The accurate parameters extraction is an important step to obtain a robust PV outputs forecasting for static or dynamic modes. For these aims, several approaches have been proposed for photovoltaic (PV) cell modeling including electrical circuit-based model, empirical models, and non-parametrical models. Moreover, numerous parameter extraction methods have been introduced in the literature depending on the proposed model and the operating mode. These methods can be classified into two main approaches including automatic numerical and analytical approaches. These approaches are commonly applied in the static mode, whereas they can be employed for dynamic parameters extraction. In this chapter, as a first stage, the static parameters extraction for both single and double diodes models is exposed wherein Genetic Algorithm and outdoor measurements are considered for fixed irradiation and temperature. In the second stage, a dynamic parameters extraction is carried out using Levenberg-Marquardt algorithm, where 1 day profile outdoor measurement is considered. After that, the robustness of the proposed approaches is evaluated and the parameters obtained by the static method and that given by the dynamic technique are compared. The test is carried out using 3 days with different weather conditions profiles. The obtained results show that the parameters extraction by dynamic techniques gives satisfactory performances in terms of agreement with the real data.
- photovoltaic module
- static parameters extraction
- dynamic parameters extraction
- empirical model
- electrical model
The increasing development of PV technologies brought out their potential to provide the energy abundance across the world. Hence, they have been interested by several research groups in the purpose to improve their behavior and extend their life-time. Meanwhile, giving an accurate forecasting of the PV outputs behavior has been always a real issue related to their nonlinearity. Two modes can characterize the PV module in terms of modeling, the first one is the static mode wherein the obtained model is characterized and validated for fixed weather conditions (irradiation and temperature), while the second is the dynamic mode where the validation is carried out using variable weather conditions. In this context, several models of the PV cell/module have been introduced in the literature based mainly on several approaches including electrical, empirical, and non-parametrical modeling. For the non-parametrical models, two approaches are introduced namely, Artificial Neural Network (ANN) and Neuron-Fuzzy based models. The accuracy and the robustness of these approaches rely strongly on the richness of the training dataset in terms of scenarios. Neuron-Fuzzy techniques has been used to predict cell short-circuit current and open-circuit voltage for the static representation . The ANN-based technique is introduced to model the PV array power for the embedded systems implementation . This technique has been tested for dynamic mode dealing . The ANN technique is also used for the prediction of the PV cell/module voltage directed to amorphous silicon PV technology wherein the obtained network has been tested by real dynamic data .
In the empirical or the analytical approaches, several models have been proposed to estimate accurately both static and dynamic modes of PV cell/modules. An analytical model is proposed by  based on manufacturer characteristic. This model provides acceptable results for both static and dynamic working. Another model has been introduced by Sandia National Laboratory , which is widely employed for PV cell/model forecasting especially for the large-scale arrays. Other empirical models have been proposed in order to estimate the PV array power under uniform shading [7, 8].
In the electrical approaches, two models widely prevail owing to their simplicity based on equivalent circuits namely: single (SDM) and double (DDM) diodes-based models. These models can deal with both static and dynamic modes with an acceptable accuracy. Besides, several enhancements have been introduced in these models to minimize parameters number and give more simplicity [9, 10]. The expressions of both photo-generation and diode saturation currents have been improved to give more accuracy in the dynamic working .
All models possess unknown parameters, which should be identified according to the module used in the practice. For this end, numerous approaches have been introduced in the literature including analytical and automatic numerical methods. The analytical methods usually rely on specific points on the I-V curve and on some value given by manufacturer. However, a significant error can be engendered if one of more of selected points is incorrect .
Owing to their adequate results, automatic numerical methods are prevail in PV models parameters extraction either through the use of deterministic algorithms such as: Newton model modified with Levenberg , Levenberg-Marquardt , Simulated Annealing algorithm (SA) , Pattern Search (PS) , Nelder-Mead Simplex algorithm (NMS) , and hybrid Nelder-Mead and modified particle swarm optimization  or by introducing metaheuristic algorithms such as: Genetic Algorithms (GA) , Particle Swarm Optimization (PSO) [20, 21, 22, 23, 24, 25, 26], Cuckoo Search (CS) , Artificial Bee Colony (ABC) [26, 28], and Artificial Bee Swarm (ABS) . Moreover, other algorithms have been introduced like (FPA) [30, 31], hybrid Bee Pollinator Flower Pollination Algorithm (BPFPA) [31, 32], Harmony Search (HS) , Artificial Fish Swarm Algorithm (AFSA) , and other algorithms. The majority of the aforementioned algorithms have been applied for static parameters extraction.
