Renewable and Sustainable Energy Reviews 170 (2022) 112968

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser

Performance analysis and modelling of a 50 MW grid-connected

photovoltaic plant in Spain after 12 years of operation
Enrique Fuster-Palop a, b, Carlos Vargas-Salgado a, Juan Carlos Ferri-Revert c, Jorge Payá a, *
a
Instituto Universitario de Investigación en Ingeniería Energética, Universitat Politècnica de València, Camí de Vera s/n, Valencia, Spain
b
ImpactE, C/ Joan Verdeguer nº 16, Valencia, 46024, Spain
c
Entoria Energy, Avenida Maisonnave, 41, 7 C D, 03003, Alicante, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: This study aims to estimate the performance and losses of a 50 MW photovoltaic (PV) utility-scale after 12 years
Photovoltaic of operation. The PV plant has monocrystalline and polycrystalline silicon modules and is located in the central
PV utility-Scale monitoring region of Spain with an annual insolation of 1976 kWh/m2. Monitoring data over the entire year 2020 has been
Performance ratio
analyzed and filtered to assess the performance results following the IEC 61724 standard guidelines. The annual
Low irradiance losses
Multiple linear regression
average reference yield, final yield, performance ratio and capacity utilization factor are of 5.44 h/d, 4.28 h/d,
Random forest 79.24%, and 19.77%, respectively. Besides the experimental analysis, this work improves the estimation of the
daily performance ratio, especially in days with low insolation. Two different modelling approaches have been
assessed and compared. In first place, a physical model has been adopted, based on the most common losses, and
including an exponential expression to account for low irradiance losses. In second place, statistical models have
been used, with either multiple linear regressions or random forest algorithms. In contrast with other published
models which require many inputs, the best accuracy has been reached with the random forest model using only
the ambient temperature and solar irradiance as predictors, obtaining a RMSE of 1% for the PR and for the
energy production.

(50 MW) for which the performance is reported in literature.

The PV energy production potential estimation is essential to provide
1. Introduction
more accuracy in the design and monitoring stages of new PV utility-
scales and to guarantee their integration to the power grid [9], and a
Photovoltaic (PV) energy systems are a key technology to increase
proper performance and reliability throughout their life-cycle [11]. For
the share of renewables in the energy mix, especially in countries with a
this purpose, commercial modelling softwares are generally employed,
high solar resource. In the last decade, the rapid cost reduction of up to
with a reliability which depends on the accuracy of the irradiance and
82% [1], together with the favorable decarbonization policies [2], has
electrical submodels [12]. The latter includes parameters such as the
increased exponentially the global PV installed capacity from a total of
power losses at different stages of the facility, namely, the performance
72 GW in 2011 to 707.5 GW in 2020 [3].
ratio (PR).
Literature on the operation of large photovoltaic plants is rather
In addition to the study of the performance of a power plant, this
recent. Most of these plants are located in hot, desert, arid or semi-arid
paper also investigates the modelling of the PR as one of the main points
climates, such as 5 MWp in Sivagangai (India) [4], 9.36 MWp in Gujarat
to estimate the AC energy yield (EAC) in PV systems using irradiance
(India) [5], 10.13 MWp in Soroti City (Uganda) [6], 11.15 MWp in
time-series. These models are widely spread in technical specification
Shagaya PV plant (Kuwait) [7], 15 MWp in Nouakchott (Mauritania) [8],
manuals [13], open-source libraries pvlib [14], research literature [15],
20.05 MWp in southwestern Algeria [9], and 23.92 MWp in El Bayad
and commercial software [16]. Generally, the main inputs are the
(Algeria) [10]. In most of the cases, these analysis were performed after
in-plane global irradiance (IPOA), the nominal capacity, along with the
only 1–5 years of operation, which provide limited insights on the
PR. The latter is introduced as a product of the different installation
long-term performance. To cover this gap, this paper investigates the PV
losses, which are strongly dependent on the technology, the system
production after 12 years of operation, for the largest PV power plant

* Corresponding author. Instituto de Ingeniería Energética, Universitat Politècnica de València, Camino de Vera s/n Edificio 8E semisótano frente acceso J,
Valencia, 46022, Spain.
E-mail address: [email protected] (J. Payá).

https://doi.org/10.1016/j.rser.2022.112968
Received 23 May 2022; Received in revised form 24 September 2022; Accepted 4 October 2022
Available online 17 October 2022
1364-0321/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Nomenclature ηsoil Soiling losses

ηtemp Temperature losses
Parameters used in equations
a, b, c Exponential fit coefficients Abbreviations
A Area of PV modules AC Alternating current
CUF Capacity Utilization Factor ANN Artificial neural network
EAC AC PV energy production ANOVA Analysis of variance
IPOA Global irradiance in the plane of array a-Si Amorphous silicon
IRR Internal rate of return DC Direct current
Isc Short-circuit current IEC International Electrotechnical Commission
GSTC Reference solar irradiance at standard conditions LIL Low Irradiance Losses
IQR Interquartile range mc-Si Mono-crystalline silicon solar cell
LCOE Levelized cost of electricity ML Machine learning
MAE Mean absolute error MLR Multiple linear regression
mtry Number of predictors selected at each split of the MPPT Maximum power point tracker
regression tree pc-Si Poly-crystalline silicon solar cell
Nmod Number of PV modules from a PV array O&M Operations and maintenance
NOCT Normal operating cell temperature PV Photovoltaic
NPV Net present value RD Royal decree
nRMSE Normalized root mean squared error RF Random forest
ntree Number of trees of the random forest algorithm STC Standard test conditions
PPV,rated Rated installed PV power of the system at standard test SCADA Supervisory control and data acquisition
conditions SVM Support vector machine
Pinv,rated Rated installed power of the inverter WS Weather station
PR Performance ratio Units
PR′ Performance ratio without considering low irradiance A Ampere
losses c€ Euro cent
PRmeasured Performance ration obtained from measurements d Day
Pmax Peak power GW Nominal Gigawatt
R2 Coefficient of determination GWh Gigawatt hour
RMSE Root mean squared error h Hour
Ta Ambient temperature Hz Hertz
Tcell Cell temperature kV Kilovolt
THD Total harmonic distorsion kVA Kilovolt-ampere
VIF Variance inflation factor kW Nominal kilowatt
Voc Open circuit voltage kWh Kilowatt hour
yi Mean measured value kWp Kilowatt peak
ŷi Predicted value m Linear metre
yi Measured value m2 Square metre
YF Final yield nm Nanometre
YR Reference yield MW Nominal Megawatt
YR1 Reference yield from weather station 1 MWh Megawatt hour
YR2 Reference yield from weather station 2 MWp Megawatt peak
β0, β1, β2, β3 MLR coefficients M€ Millions of Euros
γ Maximum power temperature coefficient ◦
C Degree Celsius
ηwiring, DC DC wiring losses V Volt
ηdeg Degradation losses W Nominal Watt
ηinv Inverter efficiency Wp Watt peak
ηLIL Low irradiance losses efficiency μV Microvolt
ηmismatch Mismatch losses Ω Ohm
ηPV,stc PV efficiency under STC

design, and the climatic conditions [17]. There is abundant literature peer-reviewed comparisons with other previous model proposals [24].
regarding losses in which affect on the PR [18], especially when facil­ Another common approach to face the non-linearity of these low irra­
ities operate far from the standard test conditions (STC) [19]. However, diances is by defining two or more empirical expressions by irradiance
there are only a few publications which quantify the impact of low ranges [25]. The simplest model which provides a LIL correction is
irradiance losses (LIL) [20] on the PR by adding a correction factor. found in the in-house program PR-FACT [26]. Nevertheless, the
Irradiances below 200–400 W/m2 cause a non-negligible drop in effi­ employed continuous efficiency curve is not publicly reported and
ciency of the modules [21], leading to an overprediction in the EAC re­ cannot be implemented by other researchers. More complex low irra­
sults when operating in such range of irradiances [22]. diance models have been developed [27], but they require detailed
Generally, the modelling of LIL is addressed with logarithmic ex­ electrical characteristics of the solar cell, which hinder the replicability
pressions to estimate either directly the PV production or the module when compared to simpler models. To cover this gap, this paper de­
efficiency [23] together with several empirical models with scarce velops a replicable method to estimate the LIL, and hereby increase the

2
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Fig. 1. Aerial view of the PV plant (500 sectors: S1–S500).

accuracy of the global PR, using the irradiance exclusively as input. networks and 8 years of recorded data from a 1.44 kWp facility. The
Besides physically based models, statistical and Machine Learning complete replicability of these models is nevertheless limited due to the
(ML) models have been proposed in recent years to estimate the PV large number of climatic variables that must be measured over a long
production [28]. However, literature in this field is scarce. An artificial period of time.
neural network (ANN) was applied to predict the PR of the PV modules The present work explores the capability of simpler ML models to
with a root mean squared error (RMSE) below 0.02 [29]. The PR was predict the global PR with only two climatic variables: IPOA and the
calculated by means of a physical expression dependent on the tem­ ambient temperature (Ta). Two regression models have been employed:
perature and irradiance. S. Bandong et al. [30] developed a Support a MLR, which is the simplest algorithm, and Random Forest (RF), which
Vector Regression (SVR) and Multiple Linear Regression (MLR) using 26 is computationally simpler than ANN and well-suited for predicting
climatic variables as predictors, obtaining a RMSE of 1.5% compared stochastic PV generation reducing bias and variance [32]. The authors
with measured data. Behzad Hashemi et al. [31] reduced the number of have not found any published research on its application to estimate the
inputs to 5, obtaining a RMSE of 0.06 with Long Short-Term Memory global PR, which is a useful alternative for prediction when there is not

Fig. 2. Photograph of the PV plant.

