method effectiveness based on experi- reduction.
mental data, we have considered the measurements of one full It is worth mentioning that the module performance can also
year; in particular, in our analysis 307 days were available for be affected by the converter. Indeed, each module is connected to
the whole year 2017 (considering only valid measurements and a micro dc–ac solar converters, as described in [26]; in this case,
removing incomplete data, related for instance to scheduled it could be difficult to distinguish the real source of degradation.
maintenance); among these, 38 days turned out to have “soft In fact, the proposed method is aimed to highlight possible faults

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
3936 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 5, MAY 2019

TABLE IV
MEAN DIAGNOSTICS INDICATORS

based on the analysis of measured output power, but it is not

able to detect a specific kind of fault or to directly identify the
reasons. Therefore, as already mentioned, a careful operator
check would be needed to determine the cause of any alert, but
the proposed algorithm is able to locate where a fault condition
happens and its degree, providing proper information to the
system maintenance operators. In particular, with respect to
a possible failure of microconverters, these generally have a
monitoring system of their status; thus, data will be completely Fig. 12. OMAE of the “faulty days” in 2017 with different training
datasets.
missing or null in case of a hard fault, which is easy to detect;
our analysis is based on measurements provided by healthy
converters: If a converter has a failure, this will be clear from
missing data. approach C3 is able to better highlight faults with respect to
normal condition. Since our target is to increase the general
forecasting accuracy and to emphasize fault conditions with
C. Training Data Selection for Diagnostic Purposes
respect to the predicted outcome of the plant, results suggest
As reported in [26], the training approach and data selection that it is better to exclude from training the previous days
are critical aspects for the effectiveness of forecasting, in partic- with detected “soft fault” condition, in order to have worse
ular when an online training is performed daily to continuously error indicators when fault conditions occur and to better detect
update the forecasting model. faults.
For this reason, here, we investigated the effect of including,
in the training of the PHANN, the previously recorded measure-
ments of inverter with alerts: This evaluation is performed by VII. CONCLUSION
analyzing the resulting forecast accuracy. An effective diagnostic method for fault detection at PV-
In particular, we conducted an analysis based on forecasting module level was presented in this paper, based on monitored
faulty modules by considering two different training datasets, data by microinverter and day-ahead power forecasting. A suit-
with previous “soft fault” condition either included or not; as able procedure has been implemented for real-time monitoring
mentioned in Section VI-B, out of 307 days available, 38 days fault conditions by comparing neighbor PV modules and for
turned out to have “soft fault” alerts. performing a final offline evaluation at the end of the day, by
In Table IV, the mean diagnostics indicators are reported with means of properly defined diagnostic indicators. In particular,
respect to the following configurations: this method was aimed to provide a useful help in the fault de-
1) this subset contains the power forecast for 269 days, tection at the PV-module level. The classification of faults and
based on a training dataset of 269 healthy days (data suggestion of possible measures were beyond the scope of this
from healthy modules); work and will be subject of further research.
2) this subset includes the forecast of 38 days with measured The proposed method has been validated on 15 PV modules,
fault conditions: forecast here is based on a mixed training and the related microinverters in a laboratory facility, consider-
dataset composed by 231 healthy and 38 faulty days; ing also one year of measurements for the definition of proper
3) this subset includes the forecast of 38 days with mea- thresholds for the identification of different fault degrees. Re-
sured fault conditions: forecast here is based on a training sults showed that the proposed approach is able to identify criti-
dataset of 269 healthy days (the same training dataset as cal failures of PV modules, avoiding systematic errors like age-
C1). ing or regular shadows, by means of a machine learning based
In all the three cases considered, the number of days con- forecasting, specific for each module. Moreover, the employ-
sidered in the training dataset was 269, for the sake of a fair ment of specifically defined diagnostic indicators can provide
comparison. an immediate comprehension of the fault status.
Moreover, Fig. 12 shows the daily OMAE of the 38 faulty The proposed approach can be easily extended to optimiz-
days forecasted with approaches C2 and C3. As reported in ers in view of future implementations, by adding the feature
this figure, the forecast error is slightly lower when previous of developing or the day-ahead forecast and the logic of the
“faulty days” are included in the training set (C2), while the comparison among different PV module output on board.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.
LEVA et al.: PV MODULE FAULT DIAGNOSIS BASED ON MICROCONVERTERS AND DAY-AHEAD FORECAST 3937