In this chapter, modeling and parameters extraction of PV cell/module are detailed. Where, comparison study among three models by applying static and dynamic identification using out-door measurement.
2. PV cell/module modeling
The PV cell presents outputs variation, which depends on weather conditions namely, irradiation and temperature. As illustrated in Figure 1, for load variation from open circuit to short circuit, the PV cell shows nonlinear characteristic that possess a maximum point of power. For an optimal working, the load should be adapted at this point. In this section, three PV cell models will be employed and improved using automatic parameters extraction namely: the empirical Sandia model and both single and double diodes electrical models.
2.1 Sandia model
This empirical model given by SANDIA National Laboratories provides relatively accurate dynamic forecast for PV cell/module by describing the thermal, the electrical, and the optical characteristics. Also, this model can be destined for any technology and can be adapted with any scale of PV arrays. Furthermore, its simplicity can qualify it to be used for real-time online prediction. Expressions (1)–(4) describe the variation of Impp, Vmpp, and Pmpp, respectively.
where, C0–3 are empirical parameters to be identified, Imp_STC, Vmp_STC are the current and the voltage in the maximum power point under standard test condition, Ee is the effective irradiation, K is the Boltzmann constant, q is the electron charge, δ(T) is the thermal voltage, αmp and βmp are, respectively, the current and the voltage temperature coefficient .
2.2 Single diode based model (SDM)
This physical model is based on the electrical approach illustrated in Figure 2 wherein the PV cell is composed of: a photo-generation current source and a diode while joule losses are represented by two serial and parallel resistors.
From this electrical representation, expression (5) and (6) can be obtained to describe the evolution of both current and voltage. The output current is expressed as a sum of the photo-generation current Iph, the diode current Id, and the shunt current Ish.
where, K is the constant of Boltzmann, q is the electron charge, T is the cell temperature, n is the diode ideality factor, and I0 is the current saturation due to diffusion and recombination.
Rs, Rsh, and n are parameters to be identified in the static study and can be adjusted in the dynamic study.
Eq. (10) describes the evolution of the energy band-gap Eg as function of the cell temperature.
where Eg0 and Egr are the energy band-gap of the silicon at 0°C and at the reference temperature Tr, respectively, α and β are constants of the material.
The photo-generation current is given by Eq. (11) as a function of the irradiation and the cell temperature.
where G is the input irradiation, T is the cell temperature, Isc is the module short-circuit current, and μ is the coefficient temperature/short-circuit current (given by the manufacturer) .
Eg0, α, β, μ, Isc, and Voc are parameters to be identified in the dynamic study.
2.3 Double diode-based model
From the electrical representation illustrated in Figure 3, the PV cell can be represented by a source of current that represents the photo-generation, two diodes and both parallel and serial resistances representing the loss of energy inside the cell.
In which I01–2 are currents saturation of the two diodes that resulted from diffusion and recombination, n1–2 are ideally factors.
n1–2 and Rs, Rsh are parameters which will be identified in the static study and they can be adjusted in the dynamic study.
The photo-generation current is represented by the same expression of the single diode model (Eq. (11)).
Parameters Eg01–2, α1–2, β1–2, μ, Isc and Voc will be identified in the dynamic study.