3
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Distance
between
PV rows

2.68
(m)

2.6

2.8
NOCT

(◦ C)

46

45

46
Number
of cells

(− )

60

96

54
Maximum

voltage
system

1000

1000
600
(V)
coefficient of
Temperature

(%/◦ C)

− 0.36

− 0.33

− 0,37
Voc
coefficient of
Temperature
Fig. 3. Installed peak power grouped by manufacturer.

(%/◦ C)

0.62

0.50

0.6p
Isc
enough information available on the facility to develop a physical
model.

coefficient of
Temperature
To sum up, given the previous literature review, this article presents
the following novelties:

(%/◦ C)

− 0.43

− 0.47

− 0.45
• Provide relevant experimental data regarding the PV performance of Pmax

a large PV system (50 MW) after 12 years of operation under Med­

current
circuit
Short-

iterranean climatic conditions. 7.68

7.76
7.80
7.88
7.94
8.00
8.06
8.12
8.18
8.25
8.32
8.41
5.02
5.08

5.24
8.45
8.57
(A)

• The development and assessment of a method to estimate the daily

LIL, based on an empirical exponential expression and using the IPOA
voltage
circuit
Open-

as input.
36.28
36.36
36.40
36.48
36.54
36.60
36.66
36.72
36.78
36.84
36.90
36.90
57.60
57.80

58.40
33.60
33.70
(V)

• The development and assessment of a MLR and a RF model to esti­

mate the global PR of the utility-scale using only as inputs Ta and
current

IPOA.
power
Peak

6.91
7.03
7.09
7.21
7.30
7.39
7.48
7.56
7.65
7.72
7.79
7.89
4.58
4.66

4.90
7.90
8.04
(A)

2. Materials
voltage
power

28.72
28.86
28.90
28.98
29.04
29.10
29.16
29.22
29.28
29.38
29.50
29.50
45.90
46.15

46.90
26.60
26.80
Peak

The grid-connected PV utility-scale of the present work is located in

(V)

the east of Olmedilla de Alarcón, Spain (39.6155◦ N, 2.0905◦ W). The

plant was commissioned in October 2008 with a nominal power of 50
Efficiency

MW, a peak power of 60.103 MWp and a total land occupation of 175.3
12.26
12.50
12.63
12.87
13.06
13.24
13.43
13.61
13.80
13.98
14.17
14.35
12.21
12.50

13.37
12.40
12.95
(%)

ha. According to the Köppen climate classification, the climate of the

power plant is classified as Csa (hot-summer Mediterranean), with daily
average temperatures that vary between 0 ◦ C and 31 ◦ C, and a horizontal
(Pmax)
power

irradiation of up to around 1050 W/m2, according to the measured data

Peak

(Wp)

199
203
205
209
212
215
218
221
224
227
230
233
210
215

230
210
220

of this study.
The solar PV power plant (Figs. 1 and 2) consists of 500 independent
Area

(m2)

1.62

1.72

1.70

sectors, each with an inverter of 100 kW and an array of different PV

modules whose total peak power varies per sector from 116.5 kWp to
Technology

127.5 kWp. The peak power distribution is shown in Fig. 3 for the
different manufacturers.
mc-Si
pc-Si

pc-Si
(− )

The PV modules have a fixed 30◦ tilt angle and are oriented towards
Characteristics of the PV modules.

the south. There are three different module manufacturers with mc-Si
and pc-Si. The characteristics of the PV modules are summarized in
Manufacturer

MULTISOL
SLK60P6 L

Table 1. Each sector contains several models of the same manufacturer

and model

Scheuten

200-P5

with different rated powers.

Siliken

YL210
Yingli
(− )

All sectors use the same inverter model INGETEAM INGECON

SUN100 with a nominal power of 100 kW and an efficiency of 96%. The
201–500

rest of the electrical parameters of the inverter are shown in Table 2. The
75–100

101–200
Sectors

energy output of the inverters is expanded to a medium voltage level of

Table 1

25–74
1–24
(− )

20 kV by means of 500 transformers of 100 kVA. The voltage is finally

4
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Table 2 increased in a substation up to 132 kW before its injection into the grid.
Summary of the characteristics of the inverters. For the present study, the hourly EAC has been measured by the
Parameter Value Units monitoring system of the inverters. The inverter measurements are
transferred to the supervisory control and data acquisition (SCADA)
Maximum input voltage 900 V
Maximum input current 286 A system, which is installed in a high-performance workstation, through
MPPT voltage range 405–750 V an industrial RS232/RS485 to Ethernet converter and an IP-based
Number of inputs 4 - network. The PV plant also has two weather stations (WS) located at
Number of maximum power trackers 1 - its northern (39.6348◦ N, 2.0867◦ W) and southern (39.6151◦ N,
Nominal output power 100 kW
Nominal operating voltage 3 × 220–3x400 V
2.0938◦ W) ends (Fig. 1). The WS are equipped with a Ta sensor and a
Nominal frequency range 50/60 Hz pyranometer which measures the IPOA. Both systems provide measure­
Maximum output current 340 A ments every 5 min, and their main specifications are summarized in
European efficiency 96 % Table 3. The instrumentation of each WS is connected to an independent
Power factor 1 -
programmable logic controller system Omron CJ1M-CPU11 to condition
THD <3 %
the signals and send them to the SCADA system. The recorded data is
stored and used for real-time monitoring, alarm management, signal
processing, report generation, as well as the integration of the SCADA to
Table 3
the web. The measurement equipment is calibrated annually by the
Summary of the pyranometer specifications.
Spanish Centre for Energy, Environmental and Technological Research.
Feature Pyranometer Temperature sensor

Value Units Value Units 3. Methods

Manufacturer Delta - E+E Elektronik -
Ohm This section describes the methodology to analyze the measured data
Model LP PYRA - EE21 - and to estimate the PR and EAC.
02
The methodology is summarized in Fig. 4. The aim is to perform an
Sensitivity 10 μV/ 10 mV/
Wm− 2 ◦
C analysis of the 50 MW PV power plant and to propose a novel method
Measuring range 0 ÷ 2000 W/m2 − 40 ÷ 60 ◦
C based on climatic data that improves the PR estimations and helps reach
Operating − 0,5 ◦
C − 40 ÷ 60 ◦
C a more accurate estimation of the EAC.
temperature range
The first step in the methodology (section 3.1) is to carry out an
Impedance 33 ÷ 45 Ω - -
Spectral range 283 ÷ nm - -
exploratory data analysis of the collected data. The EAC data of each
2800 sector and the climatic data have been initially filtered to remove po­
Type of sensor - - Pt100 (tolerance class A, - tential outliers. Afterwards, the main performance parameters of the
DIN EN 60751) utility-scale have been calculated, and an exploratory data analysis of
Accuracy - ◦
C 0.2 ÷ 0.7 ◦
C
these results has been performed. Additionally, the results are compared
with other power plants in similar climatic regions.
As a second step, two approaches have helped to model the PR with
climatic data: the first method is a physical model, considering the

Fig. 4. Workflow of the methodology.