REFERENCES [25] F. Grimaccia, S. Leva, M. Mussetta, and E. Ogliari, “Ann sizing procedure
for the day-ahead output power forecast of a PV plant,” Appl. Sci., vol. 7,
[1] R. W. Erickson and A. P. Rogers, “A microinverter for building-integrated no. 6, pp. 1–13, 2017, Art. no. 622.
photovoltaics,” in Proc. 24th Annu. IEEE Appl. Power Electron. Conf. [26] A. Dolara, F. Grimaccia, S. Leva, M. Mussetta, and E. Ogliari, “Compari-
Expo., 2009, pp. 911–917. son of training approaches for photovoltaic forecasts by means of machine
[2] H.-J. Chiu et al., “A module-integrated isolated solar microinverter,” learning,” Appl. Sci., vol. 8, no. 2, pp. 1–16, 2018, Art. no. 228.
IEEE Trans. Ind. Electron., vol. 60, no. 2, pp. 781–788, Feb. 2013. [27] M. Rana, I. Koprinska, and V. G. Agelidis, “Forecasting solar power
[3] C. T. Rodrı́guez, D. V. de la Fuente, G. Garcera, E. Figueres, and J. A. G. generated by grid connected PV systems using ensembles of neural net-
Moreno, “Reconsurable control scheme for a PV microinverter working works,” in Proc. Int. Joint Conf. Neural Netw., Jul. 2015, pp. 1–8. [Online].
in both grid-connected and island modes,” IEEE Trans. Ind. Electron., Available: http://dx.doi.org/10.1109/IJCNN.2015.7280574
vol. 60, no. 4, pp. 1582–1595, Apr. 2013. [28] M. Omar, A. Dolara, G. Magistrati, M. Mussetta, E. Ogliari, and
[4] A. Triki-Lahiani, A. B.-B. Abdelghani, and I. Slama-Belkhodja, “Fault F. Viola, “Day-ahead forecasting for photovoltaic power using arti-
detection and monitoring systems for photovoltaic installations: A review,” ficial neural networks ensembles,” in Proc. IEEE Int. Conf. Renew.
Renew. Sustain. Energy Rev., vol. 82, pp. 2680–2692, 2018. Energy Res. Appl., Nov. 2016, pp. 1152–1157. [Online]. Available:
[5] J. Bastidas-Rodriguez, G. Petrone, C. Ramos-Paja, and G. Spagnuolo, http://dx.doi.org/10.1109/ICRERA.2016.7884513
“Photovoltaic modules diagnostic: An overview,” in Proc. 39th Annu. [29] C. Monteiro, L. A. Fernandez-Jimenez, A. Ramirez-Rosado, I. J. ans
Conf. IEEE Ind. Electron. Soc., 2013, pp. 96–101. Munoz-Jimenez, and P. M. Lara-Santillan, “Short-term forecasting mod-
[6] S. Vergura, G. Acciani, V. Amoruso, G. E. Patrono, and F. Vacca, “Descrip- els for photovoltaic plants: Analytical versus soft-computing techniques,”
tive and inferential statistics for supervising and monitoring the operation Math. Problems Eng., vol. 2013, 2013, Art. no. 767284. [Online]. Avail-
of PV plants,” IEEE Trans. Ind. Electron., vol. 56, no. 11, pp. 4456–4464, able: http://dx.doi.org/10.1155/2013/767284
Nov. 2009. [30] R. Ulbricht, U. Fischer, W. Lehner, and H. Donker, “First steps towards
[7] X.-S. Si, W. Wang, C.-H. Hu, and D.-H. Zhou, “Remaining useful life a systematical optimized strategy for solar energy supply forecasting,” in
estimation–A review on the statistical data driven approaches,” Eur. J. Proc. Eur. Conf. Mach. Learn. Princ. Pract. Knowl. Discovery Databases,
Oper. Res., vol. 213, no. 1, pp. 1–14, 2011. 2013, pp. 14–25.
[8] T. Brotherton, G. Jahns, J. Jacobs, and D. Wroblewski, “Prognosis of [31] J. Kleissl, Solar Energy Forecasting and Resource Assessment. New York,
faults in gas turbine engines,” in Proc. IEEE Aerosp. Conf., 2000, vol. 6, NY, USA: Academic, 2013.
pp. 163–171. [32] A. Nespoli, E. Ogliari, F. Grimaccia, S. Leva, and M. Mussetta, “Validation
[9] G. Petrone, G. Spagnuolo, R. Teodorescu, M. Veerachary, and M. Vitelli, of ANN training approaches for day-ahead photovoltaic forecasts,” in
“Reliability issues in photovoltaic power processing systems,” IEEE Proc. Int. Joint Conf. Neural Netw., Jul. 2018, pp. 4567–4572.
Trans. Ind. Electron., vol. 55, no. 7, pp. 2569–2580, Jul. 2008. [33] IEC, Photovoltaic System Performance - Part 1: Monitoring, International
[10] M. Schwabacher, “A survey of data-driven prognostics,” in Proc. In- Electrotechnical Commission 61724-1, 2017.
fotech@ Aerosp. Conf., 2005, pp. 887–891. [34] A. Dolara, S. Leva, G. Manzolini, and E. Ogliari, “Investigation on perfor-
[11] B. Saha, K. Goebel, and J. Christophersen, “Comparison of prognostic mance decay on photovoltaic modules: Snail trails and cell microcracks,”
algorithms for estimating remaining useful life of batteries,” Trans. Inst. IEEE J. Photovolt., vol. 4, no. 5, pp. 1204–1211, Sep. 2014.