3. Static parameters extraction of PV module
For fixed irradiation and temperature, a static parameters extraction will be done to extract five parameters in SDM and seven parameters in DDM. A numerical stochastic optimization algorithm is used in this identification. This algorithm namely, Genetic Algorithm (GA), is employed to minimize the cost function given in Eq. (19) which expresses the root mean square error (RMSE) between the measured PV module I(v) characteristic and that given by the models. For this and, outdoor static measurements have been carried out using the peak measuring device tracer (PVPM 2540C), whose characteristics are illustrated in Table 1. This device has been programed to provide both I(V) and P(V) curves of 101 samples per 1 min.
where I is the simulated current, Vm is the measured voltage, N is the number of sample in I(V) characteristics. The error between the measured and simulated I(V) characteristics for the aforementioned models are expressed in Eqs. (20) and (21).
3.1 Genetic Algorithm
The Genetic Algorithm (GA) is a stochastic algorithm imitated from the biological genetic process used to find an approximate solution for optimization problems. Like in the natural concept, the chromosome is the holder of the genes that the child can probably get from his parents. By analogy, these genes represent the variables (parameters) of the function to be minimized. Five steps can characterize the GA namely, generation of initial population, evaluation of fitness, selection, crossover and mutation [37, 38].
3.1.1 Initial population
The process starts by the generation the initial population of N chromosome coded in binary. Each vector chromosome is formed of group of parameters in which its length M is given in Eq. (22) wherein n is the number of parameters and Nb is the length of the sub-string (number of bits) of each parameter as shown in Figure 4. The length of the integer part given by the vector Conv (Eq. (23)) is used to limit the research domain in which, Pi (Eq. (24)) is the parameter value in decimal code [37, 38].
In this stage, the parameters values that have been randomly generated and decoded in decimal base will be substituted in the cost function to be optimized. The fitness is the solution of the parameters in the RMSE (x) function calculated in Eq. (19). Its value is mathematically expressed in Eq. (23) [37, 38].
The chromosomes that will participate as parents to generate a new child are chosen in this step. Any chromosome in the generated population can be chosen however, the individual that presents a good fitness have a high probability. The technique used for the chromosome choice is the roulette wheel illustrated in Figure 5, wherein the selection probability Ps, expressed in Eq. (24), is calculated, consists of a cumulative sum of the fitness of each chromosome orderly relative to the sum of all fitness. After that, the process generates a random drawing probability Pr. Hence, the first chromosome corresponds to Pr < Ps is chosen for the next steps (crossover and Mutation).
After selecting the chromosomes, the algorithm gives birth to new children by performing a crossover between each two chromosomes. For this end, a drawing probability Pr is generated and compared with the crossover probability Pc (usually high probability). Hence, the parents chromosomes that corresponds to Pr < Pc will be chosen for child generation. If not, the same chromosomes are kept. As illustrated in Figure 6, the crossover by point is used wherein the bits after the point randomly chosen are swapped [37, 38].
In this step, the algorithm introduces a change in some characters of the selected chromosomes in order to expand the search space if the initial population does not fall in the optimal solution. In the binary coding, the selected bit will change from 1 to 0 and vice versa as described in Figure 7. The mutation has low probability Pm in which, it will be affected for characters that correspond to Pr < Pm in which, Pr is the drawing probability (randomly generated) [37, 38].
Our system is formed of mono-crystalline PV module SANYO technology with the characteristics listed in Table 2, peak measuring tracer and the necessary sensors. The experimental platform is illustrated in both Figure 8a and b.
|Application||DC voltage||DC current||Temperature||Irradiance||Measuring period single measurement||I-V curve samples|
|PV modules and small strings||25/50/100/250 V||2/5/10/40 A||−40°C to +120°C with Pt1000||0–1300 (W/m2) (standard-sensor)||0.02–2 (s)||101|
|Cell size||156 × 156 mm|
|PV module dimension||1319 × 894 × 35 mm|
|Nominal power||180 W|
|Open circuit voltage Voc||66.4 V|
|Short circuit current Isc||3.65 A|
|Voltage Vmpp||54 V|
|Current Impp||3.33 A|
|Nominal operating temperature NOCT||45 ± 2°C|
|Temperature coefficient (Pmax)||−0.33%/°C|
|Temperature coefficient (Isc)||1.10 mA/°C|
|Temperature coefficient (Voc)||−0.173 V/°C|
After running of GA for 1000 cycles with the parameters listed in Table 3 for both SDM and DDM using outdoor measurement of the systems, wherein extracted parameters are listed in Table 4. Figure 9a and b illustrate the agreement between the measured and simulated I(V) and P(V) characteristics for SDM model whose obtained parameters are summarized in Table 5. Figure 10a and b show the agreement between the measured and simulated I(V) and P(V) characteristics for DDM model whose extracted parameters are summarized in Table 6.