5
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Table 4 Table 6
Filtering criteria applied for the hourly measured data. Description of the losses included in the PR physical model.
Min Parameter Max Parameter Description Value Reference
2 2
20 W/m IPOA 1500 W/m ηsoil Soiling losses Optical losses due to 0.98 [38–40]
− 30 ◦ C Ta 50 ◦ C dust and particles
0 EAC 1.02 ⋅ Pinv, rated accumulated on the
PV modules’ surface
over time. The latter
fluctuate mainly with
Table 5 the frequency of the
Description of the PV performance parameters. rainfalls, the
maintenance
Parameter Description Expression Units
schedule, and dust
Reference Yield The maximum theoretical solar GPOA,d h/d type. For this facility
YR =
(YR) energy available in a specific GSTC with a 30◦ tilt, 2% of
location is defined as the ratio soiling losses has
between the total daily in-plane been assumed.
insolation (IPOA,d) and the ηdeg Degradation Power decay over 0.916 Manufacturer
reference solar irradiance at losses time in the output of 0.890 Siliken
standard conditions (GSTC = 1 the PV modules due 0.920 Scheuten
kW/m2). to different causes of Yingli
Final Yield (YF) The ratio between the EAC of the EAC,d h/d deterioration: cell
YF =
system during a certain period, PPV,rated cracks, corrosion,
in this case daily (EAC,d ), and the discoloration, glass
PV rated installed power of the breakage, etc [41].
system at standard test Values extracted from
conditions (PPV,rated ). the manufacturer
Performance The ratio between the YF and the YF % data sheet.
PR = 100⋅
Ratio (PR) YR. It can be understood as an YR ηtemp Temperature Drop in the PV ηtempi (Tcell ) [15,37] and
efficiency parameter that losses module efficiency manufacturer
measures the energy losses due to the increase of
between actual output of the the cell operating
plant with its irradiation input. temperature. The
PR allows comparing calculation of the
performance results between hourly values is
different PV systems regardless further described in
the geographical location and this section.
the installed peak power. ηLIL Low This parameter ηLIL (YR ) Manufacturer
Capacity Relationship between the YF of CUF = % irradiance gathers the nonlinear
Utilization the plant and the maximum EAC,d losses efficiency drop of the
100⋅
Factor (CUF) possible energy production, PPV,rated ⋅24 PV modules for low
defined by its installed capacity irradiance values.
in a given period. The calculation
methodology is
further explained in
product of several factors that describe the energy losses in different this section and the
value depends on the
stages of the facility. The LIL have been modelled to improve the esti­ IPOA or YR.
mated PR accuracy for low irradiances. Several regression and ML ηinv Inverter A constant value for 0.96 Manufacturer
models have been assessed. This requires examining the correlations efficiency inverter efficiency is
between the predictors (climatic data) and the predicted variable (PR) provided by the
manufacturer
and obtaining one different model per manufacturer, justified by a one-
mentioned in section
way ANOVA test. Two different models have been studied: a MLR model 2.
and a RF model with their respective k-fold cross-validations with ηmismatch Mismatch Losses due to the 0.98 [38,42]
measurements. losses interconnection of
The last step consists of predicting the total EAC of the power plant solar modules of cells
with different
through a physical model that considers the PR previously calculated
electrical properties.
from the different models, the module characteristics, and the array Following published
configuration of each sector. Finally, the production results have been literature, a typical
compared with the measurements for the different PR models. loss of 2% has been
assumed.
ηwiring,DC DC wiring Direct current losses 0.98 [38,43]
losses caused by the ohmnic
3.1. Data pre-processing resistance of the
wiring that
The performance data has been analyzed on an hourly basis for year interconnects the PV
strings with the
2020. The core of the experimental data is the EAC from the inverter, the
inverter. Due to the
Ta, and the IPOA for the tilt and azimuth angle of the PV arrays (30◦ S). To great spatial
provide stable measurements and a consistent analysis without mea­ extension of the
surement errors, the data was initially filtered according to the guide­ power plant a 2% of
lines the standard IEC 61724, as in similar PV utility-scale analysis [22]. losses has been
assumed [38].
The hourly measurements were filtered according to the ranges indi­
cated in Table 4.
In Table 4, Pinv, rated (upper filter threshold for the EAC) can be un­
derstood as the nominal power of the inverter (in this case, Pinv, rated =

6
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

100 kW). nominal operating cell temperature (NOCT in Table 1), which is defined
With respect to the data filtering, some authors employ a higher as the cell temperature obtained with Tamb 20 ◦ C and a solar irradiance
threshold for the minimum IPOA (200 W/m2) [33]. Nevertheless, to keep of 1 kW/m2. This approach is widely employed in literature and provide
as much available data as possible in this work, a less restrictive conservative loss values compared to other cell temperature models
threshold of 20 W/m2 has been considered according to the recom­ [44].
mendation of part one of the standard IEC 61724 [34]. Keeping the
NOCTi − 20
irradiance values in the range 20–200 W/m2 has enabled the develop­ Tcell,i = Ta + GPOA ⋅ (4)
800
ment of a specific characterization of the system for low irradiance
values. ηtempi = (1 − γi ⋅ (Tcelli − 25)) (5)
The typical reporting periods to assess the performance of PV utility-
scales are equal or longer than one day (e.g. annually, monthly, or daily) Since only a single complete year with measurements is available,
[35]. The hourly measurements were aggregated daily to avoid potential the absence of a cyclical component in the time series limits the use of
underestimations of module performances as reported in bibliography year-on-year and statistical methods [45], which present robust results
[36]. with time series of several years. As an alternative, instead of directly
using the degradation losses supplied by the manufacturers (Table 6),
the degradation losses were calculated with the daily PV production
3.2. PV performance parameters
balance of Eq. (1), and breaking down the PR between ηdeg,i and another
The performance of the present PV utility-scale has been evaluated factor with the rest of the losses contemplated in Table 6. The ηdeg,i daily
following the IEC 61724 standard guidelines. Their definitions and ex­ values were then averaged for the entire year, resulting in values of
pressions of the main performance parameters are described in Table 5. 0.9022 for Siliken, 0.8711 for Scheuten, and 0.8934 for Yingli. These
coefficients represent the total loss due to degradation after 12 years of
operation.
3.3. Energy production model
The efficiencies of Table 6 are typically employed to quantify the PR
[18]. In the present work a new coefficient has been added, the LIL
The hourly EAC (EAC,h ) of the entire utility-scale has been calculated
(ηLIL ), to account for the drastic drop of the module PV production at low
using Eq. (1). The daily production (EAC,d ) can be obtained aggregating
irradiances (below 200 W/m2) [20]. Additionally, this coefficient in­
the hourly production as shown in Eq. (2). cludes the drop of the inverter efficiency when the power input is low, at
500 ∑
∑ 8 low irradiance values. This helps to compensate the fact that a constant
EAC,h = PRi ⋅ηPV,stci,j ⋅Ai ⋅Nmodi,j ⋅GPOA (1) inverter efficiency had been assumed in the factor ηinv .
i=1 j=1
ηLIL has been calculated in Eq. (6) as the ratio between the PR ob­
tained from the measurements (PRmeasured ) and PR’, which does not
∑
24
EAC,d = EAC,h (2) consider the impact of the LIL.
h=1
PRmeasured
ηLIL = ′ (6)
Where: PR
Different correlations have been developed to relate the LIL with YR.
• Ai,j is the area of a PV panel of sector i in the array j, provided by the The best fitting has been achieved with the expressions indicated in (7)
manufacturer in Table 1. and (8).
• ηPV,stci,j is the PV efficiency under STC of sector i in the array j, as
ηLIL = 1 − exp (b ⋅ YR ) (7)
provided by the manufacturer in Table 1.
• Nmodi,j is the total number of modules of the sector i in the array j.
ηLIL = 1 − a⋅YcR ⋅exp (b ⋅ YR ) (8)
• GPOA is the hourly measured in-plane global irradiance.
• PRi is the performance ratio of the sector i and has been obtained Finally, the PR employed in the production model includes ηLIL , as
either with a physical quantification of the plant losses (section 3.4) shown in Eq. (9).
or with a statistical analysis (section 3.5).
(9)
′
PR = PR ⋅ηLIL

3.4. PR physical model In order to evaluate the accuracy of the developed correlations, the
daily PR obtained from the measurements has been compared with the
The physical definition of the PR is based on the determination of the calculated PR using the error metrics described in section 3.6.
losses which occur in every energy transmission or conversion stage
from IPOA to the EAC of the inverters. The losses of the transformation 3.5. PR statistical and Machine Learning models
stage have not been included since the measured data is before the grid
injection. There are no shadow losses in the PV plant. A different approach to estimate the PR is by means of statistical
According to several authors [37], the PR of each sector i can be models (e.g. MLR) and ML models (e.g. RF). Each of the developed
defined as the product of the losses indicated in Eq. (3). The term PRi
′
models employs exclusively climatic data (IPOA and Ta) as predictors.
(base PR model) refers to the PR before introducing the LIL. The data corresponds to year 2020 which is representative for the
behavior at half-life of the facility.
(3) Given the variety of equipment, a different fit is proposed for each
′
PRi = ηsoil ⋅ηdeg,i ⋅ηtemp,i ⋅ηinv ⋅ηmismatch ⋅ηwiring,DC
manufacturer to provide accurate predictions of the daily PR of the PV
Where η is the efficiency of each stage, as indicated in Table 6. Whenever utility-scale. A one-way ANOVA test has been employed to determine
the efficiency data is not available, the values have been obtained from which level of aggregation is more appropriate to define the statistical
similar facilities in literature. models. In other words, the one-way ANOVA tests helps to determine if a
The temperature losses efficiency (ηtemp ) is obtained with Eq. (5) single global model is better for all sectors, in comparison to a different
through the temperature coefficient of the PV modules (γ), defined in model for each of the module manufacturers. The null hypothesis is that
Table 1, and the PV cell temperature (Tcell ). The latter can be estimated the manufacturer groups are equal, whereas the alternative hypothesis is
with Eq. (4) using the hourly measured Ta and IPOA, as well as the that at least one of the distributions is significantly different from the

7
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

YR2
PRi = β0,i + β1,i ⋅YR1 + β2,i ⋅ + β3,i ⋅Ta (10)
YR1
The confidence of the regression parameters depends on the degree
of compliance of the MLR assumptions [50], which are evaluated
through their respective hypothesis test in section 4.2. The multi­
collinearity is quantified employing the variance inflation factor (VIF)
indicated in Eq. (11).
1
VIFi = (11)
1 − R2i,j
Fig. 5. Validation methodology of the MLR and RF models.

Where Ri,j is the correlation coefficient of the i predictor on the

others [46].
remaining explanatory variables. VIF values greater than 4 arise multi­
The confidence of the results rely on the degree the one-way ANOVA
collinearity problems [51].
assumptions are met [47]. A significance value (type-I error) of 5% has
Parallel to the MLR model, a RF model has been developed. The RF
been assumed for all the hypothesis tests. The choice is based on S.
algorithm is a non-linear ML model [52] that potentially explains the PR
Vergura [48], who indicated that medium-large PV plants present a
with a better accuracy for the range of low irradiances. Since the PR is a
larger uncertainty due to their high complexity.
continuous variable, the suggested RF model is constituted by regression
Once the PR modelling by manufacturers was justified with the one-
trees.
way ANOVA test, both MLR and RF models were trained and then tested.
The RF model was trained using the caret R package [53]. In the RF
As shown in Eq. (10), the MLR model assumes a linear relationship
algorithm, several hyperparameters need to be defined by the user. The
between the predicted variable (PR) and the predictors (YR1 from WS1,
two most relevant optimization parameters are the number of predictors
YR2 from WS2, and Ta). The stats R package has been applied. The
at each split (mtry) and the number of trees to grow for aggregation
regression parameters for each manufacturer i (β0,i , β1,i , β2,i , β3,i ) are
(ntree) [54].
estimated by ordinary least squares [49]. The mtry value is calculated by default by the algorithm as the
rounded down result of the square root of the total number of predictor

Fig. 6. Heatmaps of hourly measured IPOA, Ta and EAC in 2020.

