=Paper=
{{Paper
|id=Vol-2738/paper32
|storemode=property
|title=Comparison of Knowledge-based Feature Vector Extraction and Geometrical Parameters of Photovoltaic I-V Curves
|pdfUrl=https://ceur-ws.org/Vol-2738/LWDA2020_paper_32.pdf
|volume=Vol-2738
|authors=Cem Basoglu,Grit Behrens,Konrad Mertes,Matthias Diehl
|dblpUrl=https://dblp.org/rec/conf/lwa/BasogluBMD20
}}
==Comparison of Knowledge-based Feature Vector Extraction and Geometrical Parameters of Photovoltaic I-V Curves==
https://ceur-ws.org/Vol-2738/LWDA2020_paper_32.pdf
Comparison of knowledge based feature vector
extraction and geometrical parameters of
Photovoltaic I-V Curves
C. Basoglu1,4 , G. Behrens1 , K. Mertens2 , and M. Diehl3
1
Fachhochschule Bielefeld, Solar Computing Lab, Artilleriestraße 9, 32427 Minden,
Germany
2
Fachhochschule Münster, Fachbereich Elektrotechnik und Informatik,
Photovoltaik-Prüflabor, Steinfurt
3
photovoltaikbuero, Ternus und Diehl GbR, Rüsselsheim
4
Contact: cbasoglu@fh-bielefeld.de
Abstract. Current methods for evaluating the performance of PV mod-
ules and systems in the field are exposed to weather conditions during
system evaluation. The experimental measurement of performance nat-
urally requires a corresponding amount of solar radiation, which is not
available at all times of the year. The aim of this work is the development
of a method for the weather-independent PV plant evaluation using the
so-called dark I-V curve and an artificial neural network (ANN). The
dark I-V curve can be measured at any time of the year and in any
weather condition. In combination with the performance measurements
from conventional methods an extensive database is already available,
which was used as the ground truth for the development of the proposed
model. The results show that with the proposed method a prediction
of the power output for illumination levels above 800W/m2 a maximum
prediction error below 10% is achieved. Thus, the dark I-V curve can
be used for a weather-independent evaluation of PV systems in order to
show first indications of performance losses and further analysis.
Keywords: Machine Learning · Artificial Neural Network · I-V Curve
1 Introduction
The in-field evaluation and fault diagnosis is crucial for a high-yield operation of
photovoltaic plants. Analyzing the light I-V curve (current-voltage curve) of a
PV array is the commonly used method for in-field evaluation and characterisa-
tion [2,3,4,5,9,12]. The I-V curve (see Figure 1) describes the energy conversion
capacity under given conditions of irradiation and temperature. Only the ex-
perimental measurement of the I-V curve is able to specify with precision the
electrical parameters of a photovoltaic cell, module or array.
Copyright c 2020 by the paper’s authors. Use permitted under Creative Commons
License Attribution 4.0 International (CC BY 4.0).
I-V
Power
12 (V MPP, PMPP) 5
ISC
10
4
Current (I)
8
Power (P)
3
6 (V MPP, IMPP)
2
4
1
2
0 0
0 100 200 300 400 500 600 700
Voltage (V) V OC
Fig. 1: I-V curve of a solar cell
The I-V curve starts at the short-circuit condition, ISC , where the voltage is
zero. The current decreases slightly as the voltage is increased, until the curve
nears the open-circuit condition where the current rapidly drops off. The curve
ends at the open-circuit condition, VOC , with the current at zero. At some point
on the I-V curve, the power of the cell is at its maximum. This point is known
as the maximum power point (MPP), and solar cells are the most efficient at
converting light energy into electrical energy at this point.
The in-field experimental measurement of the I-V curve is highly dependent
on the weather conditions during the evaluation of the system. Passing clouds or
shadows from other objects at certain times of the day result in a not negligible
loss of time. Furthermore, the measurement requires sufficient light irradiation
which is not available in all seasons of the year. To address these issues, this
study proposes a novel method by predicting the light I-V curve using the so
called dark I-V curve and an ANN. The dark I-V curve is measured without
illumination by using an external power supply as reverse current source and
is commonly used in the manufacturing process of solar modules [1,6,8]. Thus
makes the dark I-V curve independent of weather conditions while maintaining
many of the previously described electric characteristics.
2 Related Work
There are already several approaches which can identify faults using dark I-V
curves for diagnostic purposes. For example, the method proposed by Mertens,
K. et al. [10] is able to detect potential induced degradation (PID) and diode
errors using numerical analysis of the dark I-V curve.