|Number of cycle||1000|
|The electrical parameter||Iph [A]||I0 [A]||n||Rs [Ω]||Rsh [KΩ]|
|The identified value||3.0195||49591e−005||1.874*96||0.3273||8.1514|
|The electrical parameter||Iph [A]||I01 [A]||I02 [A]||n1||n2||Rs [Ω]||Rsh [KΩ]|
|The identified value||3.0289||6.1035e−005||3.8147–006||1.3658*96||1.9179*96||0.1017||5.992|
Some parameters will be identified again in the dynamic study including the parameters involved in I0 and Iph equations, while the remaining will be adjusted to give more accuracy under variable weather conditions.
4. Dynamic parameters extraction of PV module
In this section, dynamic parameters identification will be described wherein the process is done by using 1 day profile of measurement. This allows to improve the nominal values given by the manufacturer, which can cause a significant error due to operating conditions and the consumed lifetime. Moreover, parameters obtained by static method can be adjusted by dynamic identification. For this end, automatic parameters adjustment using Levenberg-Marquardt optimization algorithm is employed.
As illustrated in Figure 11, the main idea is to take both PV module model and the MPPT as a single system with three outputs namely, Impp, Vmpp, and Pmpp. These outputs will be compared with 1 day profile of outdoor measurements. The process consists in minimizing the error between the model outputs and the real data. The whole system has been implemented in Matlab/Simulink tool.
For Sandia model, the process is carried out without the use of MPPT considering that this model has been established to the dynamic forecasting.
4.1 The MPPT used
An Accurate fuzzy logic MPPT algorithm is employed in our system (for SDM and DDM) in order to get satisfactory results in terms of precision and accuracy. The algorithm is used to control a DC/DC boost converter for the purpose to keep the PV module working at the maximum point of power. Mamdani inference model is used with two inputs namely, the error E and the variation of the error ΔE. The calculation of these attributes is expressed in Eqs. (27)–(30).
Three steps can characterize the fuzzy algorithm; the first one is the fuzzification process that consists on the conversion of the numerical inputs values (E and ΔE) into linguistic values by the substitution in the membership functions. The second step is the inference process, which is considered as the main stage in the fuzzy algorithm wherein the relation between the inputs and the output is done. The third step is the defuzzification where the process converts the linguistic decision into numerical output. Figure 12 describes briefly the fuzzy processing steps .
For both inputs and output, five trapezoidal and triangular membership functions have been employed namely: NS (negative strong), NL (negative low), ZE (zero), PL (positive low) and PS (positive strong). The center of gravity based method is used for the defuzzification to provide the control duty cycle after applying the Mamdani inference model given in Table 6. Figure 13a,b and d describes the used membership functions .
4.2 Simulation study
The Levenberg-Marquardt algorithm is implemented using 1 day profile of outdoor real measurement of dynamic PV outputs (Pmpp, Impp and Vmpp). The process consists in minimizing the error between simulated outputs of both SDM and DDM and 8 h of real data (09:00 am–05:00 pm). The peak measuring device tracer (PVPM 2540C) has been programmed to provide 1 sample per minute. Table 7 lists the lower and upper limits search of the extracted parameters. The extracted parameters using the dynamic method are summarized in Table 8. The inputs measurement of the irradiation and the temperature are illustrated in Figure 14a and b, respectively. Satisfactory results have been obtained in terms of matching between the real data and the simulated outputs Pmpp, Impp and Vmpp for SDM, DDM and Sandia as shown in Figure 15a–c, respectively.