8
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Fig. 7. Monthly average of the Ta and PV performance parameters (YR, YF, PR and CUF) in 2020. The interquartile ranges are represented by error bars and
dashed lines.

variables. In this case, since there are three predictors, the mtry value is √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑N
2. The ntree value was fixed once the increase of ntree improved the 1
i=1 (yi − ̂ y i )2
(13)
N
RMSE in less than 1%. The previous criteria is commonly used by re­ nRMSE =
yi
searchers [55] and employed in RF models applied to PV applications
[56]. 1 ∑N
MAE = |y − ŷi | (14)
To assess the accuracy and robustness of the MLR and RF models, the N i=1 i
validation was performed as depicted in Fig. 5. The pre-processed ∑N
dataset obtained in section 3.1 was randomly split using an 80:20 y i )2
(yi − ̂
R2 = 1 − ∑i=1 (15)
ratio to create a train dataset and test dataset. This partition is done to N
i=1 (yi − yi )
2

perform an external validation of the models with unseen data from the
train set [57]. Then, a k-fold cross validation was conducted with the Where yi , ŷi , yi are the measured, predicted and the mean measured
train set formed by the 80% of the original dataset to obtain the opti­ values, respectively, and N the number of samples of the dataset.
mized regressors coefficients of the MLR model and build the RF model.
For this work a k value of 10 was considered, which is commonly used in 4. Results
literature [58]. Finally, external model validations with the remaining
20% of the original dataset were employed. Following the methodology described in Fig. 4, this section presents
the performance analysis of the PV utility-scale, the results and valida­
3.6. Model deviation tion of the PR models, and finally in the impact on the PV production.

The accuracy of the three models for the daily PR and the derived PV
4.1. Performance results of the PV facility
production has been compared. For the MLR and RF models, the trained
model performance is evaluated on the test set using as error metrics the
The raw measured data consists of 8727 hourly measurements of
root mean squared error (RMSE), the normalized root mean squared
three variables: IPOA, Ta and EAC (in 500 sectors), during 364 days of
error (nRMSE), the mean absolute error (MAE) and the coefficient of
2020. After applying the data filtering explained in section 3.1, the
determination (R2) as defined in equations (12)–(15) [59]:
resulting dataset was reduced by 59.33% of the original hours. The
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑N minimum irradiation threshold was the main effective filter since
i=1 (yi − ̂ y i )2
RMSE = (12) 49.74% of the original data was removed due to nighttime hours, and
N 8.30% during the sunrise and sunset hours. Additionally, there were 10
days (1.29% of the raw data) when the PV production stopped (non-
productive hours in Fig. 6. Stops on individual days are mainly due to
inverter failures caused by high temperatures, blown fuses and powered

Table 7
Statistical results of the global daily performance parameters in 2020.
Parameter Units Minimum 1st Quartile Median Mean 3rd Quartile Maximum Standard deviation Skewnessa

YR h/d 0.59 4.10 5.90 5.44 7.30 8.03 2.09 − 0.68

YF h/d 0.43 3.31 4.74 4.28 5.63 6.33 1.59 − 0.85
PR % 69.11% 75.93% 79.15% 79.24% 82.20% 93.44% 4.08% 0.23
CUF % 1.81% 13.81% 19.77% 17.88% 23.52% 26.43% 6.64% − 0.85
a
Dimensionless Value.

9
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Fig. 8. Boxplots of the average YF (a), PR (b) and CUF (c) of each sector and manufacturer.

surge protection devices, which need staff intervention before starting- 7.35 h and of 3.63 h of irradiance equivalent to 1 kW/m2, respectively.
up again. Fig. 6 shows the hourly values of the total measured EAC of Regarding the YF, the yearly average was 5.44 h/d. The minimum
the utility-scale, IPOA, and Ta after the data filtering. was 2.98 h/d in December and the maximum 5.49 h/d in August.
The annual IPOA was 1975.52 kWh/m2 and the EAC of the complete However, the maximum daily values, up to 6.33 h/d (in March), were
utility-scale in 2020 was 91.32 GWh, with a monthly average of 7.61 registered during the Spring season, where the lower Ta compared to
GWh. summer provided a higher efficiency. There are also noticeable differ­
At the top of Fig. 7 the average Ta for each month is shown. The ences in the yearly average YF for the different module manufacturers.
average daily Ta during the year is 16.61 ◦ C, fluctuating from a mini­ The best performance is obtained by Siliken with 4.32 h/d, followed by
mum average of 7.43 ◦ C in January, and a maximum average of 28.28 ◦ C Yingli with 4.26 and, lastly, Scheuten with 4.19 h/d. The latter also
in July. provides the lowest performances within some sectors, reaching lower
The monthly variation of the PV performance parameters is also values than 4 h/d, as shown in Fig. 8.
presented in Fig. 7 and Table 7. According to section 3.2, the YR reached The PR ranges between in 74.97% in August and 83.46% in
its maximum in August and its minimum in December with a value of February, with an annual average of 79.24%. The lowest PRr are

Table 8
Comparison of the PV performance in different PV facilities with hot-summer Mediterranean climate (Csa).
Location Commisioning Monitoring Module Peak power YF (h/d) min/ PR (%) min/ CUF (%) min/ Ref.
Year period type (kWp) mean/max mean/max mean/max

Olmedilla de Alarcón, 2008 2020 mc-Si/pc- 60,103.4 0.43/4.28/6.33 69.11/79.24/ 1.81/17.83/26.43 Present
Spain Si 93.44 study
Ar Ramtha, Jordan - 2017–2018 pc-Si 5000 2.80/4.60/5.80 74.00/80.00/ 11.00/18.10/ [62]
90.00 23.00
Albacete, Spain 2010 2010–2013 pc-Si 4600 2.20/-/4.37 59.46/-/78.84 9.16/-/18.22 [63]
Mugla, Turkey - 2008 pc-Si 2730 2.53/4.77/6.65 -/72/- - [64]
Albacete, Spain 2008 2012–2016 pc-Si 2700 4.71/-/7.92 78.93/-/84.86 19.64/-/33.02 [63]
Şahinkaya, Turkey 2016 2017 pc-Si 2130.7 2.32/4.53/6.28 73.92/81.15/ 9.65/18.86/26.16 [37]
91.78
Albacete, Spain 2007 2012–2016 pc-Si 1400 7.92/-/2.2 83.56/-/85.60 20.14/-/29.87 [63]
Albacete, Spain 2007 2010–2013 mc-Si 1300 3.10/-/5.69 80.36/-/85.66 12.91/-/18.66 [63]
Albacete, Spain 2008 2010–2013 mc-Si 1300 3.50/-/6.69 64.91/-/68.83 14.65/-/27.29 [63]
Monteroni, Italy 2011 2012–2015 mc-Si 960 1.70/3.80/6.20 75.00/84.4/ 6.90/15.60/25.60 [11]
94.00
Ciudad Real, Spain 2013 2013–2016 pc-Si 370 4.29/-/4.63 80.39/-/81.39 17.86/-/19.30 [63]
Sitia, Crete, Greece 2002 2007 pc-Si 171.36 1.96/3.66/5.07 58.00/67.36/ -/15.26/- [15]
73.00
Manisa, Turkey 2018 1 year mc-Si 30 1.53/4.16/6.09 81.22/83.61/ 6.38/17.35/25.39 [65]
86.15
Bouzareaha, Algeria 2004 2016–2018 pc-Si 9.5 -/3.37/- -/70.00/- - [66]
Tangiers, Morocco - 2015 pc-Si 5 1.96/4.45/6.42 58.00/79.00/ 6.55/14.84/21.42 [67]
98.00
Chania, Crete, Greece - 2010–2012 a-Si/mc-Si 2.18 1.83/-/6.55 80.40/-/95.40 - [68]
Tangiers, Morocco - 2016 pc-Si 2 3.38/4.72/5.90 71.23/77.24/ 10.83/11.76/ [69]
84.00 12.78
Los Angeles, United - - - 0 -/4.22/- -/72.10/- - [70]
States
Casablanca, Morocco - - - 0 -/4.29/- -/71.90/- - [70]

10
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Table 9
Parameters of the economic analysis.
Variable Value Units Reference

Total investment cost 384 M€ Present

study
Averaged yearly EAC measured for 91,967 MWh Present
the life cycle study
Fixed electricity price (2008–2014) 22.976 c€/kWh [71]
Averaged electricity market price 6.186 c€/kWh [75]
(2014-Present)
Specific remuneration for the 31.754 c€/kWh [76]
operation (2014-Present)
Specific remuneration for return on 244,850 €/MW⋅year [76]
the investment (2014-Present)
Average degradation rate Siliken: %/year Present
0.816 study
Scheuten:
1.074
Yingli: 0.888
O&M cost 11.6 €/kWp⋅year [77]
Annual Spanish inflation rate 1.062 % [78]
(averaged between 2008 and
2020)
Annual discount rate 7.090 % [79]
Life cycle of the facility 25 years Present
study