Besides the diagnostic value of the dark I-V curves, there are only a few meth-
ods which focus on predicting the light conversion performance under different
illumination levels. King et al. [8] uses the two diode model (see Equation 1)
with some experience based parameter assumptions to extract the remaining
parameters of the model. Mertens et al. [11] uses also the two diode model and
module parameters from the manufacturer data-sheet to solve remaining param-
eters of the model. Both methods extract the two diode model parameters using
the dark I-V curve and use these parameters to calculate the light I-V curve.
In this work, one approach also use the two diode model to extract the model
parameters of both, the dark and light I-V curve, and uses this knowledge to
learn the relationship between them.
3 Method
In cooperation with our research partners, a database with 3424 light and 1656
dark I-V curves from field and laboratory measurements of 131 different module
types has been collected. Each I-V curve consists of 200 current-voltage pairs and
additional metadata like irradiance level and temperature. To build the ground
truth for the proposed method, the dark I-V curves of each module type are
combined with all light I-V curves of the same module type, which results in
37686 training and validation data sets for the neural network. Three different
feature extraction approaches for the I-V curve are considered. Each approach
generates a feature set of the dark and light I-V curve. The feature set of the
dark I-V curve is used as the input vector and the light I-V curve as the output
vector for the neural network.
The first approach (E1) uses the the electrical two diode model [7] and per-
forms the levenberg-marquardt curve fitting algorithm to fit the following func-
tion to the measured data points.
U +I·RS U +I·RS U + I · RS
I = IP H − IS1 · (e η1 ·UT − 1) − IS2 · (e η2 ·UT − 1) − (1)
RP
The extracted series resistance (RS ), parallel shunt resistance (RP ), saturation
currents (IS1,2 ), diode ideality factors (η1,2 ) and temperature coefficient (UT )
of the dark and light I-V curves, are used as an input and respectively output
vector for the neural network.
The second approach (E2) for feature extraction performs a principal compo-
nent analysis of the I-V curves while retaining 95% of its variance. This approach
generates 64 components for the dark I-V curve and 76 components for the light
I-V curve.
The last approach (E3) uses a barycentric lagrange interpolation to extract
20 equidistant points for each I-V curve. Since the points are equidistant, the
values on the x-axis (voltage) are removed and only implied by its position in
the vector.
The architecture of the neural network for all approaches consists of the
input, output and a single hidden layer. The number of nodes in the hidden
layer equals half the average of the input and output layers. The output of the
nodes in the hidden layer are controlled by the tangens hyperbolicus (tanh)
activation function and the network is optimized using the RMSE error function
with RMSprop optimization algorithm. Finally the dense networks are trained
with 75% of the combined data sets, using 80% of it for the training and 20%
for the optimizer validation to avoid over fitting.
4 Experimental Results
For the final validation of the different approaches, 25% of the combined data
sets are used. Table 1 shows the mean percentage error (MPE) using min/max
and percentiles for the error distribution. For this purpose the light I-V curves
are reconstructed from the features described and compared to the measured I-V
curves. Overall, the third approach (E3) using the equidistant interpolation of
Mean SD (σ) Min. 25% 50% 75% Max.
E1 6.08 4.82 0.86 2.89 4.80 8.30 25.28
E2 9.12 6.41 1.45 4.53 8.67 12.58 32.94
E3 4.12 3.44 0.39 1.66 3.28 5.58 21.19
Table 1: Mean percentage error (MPE) and percentiles using the different pre-
processing methods
the I-V curve yields the smallest prediction error. Figure 2 shows the predicted
I-V curve (green) and the measured I-V curve (dashed black) for the third ap-
proach (E3). With some exceptions, the prediction becomes more accurate with
6
Messung
Measure Measure
4 E3 E3
950 W/m² 36 °C 5
917 W/m² 24 °C
3 740 W/m² 37 °C 4 820 W/m² 45 °C
Current (I)
Current (I)
3 645 W/m² 42 °C
575 W/m² 37 °C
2
2 364 W/m² 26 °C
269 W/m² 24 °C
1
1
32 W/m² 14 °C
0 0
0 5 10 15 20 25 30 35 40 0 100 200 300 400 500 600
Voltage (U) Voltage (U)
(a) Solon P165/5 (b) Sunpower E19 240 (14 in series)
8 8
Messung
Measure Measure
E3 E3
980 W/m² 44 °C
6 875 W/m² 44 °C 6 868 W/m² 46 °C
765 W/m² 22 °C 746 W/m² 47 °C
Current (I)
Current (I)
4 600 W/m² 40 °C 4
394 W/m² 36 °C
2 2
218 W/m² 30 °C
43 W/m² 26 °C
0 0
0 100 200 300 400 500 0 100 200 300 400 500 600 700 800
Voltage (U) Voltage (U)
(c) Kyocera KC180 (20 in series) (d) Winaico WST-250P6 (22 in se-
ries)
Fig. 2: Prediction examples for different string configurations of the 3. approach
increasing irradiation. Above 800 W/m2 the maximum prediction error is be-
low 10%, which is in the range of peak performance prediction of commercially
available light I-V curve measurement devices [13].