Meas. Control, vol. 31, no. 3–4, pp. 293–308, 2009.
[12] K. Goebel, B. Saha, and A. Saxena, “A comparison of three data-driven
techniques for prognostics,” in Proc. 62nd Meeting Soc. Machinery Failure
Sonia Leva (SM’13) received the Ph.D. degree
Prevention Technol., 2008, pp. 119–131.
in electrical engineering from the Politecnico di
[13] L. Peel, “Data driven prognostics using a Kalman filter ensemble of neural Milano, Milano, Italy, in 2001.
network models,” in Proc. Int. Conf. IEEE Prognostics Health Manage.,
She is currently a Full Professor of electrical
2008, pp. 1–6.
engineering with the Department of Energy, Po-
[14] A. K. Mahamad, S. Saon, and T. Hiyama, “Predicting remaining useful
litecnico di Milano, where she is also a Director
life of rotating machinery based artificial neural network,” Comput. Math. of the Solar Tech Lab and the Laboratory of Mi-
Appl., vol. 60, no. 4, pp. 1078–1087, 2010.
cro Grids.
[15] S. Morando, S. Jemei, R. Gouriveau, N. Zerhouni, and D. Hissel, “Fuel
Prof. Leva is a member of the IEEE Work-
cells prognostics using echo state network,” in Proc. 39th Annu. Conf.
ing Group “Distributed Resources: Modeling and
IEEE Ind. Electron. Soc., 2013, pp. 1632–1637.
Analysis” and a senior member of the IEEE
[16] S. Rosiek, J. Alonso-Montesinos, and F. Batlles, “Online 3-h forecasting
Power and Energy Society.
of the power output from a BIPV system using satellite observations and
ANN,” Int. J. Elect. Power Energy Syst., vol. 99, pp. 261–272, Jul. 2018.
[17] M. K. Alam, F. Khan, J. Johnson, and J. Flicker, “A comprehensive re-
view of catastrophic faults in PV arrays: Types, detection, and mitigation
techniques,” IEEE J. Photovolt., vol. 5, no. 3, pp. 982–997, May 2015. Marco Mussetta (SM’17) received the Ph.D. de-
[18] M. K. Alam, F. H. Khan, J. Johnson, and J. Flicker, “PV faults: Overview, gree in electrical engineering from the Politec-
modeling, prevention and detection techniques,” in Proc. IEEE 14th Work- nico di Milano, Milano, Italy, in 2007.
shop Control Model. Power Electron., 2013, pp. 1–7. He is currently an Associate Professor of elec-
[19] M. Aghaei, F. Grimaccia, C. A. Gonano, and S. Leva, “Innovative auto- trical engineering with the Department of En-
mated control system for PV fields inspection and remote control,” IEEE ergy, Politecnico di Milano.
Trans. Ind. Electron., vol. 62, no. 11, pp. 7287–7296, Nov. 2015. [Online]. Prof. Mussetta is the Chair of the IEEE
Available: http://dx.doi.org/10.1109/TIE.2015.2475235 Computational Intelligence Society Task Force
[20] A. Dolara, G. C. Lazaroiu, S. Leva, G. Manzolini, and L. Votta, “Snail on “Fuzzy Systems in Renewable Energy and
trails and cell microcrack impact on PV module maximum power and Smart Grid.” He is a senior member of the IEEE
energy production,” IEEE J. Photovolt., vol. 6, no. 5, pp. 1269–1277, Industrial Electronics Society.
Sep. 2016.
[21] P. Baraldi, F. Mangili, and E. Zio, “Investigation of uncertainty treatment
capability of model-based and data-driven prognostic methods using sim-
ulated data,” Rel. Eng. Syst. Saf., vol. 112, pp. 94–108, 2013. Emanuele Ogliari received the Ph.D. degree in
[22] E. Ogliari, A. Dolara, G. Manzolini, and S. Leva, “Physical and hybrid electrical engineering from the Politecnico di Mi-
methods comparison for the day ahead PV output power forecast,” Renew. lano, Milano, Italy, in 2016.
Energy, vol. 113, pp. 11–21, 2017. He has been working on photovoltaics power
[23] A. Dolara, F. Grimaccia, S. Leva, M. Mussetta, and E. Ogliari, “A physical plant design and their optimization since 2010
hybrid artificial neural network for short term forecasting of PV plant and RES expected power by means of compu-
power output,” Energies, vol. 8, no. 2, pp. 1138–1153, 2015. tational intelligence techniques since 2012. He
[24] F. Kasten and G. Czeplak, “Solar and terrestrial radiation dependent on is currently an Assistant Professor with the De-
the amount and type of cloud,” Sol. Energy, vol. 24, no. 2, pp. 177–189, partment of Energy, Politecnico di Milano.
1980.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on February 12,2021 at 13:54:22 UTC from IEEE Xplore. Restrictions apply.