The NPV, IRR and payback period are 93.02 M€, 9.19%, and 17.61 years, and the
LCOE is 0.359 €/kWh.
Fig. 9. Relationship between LIL (ηLIL ) and the YR.
obtained in summer due to the higher temperatures. There is a clear
correlation with the temperature. The global PR was above 75% for on two different Spanish legislative frameworks during its operation.
84.2% of the days with measurements, proving that the system has been The first period, from its commissioning in 2008 until July 2014, fol­
working correctly in global terms. There are significant fluctuations in lowed RD 661/2007 [71], with a fixed price. The second period, which
the PR when comparing the different manufacturers: the Siliken and is still in force, follows RD 413/2014 [72]. The remuneration calcula­
Yingli (pc-Si) sectors provide an annual average of 79.61% and 79.01%, tions for this period are described in detail for other plants in literature
respectively, while the Scheuten (mc-Si) sectors yield 77.09%. [73].
The PR of the mc-Si sectors is on average around 2% lower than the The net present value (NPV), internal rate of return (IRR), payback
pc-Si sectors. Considering the similarities among manufacturer charac­ period and the levelized cost of electricity (LCOE), were estimated ac­
teristics in the STC efficiencies, the performance difference is mainly cording to N. Bansal et al. [74], considering the initial investment cost,
caused by a greater drop in efficiency when the temperature increases. the O&M costs, the cashflows generated by the energy selling, the
This issue can be observed in Table 1, since the temperature coefficient annual degradation rate of the modules, the inflation rate and the dis­
in the mc-Si modules is higher than for the pc-Si modules. The same count rate summarized in Table 9.
phenomenon was also found in similar climate conditions, both in
northern Algeria [60] and in Morocco [61]. 4.2. PR modelling results
The monthly average CUF ranges between 12.44% in December and
22.93% in August, with an annual average of 17.88%. The CUF Fig. 9 shows the deviations between the modelled daily PR consid­
dispersion decreases during the months with more sunny hours and ering the physical losses and the measured daily PR, as calculated with
stable weather. The annual average CUF among module manufacturers Eq. (6). While for regular days with daily YF higher than 3 h/d the de­
differs slightly: 17.99% for Siliken, 17.44% for Scheuten and 17.74% for viations fluctuate around 1, which means that no significant corrections
Yingli, following a similar distribution scheme as the PR. are required, there is a clear drop for daily YR values below 2 h/d.
According to the assumptions described in section 3.4, the estimated These deviations were modelled with two nonlinear exponential fits,
degradation losses are generally 1–2% higher than the rates provided by whose coefficients and error metrics are given in Table 10. The
the manufacturer datasheets. After 12 years of operation, the average employment of an exponential fit allows reducing selectively these dif­
degradation loss for Siliken modules is 9.79%, 12.89% for Scheuten, and ferences only for low YR values. A linear regression fit would tend to
10.66% for Yingli, and their respective averaged yearly degradation overestimate the LIL. The exponential 1 fit, despite its simplicity, tends
rates are 0.816%/year, 1.074%/year, and 0.888%/year. Consequently, to excessively reduce the PRs with low YR values s. The exponential fit 2
the averaged modules efficiency at STC drops to 12.13% for Siliken, presents a more moderate fit and reduces the error compared with the
10.85% for Scheuten, and 11.02% for Yingli. The highest degradation is exponential 1 for YR values around 2 h/d. The exponential 2 was
suffered by the mc-Si technology. consequently selected for the comparison with the PR models.
As indicated in Table 8, the performance of the present utility-scale is Adding the correction of the exponential fit 2 in Eq. (9) clearly im­
comparable with other PV power plants reported in scientific literature proves the results, as may be inferred by comparing Fig. 10a and b.
under Csa Mediterranean climate. However, due to the long operating There are significant overpredictions with the base PR model (up to 15%
period and consequent degradation losses, the average YF, PR, and CUF of relative error) which are mitigated when ηLIL is introduced. With the
are slightly lower than in the other plants, where the performance was exponential fit 2, the nRMSE decreases by 48.22%.
measured a few years after their commissioning. Among the registered The compliance with the one-way ANOVA assumptions has been
PV utility-scales the CUF is generally greater for bigger installed powers, verified prior to its application. To meet the normality of the annual
and mc-Si technologies on average provide better PR results. average PR, the inferior outliers below the limit defined by Tukey (Q1-
Besides the energy performance assessment, an economic analysis 1.5⋅IQR = 0.757) were filtered applying the same method as in other PR
has also been conducted. The remuneration of the facility has depended analyses [46]. As a result, 22 facilities were omitted and the remaining

11
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Table 10
LIL exponential model coefficients and error metrics.
Model Expression a b c RMSE MAE R2

Exponential 1 ηLIL = 1 − exp (b ⋅YR ) - − 1.688 - 0.047 0.033 0.514

Exponential 2 ηLIL = 1 − a⋅YcR ⋅exp (b ⋅YR ) 0.539 − 1.067 0.273 0.043 0.032 0.585

Fig. 10. PR results validated with the base (a) and exponential (b) models.

PR followed a normal distribution with a mean of 0.795 and a standard

deviation 0.014, as verified with the Kolmogorov-Smirnov test with a
p-value of 0.239.
The normality of each PR distribution was tested with the
Kolmogorov-Smirnov test for the Siliken, Scheuten and Yingli modules,
providing p-values of 0.724, 0.516 and 0.0806, respectively. All the p-
values are consequently higher than the type-I error threshold (0.05). A
p-value of 0.239 was obtained. However, the homoscedasticity among
the three PR samples was not met applying the Barlett’s test. To reduce
the heterogeneity of variances, the Welch’s correction factor [80] was
included in the one-way ANOVA test [81], which provided a p-value of
2.7e-14. Thus, the null hypothesis was rejected, which led to develop
three independent PR statistical models.
For the MLR and RF model, the predictors were selected considering
the global Pearson correlation coefficients between the climatic data and
the global PR of the plant (Fig. 11) and the multicollinearity among
predictors measured with the VIF. There is a low negative correlation
with the measured irradiance from the two WS and a moderate corre­
lation with Ta. This reveals that the higher PR is generally reached in
cold days. The VIF values between Ta and the irradiances are below 1.6,
presenting reduced multicollinearity. However, there is high multi­
Fig. 11. Correlation matrix of the PR and the climatic variables as predictors. collinearity between YR1 and YR2 with a VIF value of 50.25. The selected
predictors are Ta, YR1, and the ratio YR2/YR1 to consider weather fluc­
tuations and have both irradiance predictors uncorrelated. This ratio
presents a correlation coefficient with PR and Ta of 0.51 and − 0.11,

Table 11
Coefficients and p-values of the MLR models to estimate the PR of the three manufacturers.
Manufaturer MLR coefficients p-values

β0 β1 β2 β3 β0 β1 β2 β3

Siliken 0.403 0.003 0.435 − 0.004 <2e-16 1e-04 <2e-16 <2e-16

Scheuten 0.687 0.008 0.111 − 0.005 <2e-16 2–14 <2e-16 1e-04
Yingli 0.933 0.004 − 0.114 − 0.004 <2e-16 3e-07 <2e-16 8e-06

12
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Fig. 12. Validation of the PR results obtained with the MLR (a) and RF (b) models compared with the measured PR.

Table 12
Error metrics of the PR for the MLR and RF models.
Manufacturer MLR model RF model

RMSE nRMSE MAE R2 RMSE nRMSE MAE R2

Siliken 0.026 0.033 0.019 0.689 0.024 0.030 0.017 0.741

Scheuten 0.028 0.037 0.020 0.547 0.023 0.029 0.016 0.701
Yingli 0.025 0.031 0.018 0.497 0.019 0.024 0.014 0.683

respectively. 3%. Fig. 12a, shows the global PR obtained by weighting the estimations
The MLR expression results from the linear combination of the var­ of each regression with the number of sectors associated for each
iables shown in Eq. (10), and the fitted coefficients gathered in Table 11. manufacturer, providing an accuracy of 87.85%. The accuracy repre­
The intercept (β0 ) value is the most influential coefficient for every sents the number of estimations with a relative error lower than 5%. The
manufacturer, followed by the fluctuations of the measured irradiances global nRMSE is 0.0324, which is close to the value obtained for the
between both WSs. The negative coefficients of β3 explain the reduction Siliken sectors which are the most frequent sectors. The global nRMSE
of the PR for increasing Ta values. The regression fit provides p-values improves by 41.54% and 15.29% compared with the base model and the
below 0.05 for all the manufacturers (see Table 11). This supports the exponential model, respectively. However, there is a trend to over­
null hypothesis that the independent variables do not affect significantly estimate the lower PR values due to the linearity of the model similar to
the dependent variable. the base model.
Another conclusion is that including nonlinear combinations of the For the RF regression models, the hyperparameters were first tuned
predictors have any power in explaining the PR, as verified with the to provide the lower RMSE. The Siliken sectors do not require more than
Ramsey’s RESET test (p-value of 0.5671). The residuals of each regres­ 50 trees, and the other manufacturers require up to 100 trees to provide
sion fit follow a normal distribution according to the Kolmorov-Smirnov stability in RMSE. The nRMSE was below 3%. This is a major
test with an averaged p-value of 0.0998. The residuals are uncorrelated, improvement compared to the MLR, as shown in Fig. 12b, where the
given the averaged Durbin-Watson D statistic of 1.839 (p-value of 0.198) number of outliers has been reduced, especially for low PR values. The
[82]. greater deviations are found in the extreme PR values. Nevertheless, the
Fig. 12a compares the measured and the predicted PR of the three global accuracy rises up to 99.44%. Weighting all the sectors the global
MLR models. There is a higher overprediction for lower PR values, as PR yields a nRMSE of 0.013, shown in Table 13. This represents a
happened with the base PR model and this is not explained with linear reduction of 77.04% with respect to the base model and is similar or
relationships. The nRMSE represents for the three regressions around lower than the SVR model found in literature [30].