5 Conclusion and Future Work
In summary we tested different pre-processing approaches in order to predict the
light I-V curve using the dark I-V curve and an ANN. The experimental results
shows that using interpolated points of the I-V curve yields better results than
using a PCA or the electrical two diode model for feature extraction. We plan to
further investigate recurrent neural networks with an increased number of I-V
curve points to further improve the prediction accuracy.
Acknowledgements
This work has been developed in the project PVServ 2.0 (reference number:
ZF4401205LT7) and is funded by the German ministry of economic and energy
(BMWi) within the research programme ZIM 2018.
References
1. Beier, J., Voss, B.: Humps in dark i-v-curves-analysis and explanation. In: Confer-
ence Record of the Twenty Third IEEE Photovoltaic Specialists Conference - 1993
(Cat. No.93CH3283-9). pp. 321–326 (1993)
2. Bressan, M., El Basri, Y., Galeano, A., Alonso, C.: A shadow fault detection
method based on the standard error analysis of iv curves. Renewable energy 99,
1181–1190 (2016)
3. Fadhel, S., Delpha, C., Diallo, D., Bahri, I., Migan, A., Trabelsi, M., Mimouni,
M.: Pv shading fault detection and classification based on iv curve using principal
component analysis: Application to isolated pv system. Solar Energy 179, 1–10
(2019)
4. Huang, J.M., Wai, R.J., Gao, W.: Newly-designed fault diagnostic method for solar
photovoltaic generation system based on iv-curve measurement. IEEE Access 7,
70919–70932 (2019)
5. Jones, C.B., Mart?nez-Ram?n, M., Smith, R., Carmignani, C.K., Lavrova, O.,
Robinson, C., Stein, J.S.: Automatic fault classification of photovoltaic strings
based on an in situ iv characterization system and a gaussian process algorithm.
In: 2016 IEEE 43rd Photovoltaic Specialists Conference (PVSC). pp. 1708–1713
(2016)
6. Kaminski, A., Marchand, J.J., Fave, A., Laugier, A.: New method of parameters
extraction from dark i-v curve. In: Conference Record of the Twenty Sixth IEEE
Photovoltaic Specialists Conference - 1997. pp. 203–206 (1997)
7. Kawamura, H., Naka, K., Yonekura, N., Yamanaka, S., Kawamura, H., Ohno, H.,
Naito, K.: Simulation of i–v characteristics of a pv module with shaded pv cells.
Solar Energy Materials and Solar Cells 75(3-4), 613–621 (2003)
8. King, D.L., Hansen, B.R., Kratochvil, J.A., Quintana, M.A.: Dark current-voltage
measurements on photovoltaic modules as a diagnostic or manufacturing tool. In:
Conference Record of the Twenty Sixth IEEE Photovoltaic Specialists Conference
- 1997. pp. 1125–1128 (1997)
9. Mellit, A., Tina, G., Kalogirou, S.: Fault detection and diagnosis methods for
photovoltaic systems: A review. Renewable and Sustainable Energy Reviews 91,
1 – 17 (2018). https://doi.org/https://doi.org/10.1016/j.rser.2018.03.062, http:
//www.sciencedirect.com/science/article/pii/S1364032118301370
10. Mertens, K., Arnds, A., Diehl, M.: Quick and effective plant evalua-
tion using dark-iv string curves. In: Proceedings of 33st European Photo-
voltaic Solar Energy Conference, Amsterdam, 2017. pp. 2346 – 2348 (2017).
https://doi.org/10.4229/EUPVSEC201
11. Mertens, K.: String-dunkelkennlinien: Eine neue effiziente methode zur anla-
genevaluation, 33. symposium photovoltaische solarenergie, staffelstein, 27.04.2018
(2019)
12. Sarikh, S., Raoufi, M., Bennouna, A., Benlarabi, A., Ikken, B.: Fault diagnosis in a
photovoltaic system through iv characteristics analysis. In: 2018 9th International
Renewable Energy Congress (IREC). pp. 1–6. IEEE (2018)
13. Wagner, A.: Photovoltaik Engineering. Springer (2006)