|SDM parameters||Boundaries||DDM parameters||Boundaries||Sandia parameters||Boundaries|
|n||[0, 2]*96||n1, n2||[0, 2]*96||C0||[0, 2]|
|Rs [Ω]||[0, 1]||Rs [Ω]||[0, 1]||C1||[−1, 1]|
|Rsh [Ω]||[0, 104]||Rsh [Ω]||[0, 104]||C2||[−10, 10]|
|Eg0||[0, 1]||Eg01–2||[0, 2]||C3||[−10, 50]|
|A||[0, 1]||α1–2||[0, 1]||αImp [°C−1]||[0, 1]|
|B||[0, 104]||β1–2||[0, 104]||βVmp [V/°C]||[−1, 0]|
|μ||[0, 1]||μ||[0, 1]||n|
|Isc||[3, 3.7]||Isc||[3, 3.7]|
|Voc||[60, 66.8]||Voc||[60, 66.8]|
|SDM parameters||Values||DDM parameters||Values||Sandia parameters||Values|
|n||105.73/96||n1, n2||90.73/96; 73.39/96||C0||1.058|
|Rs [Ω]||0.82495||Rs [Ω]||0.3219||C1||0.020|
|Rsh [Ω]||8.371 × 103||Rsh [Ω]||4.9664 × 103||C2||−0.341|
|Eg0 [ev]||1.4525||Eg01–2 [ev]||1.649; 1.31||C3||−9.997|
|α||6.56 × 10−4||α1–2||0.0018; 0.0132||αImp [°C−1]||2.53 × 10−14|
|β||126.11||β1–2||694.84; 1020.76||βVmp [V/°C]||−0.203|
5. Experimental validation
In this section, a validation with an unseen data is carried out to test and compare the effectiveness of the proposed enhancement. The three developed models will be compared with real measurement profile (09:00 am–05:00 pm) of irradiation and temperature for different weather conditions. Wherein, the SDM and DDM models using the developed parameters are compared with the former nominal parameters listed in Table 9, Sandia model and the real data. Three different skies of real measurement have been used for this validation namely, clear day, semi-cloudy day and cloudy day. The matching in the power (Pmpp) between and the real data and SDM and DDM with nominal parameters, SDM and DDM with the new parameters and Sandia model is illustrated in Figure 16a–c for clear day, semi-cloudy day and cloudy day respectively. Besides, the agreement in the voltage and the current (Vmpp and Impp) for these models with the real data is shown in Figure 17a–c and Figure 18a–c, respectively.
|α||4.73 × 10−4|
It is clearly found that an improved agreement has been shown by models with new parameters compared to that given by the nominal parameters and the static method.
For more clarity, the hourly power efficiency given by the presented models and the real data has been calculated. It consists on the average of the power during 1 h versus the optimal PV module power (Eq. (31)) .
in which, N is the number samples per hour.
Figure 19a–c show the bar-graph of the hourly power efficiency of the proposed models for the three weather conditions, namely the clear day, semi-cloudy day and cloudy day, respectively. The enhanced models present higher hourly power efficiency versus models with the former parameters and those given by the static technique. Furthermore, root mean square error (RMS) and the mean absolute error (MAE) between the real data and the studied models are calculated by Eqs. (19) and (32) to show the enhancement of the proposed method.
where N is the number of samples .
The aforementioned results show clearly that the extracted parameters of the PV module using dynamic techniques present more accuracy compared with the static method and the parameters given by the manufacturer. Indeed, the parameters obtained by the static method are clearly improved for variable weather conditions (irradiation and temperature), which is confirmed using different skies. Table 10 summarizes the calculated RMS and MAE errors values which show obviously that the developed models present advantages comparing with real outdoor data of different weather conditions.
In this chapter, both dynamic and static parameters identification methods have been highlighted and compared with real measurement. The SDM and DDM nominal parameters involved in I0 and Iph equations have been developed by dynamic method. This improved result has been compared with that given by the static technique and Sandia model versus out-door real data for different skies (clear day, semi-cloudy day and cloudy day). It was found that SDM and DDM based on the parameters extracted by dynamic method give satisfactory accuracy, which is confirmed by some calculated indicator such as: the hourly efficiency and both root mean square error (RMS) and the mean absolute error (MAE). This allows to solve modeling problems of PV module that apply for several applications such as fault detection.