Table 13
General error metrics of the PR and EAC models.
Model PR EAC
2
RMSE (− ) nRMSE (− ) MAE (− ) R (− ) RMSE nRMSE (− ) MAE R2 (− ) Annual error Annual error (YR < 3 h/d) (%)
(MWh) (MWh) (%)

Base case 0.044 0.055 0.028 0.144 8.268 0.032 6.334 0.993 1.814 7.280
Exponential 1 0.034 0.043 0.023 0.432 7.973 0.031 5.968 0.993 − 1.141 − 1.006
Exponential 2 0.029 0.037 0.021 0.517 7.924 0.031 5.870 0.994 − 0.706 − 0.474
MLR 0.026 0.032 0.018 0.617 6.301 0.024 4.978 0.996 − 0.116 − 0.787
RF 0.010 0.013 0.007 0.945 2.626 0.010 1.973 0.999 0.099 2e-04

13
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

The manufacturer’s error metrics of both MLR and RF models are

shown in Table 12.

4.3. PV production modelling results

The global EAC of the PV power plant has been obtained for the base
case, with the theoretical PR values, and compared with the production
obtained with the three PR models (the exponential fit, the MLR and the
RF model).
All the energy losses estimated in the energy balance are quantified
in Fig. 13 by means of a Sankey diagram. The annual in-plane global
irradiance does not consider the non-productive days since they have
been filtered. LIL represent 0.78% of the annual array nominal energy at
STC and the degradation losses have the biggest weight due to the long
operating time of the utility-scale. The estimated annual EAC with the
physical model differs by − 0.71% compared with the measurements.
The annual production would rise 2.61% (up to 93.03 GWh) if the 10
Fig. 13. Sankey diagram of the annual losses in the PV utility-scale according non-productive days were considered.
to the PR physical model. The validation results for each model are shown in Fig. 14. The base
model systematically overpredicts the production when the daily

Fig. 14. Validation of the global PV production obtained considering the PR for the base (a), exponential (b), MLR (c) and RF (d) models.

14
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

yield the greatest weight in the energy balance, representing a global

energy loss of 13% of the global energy at STC.
After the performance analysis, a more in-depth study has been
performed to reduce the outliers in the predictions in low irradiance
days. A physical model was developed, as the product of the different
losses of the PV system, including the LIL through irradiance measure­
ments and an exponential fit. The results improved the nRMSE by 1.9%
compared with the conventional model, increasing the R2 from 0.144 to
0.553 for low irradiances.
A second approach has been applied using two statistical methods
using only Ta and YR as predictors. The RF model has provided the best
performance with a nRMSE of 1.27%. These results indicate a better
performance than SVR models found in literature, which require
significantly more predictors. In contrast, the MLR model has reduced
the nRMSE by 2.30% with an accuracy of 87.85%.
The inclusion of the improvements in the PR and in the PV daily
production model has provided improvements in nRMSE of 0.11%,
0.76%, and 2.19% for the exponential, MLR, and the RF models,
respectively. These improvements are significantly greater for the low
irradiance days, providing reductions in the nRMSE up to 7.27%, 7.75%
Fig. 15. nRMSE of the estimated daily production of the PV plant for days with and 11.07% for the exponential, MLR and RF models, respectively. In
irradiances equal or lower than the irradiance shown on the x-axis. any case, both statistical models provided a better PR accuracy than the
physical model, and are recommended to forecast the PR whenever
irradiance is low; however, the three PR models significantly reduce the measured data is available. Moreover, they constitute an alternative to
number of outliers for low daily irradiances. Compared with the base model and predict the PR when there is scarce technical data of the
case, the total RMSE is reduced up to 68.24% with the RF, while the plant.
exponential and the MLR models provide moderate improvements in As future work, the degradation of the PV modules will be studied in
RMSE of 4.16% and 23.79%, respectively. The relative error of the more detail by analyzing the performance after more years of operation.
estimated annual production was reduced up to 0.09% with the RF
model. Filtering only the days of the year in which the YR is below 3 h/d, Credit author statement
an improvement in the annual production error of close to 7% is
observed for the three previous models, and the RF provides the best Enrique Fuster-Palop: Methodology, Software, Formal analysis,
results. Table 13 provides the general error metrics in the estimation of Resources, Validation, Investigation, Writing - Original Draft, Visuali­
the production. zation, Carlos Vargas-Salgado: Conceptualization, Writing - Review &
Fig. 15 represents the performance of each model, by means of the Editing, Juan Carlos Ferri-Revert: Resources, Investigation, Jorge
nRMSE of the daily EAC. The nRMSE is clearly bigger in low irradiance Payá: Conceptualization, Writing - Review & Editing, Supervision,
days. Daily irradiation measurements lower than 2 kWh present a Project administration, Valencia, May 23, 2022.
nRMSE of 13.14%, which is significantly higher than the nRMSE of
3.21% which is obtained for the full irradiance range. By incorporating
the PR the LIL factor, the nRSME is reduced in the full range up to 5.87%. Declaration of competing interest
This value drops to 2.44% with the MLR model and 1.01% with the RF
model. These three models reduce the nRMSE for low irradiances by The authors declare that they have no known competing financial
more than half compared with the base model. Nevertheless, the RF interests or personal relationships that could have appeared to influence
model provides the lowest fluctuations and confident production pre­ the work reported in this paper.
dictions for the complete range of daily irradiances.
Data availability
5. Conclusions
The data that has been used is confidential.
The present work involves the analysis of a 50 MW PV utility-scale
plant in Olmedilla de Alarcón (Spain) after 12 years of operation Acknowledgements
under Mediterranean climatic conditions. The experimental campaign
consists of a monitoring period of one year with measurements of cli­ The authors gratefully acknowledge the operation & maintenance
matic data and EAC from the inverters. Using this data, the main PV staff of the PV Power Plant in Olmedilla de Alarcón for providing the
performance parameters have been obtained. measured data of the solar PV power plant.
The annual average and the minimum and maximum monthly
average registered for the YR, YF, PR and CUF respectively are: 5.44 h/d, References
4.28 h/d, 79.24%, and 19.77%. These results provided a clear season­
[1] David Feldman RM, Ramasamy Vignesh, Fu Ran, Ramdas Ashwin, Jal Desai, “U.S.
ality, with lower system efficiencies during the summer due to the high Solar photovoltaic system and energy storage cost benchmark. 2020. p. 2021.
temperatures. The performance is slightly lower than other PV power [2] Mitrašinović AM. Photovoltaics advancements for transition from renewable to
plants in the Mediterranean, although with more years of operation. clean energy. Energy 2021;237:121510. https://doi.org/10.1016/J.
ENERGY.2021.121510. Dec.
Nevertheless, the PR is over 80% for almost 42% of the measured days,
[3] IRENA. Renewable Capacity Statistics 2021:2021.
proving a correct performance. Furthermore, the pc-Si sectors provided [4] Sundaram S, Babu JSC. Performance evaluation and validation of 5 MWp grid
PR values around 2% greater than the mc-Si sectors, mainly due to the connected solar photovoltaic plant in South India. Energy Convers Manag Aug.
higher PV temperature and degradation losses of the mc-Si sectors. The 2015;100:429–39. https://doi.org/10.1016/J.ENCONMAN.2015.04.069.
[5] Bansal N, Jaiswal SP, Singh G. Long term performance assessment and loss analysis
estimated degradation losses of the modules are approximately 2–3% of 9 MW grid tied PV plant in India,” Mater. Today Proc., Feb. 2022. https://doi.
lower than according to the manufacturer data. The degradation losses org/10.1016/J.MATPR.2022.01.263.

15
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

[6] Oloya IT, Gutu TJL, Adaramola MS. Techno-economic assessment of 10 MW [32] Ahmad MW, Mourshed M, Rezgui Y. Tree-based ensemble methods for predicting
centralised grid-tied solar photovoltaic system in Uganda. Case Stud Therm Eng PV power generation and their comparison with support vector regression. Energy
2021;25:100928. https://doi.org/10.1016/J.CSITE.2021.100928. Jun. Dec. 2018;164:465–74. https://doi.org/10.1016/J.ENERGY.2018.08.207.
[7] AL-Rasheedi M, Gueymard CA, Al-Khayat M, Ismail A, Lee JA, Al-Duaj H. [33] Kunaifi K, Reinders A, Lindig S, Jaeger M, Moser D. Operational performance and
Performance evaluation of a utility-scale dual-technology photovoltaic power plant degradation of PV systems consisting of six technologies in three climates. Appl Sci
at the Shagaya Renewable Energy Park in Kuwait. Renew Sustain Energy Rev 2020; 2020;10. https://doi.org/10.3390/APP10165412. Page 5412, vol. 10, no. 16, p.
133:110139. https://doi.org/10.1016/J.RSER.2020.110139. Nov. 5412, Aug. 2020.
[8] Elhadj Sidi CEB, Ndiaye ML, El Bah M, Mbodji A, Ndiaye A, Ndiaye PA. [34] International Electrotechnical Committee. IEC 61724 photovoltaic system
Performance analysis of the first large-scale (15 MWp) grid-connected photovoltaic performance monitoring - guidelines for measurement, data exchange and analysis.
plant in Mauritania. Energy Convers Manag Jul. 2016;119:411–21. https://doi. 1998 [Online]. Available: https://global.ihs.com/doc_detail.cfm?document_
org/10.1016/J.ENCONMAN.2016.04.070. name=IEC 61724& item_s_key=00290218. [Accessed 27 February 2022].
[9] Bentouba S, Bourouis M, Zioui N, Pirashanthan A, Velauthapillai D. Performance Accessed.
assessment of a 20 MW photovoltaic power plant in a hot climate using real data [35] Dhimish M. Thermal impact on the performance ratio of photovoltaic systems: a
and simulation tools. Energy Rep Nov. 2021;7:7297–314. https://doi.org/ case study of 8000 photovoltaic installations. Case Stud Therm Eng 2020;21:
10.1016/J.EGYR.2021.10.082. 100693. https://doi.org/10.1016/J.CSITE.2020.100693. Oct.
[10] Dahmoun MEH, Bekkouche B, Sudhakar K, Guezgouz M, Chenafi A, Chaouch A. [36] Ransome S, Funtan P. Why hourly averaged measurement data is insufficient to
Performance evaluation and analysis of grid-tied large scale PV plant in Algeria. model PV system performance accurately. In: 20th European photovoltaic solar
Energy Sustain. Dev. Apr. 2021;61:181–95. https://doi.org/10.1016/J. energy conference; 2005. p. 2752–5.
ESD.2021.02.004. [37] Cubukcu M, Gumus H. Performance analysis of a grid-connected photovoltaic plant
[11] Malvoni M, Leggieri A, Maggiotto G, Congedo PM, De Giorgi MG. Long term in eastern Turkey. Sustain Energy Technol Assessments 2020;39:100724. https://
performance, losses and efficiency analysis of a 960 kWP photovoltaic system in doi.org/10.1016/J.SETA.2020.100724. Jun.
the Mediterranean climate. Energy Convers Manag Aug. 2017;145:169–81. [38] Roberts JJ, Mendiburu Zevallos AA, Cassula AM. Assessment of photovoltaic
https://doi.org/10.1016/J.ENCONMAN.2017.04.075. performance models for system simulation. Renew Sustain Energy Rev May 2017;
[12] Copper JK, Sproul AB, Jarnason S. Photovoltaic (PV) performance modelling in the 72:1104–23. https://doi.org/10.1016/J.RSER.2016.10.022.
absence of onsite measured plane of array irradiance (POA) and module [39] Pérez NS, Alonso-Montesinos J, Batlles FJ. Estimation of soiling losses from an
temperature. Renew Energy Feb. 2016;86:760–9. https://doi.org/10.1016/J. experimental photovoltaic plant using artificial intelligence techniques. Appl Sci
RENENE.2015.09.005. 2021;11. https://doi.org/10.3390/APP11041516. Page 1516, vol. 11, no. 4, p.
[13] IDAE. “Pliego de Condiciones técnicas de Instalaciones conectadas a red. 2011. 1516, Feb. 2021.
[14] Holmgren WF, Hansen CW, Mikofski MA. Pvlib python: a python package for [40] Polo J, et al. Characterization of PV soiling losses in urban mediterranean
modeling solar energy systems. J. Open Source Softw. Sep. 2018;3(29):884. environment. 2019. https://doi.org/10.18086/swc.2019.15.03.
https://doi.org/10.21105/JOSS.00884. [41] Kim J, Rabelo M, Padi SP, Yousuf H, Cho EC, Yi J. A review of the degradation of
[15] Kymakis E, Kalykakis S, Papazoglou TM. Performance analysis of a grid connected photovoltaic modules for life expectancy. Energies 2021;14. https://doi.org/
photovoltaic park on the island of Crete. Energy Convers Manag Mar. 2009;50(3): 10.3390/EN14144278. Page 4278, vol. 14, no. 14, p. 4278, Jul. 2021.
433–8. https://doi.org/10.1016/J.ENCONMAN.2008.12.009. [42] Shiva Kumar B, Sudhakar K. Performance evaluation of 10 MW grid connected
[16] Gilman P, Dobos A, Diorio N, Freeman J, Janzou S, Ryberg D. SAM photovoltaic solar photovoltaic power plant in India. Energy Rep Nov. 2015;1:184–92. https://
model technical reference update. 2016. doi.org/10.1016/J.EGYR.2015.10.001.
[17] Ozden T. A countrywide analysis of 27 solar power plants installed at different [43] Marion B, et al. Performance parameters for grid-connected PV systems. Conf Rec
climates. Sci. Reports 2022;12(1):1–11. https://doi.org/10.1038/s41598-021- IEEE Photovolt Spec Conf 2005:1601–6. https://doi.org/10.1109/
04551-7. 2022 121. PVSC.2005.1488451.
[18] Khalid AM, Mitra I, Warmuth W, Schacht V. Performance ratio – crucial parameter [44] Santiago I, Trillo-Montero D, Moreno-Garcia IM, Pallarés-López V, Luna-
for grid connected PV plants. Renew Sustain Energy Rev Nov. 2016;65:1139–58. Rodríguez JJ. Modeling of photovoltaic cell temperature losses: a review and a
https://doi.org/10.1016/J.RSER.2016.07.066. practice case in South Spain. Renew Sustain Energy Rev Jul. 2018;90:70–89.
[19] Quesada B, Sánchez C, Cañada J, Royo R, Payá J. Experimental results and https://doi.org/10.1016/J.RSER.2018.03.054.
simulation with TRNSYS of a 7.2kWp grid-connected photovoltaic system. Appl [45] Malvoni M, Kumar NM, Chopra SS, Hatziargyriou N. Performance and degradation
Energy May 2011;88(5):1772–83. https://doi.org/10.1016/j. assessment of large-scale grid-connected solar photovoltaic power plant in tropical
apenergy.2010.12.011. semi-arid environment of India. Sol Energy Jun. 2020;203:101–13. https://doi.
[20] Mavromatakis F, Vignola F, Marion B. Low irradiance losses of photovoltaic org/10.1016/J.SOLENER.2020.04.011.
modules. Sol Energy Nov. 2017;157:496–506. https://doi.org/10.1016/J. [46] Taylor J, Leloux J, Hall LMH, Everard AM, Briggs J, Buckley A. Performance of
SOLENER.2017.08.062. distributed PV in the UK: a statistical analysis of over 7000 systems. Conf. 31st Eur.
[21] Soler-Castillo Y, Rimada JC, Hernández L, Martínez-Criado G. Modelling of the Photovolt. Sol. Energy Conf. Exhib. 2015. https://doi.org/10.13140/
efficiency of the photovoltaic modules: grid-connected plants to the Cuban national RG.2.1.2019.6568.
electrical system. Sol Energy Jul. 2021;223:150–7. https://doi.org/10.1016/J. [47] Seltman HJ. Experimental design and analysis. 2018.
SOLENER.2021.05.052. [48] Vergura S. A statistical tool to detect and locate abnormal operating conditions in
[22] Lindig S, Louwen A, Moser D, Topic M. Outdoor PV system monitoring—input data photovoltaic systems. Sustain Times 2018;10. https://doi.org/10.3390/
quality, data imputation and filtering approaches. Energies 2020;13. https://doi. SU10030608. Page 608, vol. 10, no. 3, p. 608, Feb. 2018.
org/10.3390/EN13195099. Page 5099, vol. 13, no. 19, p. 5099, Sep. 2020. [49] Kabacoff RI. R in Action SECOND EDITION Data analysis and graphics with R.
[23] Huld T, Gottschalg R, Beyer HG, Topič M. Mapping the performance of PV 2015.
modules, effects of module type and data averaging. Sol Energy Feb. 2010;84(2): [50] Flatt C, Jacobs RL. Principle assumptions of regression analysis: testing,
324–38. https://doi.org/10.1016/J.SOLENER.2009.12.002. techniques, and statistical reporting of imperfect data sets. vol. 21, no. 4, pp.
[24] de la Parra I, Muñoz M, Lorenzo E, García M, Marcos J, Martínez-Moreno F. PV 484–502, https://doi.org/10.1177/1523422319869915; 2019. 10.1177/
performance modelling: a review in the light of quality assurance for large PV 1523422319869915.
plants. Renew Sustain Energy Rev Oct. 2017;78:780–97. https://doi.org/10.1016/ [51] Fávero LP, Belfiore P. Simple and multiple regression models. Data Sci. Bus. Decis.
J.RSER.2017.04.080. Mak. Jan. 2019:443–538. https://doi.org/10.1016/B978-0-12-811216-8.00013-6.
[25] Roumpakias E, Stamatelos A. Comparative performance analysis of grid-connected [52] Breiman L. Random forests. Mach Learn 2001;451. https://doi.org/10.1023/A:
photovoltaic system by use of existing performance models. Energy Convers Manag 1010933404324. 45, no. 1, pp. 5–32, Oct. 2001.
Oct. 2017;150:14–25. https://doi.org/10.1016/J.ENCONMAN.2017.08.001. [53] Kuhn M. Building predictive models in R using the caret package. J Stat Software
[26] Busch L, Schäfer T, Song W, Mack M, Egler M. PV system energy yield calculation Nov. 2008;28(5):1–26. https://doi.org/10.18637/JSS.V028.I05.
program PR-FACT,” 28th. Eur. Photovolt. Sol. Energy Conf. Exhib. Nov. 2013: [54] Probst P, Wright MN, Boulesteix AL. Hyperparameters and tuning strategies for
3699–708. https://doi.org/10.4229/28THEUPVSEC2013-5CO.6.3. random forest. Wiley Interdiscip. Rev. Data Min. Knowl. Discov. May 2019;9(3):
[27] Parretta A, Sarno A, Vicari LRM. Effects of solar irradiation conditions on the e1301. https://doi.org/10.1002/WIDM.1301.
outdoor performance of photovoltaic modules. Opt Commun 1998;153(1–3): [55] Belgiu M, Drăgu L. Random forest in remote sensing: a review of applications and
153–63. https://doi.org/10.1016/S0030-4018(98)00192-8. future directions. ISPRS J Photogrammetry Remote Sens Apr. 2016;114:24–31.
[28] Trigo-Gonzalez M, et al. Development and comparison of PV production estimation https://doi.org/10.1016/J.ISPRSJPRS.2016.01.011.
models for mc-Si technologies in Chile and Spain. J Clean Prod 2021;281:125360. [56] Assouline D, Mohajeri N, Scartezzini JL. Large-scale rooftop solar photovoltaic
https://doi.org/10.1016/j.jclepro.2020.125360. Jan. technical potential estimation using Random Forests. Appl Energy May 2018;217:
[29] Tossa AK, et al. Artificial intelligence technique for estimating PV modules 189–211. https://doi.org/10.1016/J.APENERGY.2018.02.118.
performance ratio under outdoor operating conditions. J Renew Sustain Energy [57] Mathai N, Chen Y, Kirchmair J. Validation strategies for target prediction methods.
Oct. 2018;10(5):053505. https://doi.org/10.1063/1.5042217. Briefings Bioinf May 2020;21(3):791–802. https://doi.org/10.1093/BIB/BBZ026.
[30] Bandong S, Leksono E, Purwarianti A, Joelianto E. Performance ratio estimation [58] Kuhn M, Johnson K. Applied predictive modeling. Appl. Predict. Model. Jan. 2013:
and prediction of solar power plants using machine learning to improve energy 1–600. https://doi.org/10.1007/978-1-4614-6849-3.
reliability. SAVE Proc 2019:36–41. https://doi.org/10.1109/ICA.2019.8916687. [59] Bounoua Z, Ouazzani Chahidi L, Mechaqrane A. Estimation of daily global solar
2019 6th Int. Conf. Instrumentation, Control. Autom. ICA, . [Accessed July 2019]. radiation using empirical and machine-learning methods: a case study of five
[31] Hashemi B, Taheri S, Cretu AM, Pouresmaeil E. Systematic photovoltaic system Moroccan locations. Sustain. Mater. Technol. 2021;28:e00261. https://doi.org/
power losses calculation and modeling using computational intelligence 10.1016/J.SUSMAT.2021.E00261. Jul.
techniques. Appl Energy 2021;284. https://doi.org/10.1016/J. [60] Guenounou A, Malek A, Aillerie M. Comparative performance of PV panels of
APENERGY.2020.116396. Feb. different technologies over one year of exposure: application to a coastal

16
E. Fuster-Palop et al. Renewable and Sustainable Energy Reviews 170 (2022) 112968

Mediterranean region of Algeria. Energy Convers Manag Apr. 2016;114:356–63. [71] Government of Spain. Ministry of Industry Tourism and Trade, “Royal Decree 661/
https://doi.org/10.1016/J.ENCONMAN.2016.02.044. 2007,” BOE, 2007. [Online]. Available: https://www.boe.es/eli/es/rd/2007/05/
[61] Elamim A, Hartiti B, Haibaoui A, Lfakir A, Thevenin P. Analysis and comparison of 25/661. [Accessed 14 September 2022]. Accessed.
different PV technologies for determining the optimal PV panels- A case study in [72] Government of Spain. Ministry of industry tourism and trade. Royal Decree 413/
Mohammedia , Morocco. IOSR J Electr Electron Eng Jan. 2017;12:37–45. https:// 2014,” BOE, 2014. [Online]. Available: https://www.boe.es/buscar/doc.php?
doi.org/10.9790/1676-1201013745. 01. id=BOE-A-2014-6123. Accessed: 18-Sep-2022.
[62] Alshare A, Tashtoush B, Altarazi S, El-Khalil H. Energy and economic analysis of a [73] Blanco-Díez P, Díez-Mediavilla M, Alonso-Tristán C. Review of the legislative
5 MW photovoltaic system in northern Jordan. Case Stud Therm Eng 2020;21: framework for the remuneration of photovoltaic production in Spain: a case study.
100722. https://doi.org/10.1016/J.CSITE.2020.100722. Oct. Sustain Times 2020;12. https://doi.org/10.3390/SU12031214. Page 1214, vol. 12,
[63] Martín-Martínez S, Cañas-Carretón M, Honrubia-Escribano A, Gómez-Lázaro E. no. 3, p. 1214, Feb. 2020.
Performance evaluation of large solar photovoltaic power plants in Spain. Energy [74] Bansal N, Pany P, Singh G. Visual degradation and performance evaluation of
Convers Manag Mar. 2019;183:515–28. https://doi.org/10.1016/J. utility scale solar photovoltaic power plant in hot and dry climate in western India.
ENCONMAN.2018.12.116. Case Stud Therm Eng 2021;26:101010. https://doi.org/10.1016/J.
[64] Eke R, Demircan H. Performance analysis of a multi crystalline Si photovoltaic CSITE.2021.101010. Aug.
module under Mugla climatic conditions in Turkey. Energy Convers Manag Jan. [75] Comisión Nacional de los Mercados y la Competencia (CNMC). Estadísticas. Precios
2013;65:580–6. https://doi.org/10.1016/J.ENCONMAN.2012.09.007. del Mercado de Producción de Energía Eléctrica [Online]. Available: https://www.
[65] Murat Ates A, Singh H. Rooftop solar Photovoltaic (PV) plant – one year measured cnmc.es/estadisticas?hidtipo=12750. [Accessed 18 September 2022]. Accessed.
performance and simulations. J King Saud Univ Sci May 2021;33(3):101361. [76] Comisión Nacional de los Mercados y la Competencia (CNMC). Evaluación del
https://doi.org/10.1016/J.JKSUS.2021.101361. régimen retributivo específico correspondiente al periodo 2014-2020. Expediente
[66] Bouacha S, et al. Performance analysis of the first photovoltaic grid-connected núm. INF/DE/037/21 2021;1. https://www.cnmc.es/expedientes/infde03721.
system in Algeria. Energy Sustain. Dev. Aug. 2020;57:1–11. https://doi.org/ [Accessed 12 September 2022]. 8-8.
10.1016/J.ESD.2020.04.002. [77] Walker A, et al. Model of operation-and-maintenance costs for photovoltaic
[67] Attari K, Elyaakoubi A, Asselman A. Performance analysis and investigation of a systems. 2020.
grid-connected photovoltaic installation in Morocco. Energy Rep Nov. 2016;2: [78] inflation.eu. Historic inflation Spain – historic CPI inflation Spain [Online].
261–6. https://doi.org/10.1016/J.EGYR.2016.10.004. Available: https://www.inflation.eu/en/inflation-rates/spain/historic-inflation/
[68] Savvakis N, Tsoutsos T. Performance assessment of a thin film photovoltaic system cpi-inflation-spain.aspx [Accessed: 18-Sep-2022.
under actual Mediterranean climate conditions in the island of Crete. Energy 2015; [79] Ministry for the ecological transition and the demographic challenge, “orden TED/
90:1435–1455, Oct. https://doi.org/10.1016/J.ENERGY.2015.06.098. 171/2020. BOE, 2021. [Online]. Available: https://www.boe.es/buscar/act.php?
[69] Herbazi R, et al. Performance evaluation and analysis of polycrystalline id=BOE-A-2020-2838. [Accessed 12 September 2022]. Accessed.
photovoltaic plant located in Northern Morocco. https://doi.org/10.1080/01 [80] Wilcox RR. ONE-WAY ANOVA. Appl. Contemp. Stat. Tech. Jan. 2003:285–328.
430750.2019.1694985; 2019. 10.1080/01430750.2019.1694985. https://doi.org/10.1016/B978-012751541-0/50030-4.
[70] A. A. Merrouni, A. I. Amrani, and A. Mezrhab, “Electricity production from large [81] Liu H. Comparing Welch’s ANOVA, a kruskal-wallis test and traditional ANOVA in
scale PV plants: benchmarking the potential of Morocco against California, US,” case of heterogeneity of variance. Theses Diss.; 2015. https://doi.org/10.25772/
Energy Procedia, vol. 119, pp. 346–355, Jul. 2017, doi: 10.1016/J. BWFP-YE95.
EGYPRO.2017.07.118. [82] Schreiber-Gregory DN, Jackson HM, Bader K. Logistic and linear regression
assumptions : violation recognition and control. 2018.

17