Feasibility of Artificial Neural Network for Maximum Power Point Estimation of

Non Crystalline-Si Photovoltaic Modules

Syafaruddina, Takashi Hiyamab

Department Computer Science and Electrical Engineering Kumamoto University

2-39-1 Kurokami, Kumamoto 860-8555 Japan asyafa@st.eecs.kumamoto-u.ac.jp bhiyama@eecs.kumamoto-u.ac.jp

Engin Karatepe Department of Electrical and Electronics Engineering

Ege University 35100 Bornova-Izmir, Turkey

engin.karatepe@ege.edu.tr

—Solar cell markets are growing favorably. The emerging non crystalline silicon (c-Si) technologies are starting to make significant in-roads into solar cell markets. The most of the artificial neural network (ANN) have been used in maximum power points tracking applications for c-Si solar cell technology. However, the characteristics of different solar cell technologies at maximum power point (MPP) have different trends in current-voltage characteristic. In this reason, the investigation of feasibility using neural networks is very important for different solar cell technologies to increase the efficiency of photovoltaic (PV) systems. The paper investigates three different ANN structures, such as radial basis function (RBF), adaptive neuro-fuzzy inference system (ANFIS) and three layered feed-forward neural network (TFFN) for identification the optimum operating voltage of non c-Si PV modules. These ANN models have been trained and verified for double junction amorphous Si (2j a-Si), triple junction amorphous Si (3j a-Si), Cadmium Indium Diselenide (CIS) and thin film Cadmium Telluride (CdTe) solar cell technologies. The results show that the flexibility of training process, the simplicity of network structure and the accuracy of validation error are important factors to select a suitable ANN model.

Keywords-ANN; RBF; ANFIS; TFFN; Solar Cell; 2j a-S; 3j a-Si; CIS; thin film CdTe

I. INTRODUCTION

Intelligent system by means the artificial neural network (ANN) has been satisfactorily used to solve the optimization tasks of engineering problems. The advantages of ANN methods over the other optimization methods are simple computational techniques and high pattern recognition abilities [1-4]. As the mature optimization techniques, there are many different ANN architectures. These may have their own advantages as well as disadvantages over the others in different applications. For example, radial basis function (RBF) neural network is found very fast during the training process and the structure is directly confirmed after training. However, there might be trivial errors occur in RBF method during the validation process. On the other hand, adaptive neuro-fuzzy inference system (ANFIS) method is also a very

strong network with high accuracy output during training and validation process. The accuracy of this method is likely depending on the type and number of membership function for the input signals. However, this network is generally designed for a single output. Therefore, for multi-objective optimization problems, it requires multi ANFIS network and each network must be separately trained. In comparison, the three layered feed-forward neural (TFFN) method needs intuitive decision of users to determine the best network structure. The problem of this method is too many possible network structures that can be selected during the training process [5].

The artificial neural network is successfully compatible with the optimization problem of PV system due to their abilities to deal with non-linear characteristics [6]. In the maximum power point tracking (MPPT) control application, the ANN method is very useful to identify the global MPP points directly without solving any non-linear equations. One of the MPPT methods utilizes the optimum voltage as the reference signal [7-9]. The optimum voltage is taken into consideration because this parameter is varied due to the environmental factors and their trends are unique in different type of solar cells. Moreover, under partially shaded or mismatched cell conditions, the optimum voltage might move significantly in the I-V characteristic. From this point of view, much more attentions are still necessary to deal with MPP optimization problems of PV systems under any scenarios.

Enormous studies of using ANN in the PV system application have been proposed so far [10-26]. Mostly, the ANN methods are utilized to forecast the generating power of system by estimating the insolation level [10-12]. In other papers, the ANN methods were combined with other optimization techniques [13, 14] and to determine the optimal power dispatch under random load [15]; with heuristic optimization approach to optimize operating costs of a representative PV based microgrid system [16] and with evolutionary programming to quantify the optimum values for the number of hidden nodes [17]. More specific study using ANN to optimize the solar power battery charger operation for MPPT controller was explained in [18]. The ANN was also

978-1-4244-5098-5/09/$26.00 ©2009 IEEE

used to optimize the duty ratio of buck-boost converter taking the environmental factors as the input signals [19, 20]. Most of the proposed ANN methods were using multi layer perceptron (MLP) structure. This structure has sometimes higher computational burden, especially for the feed-forward network. The RBF method can be found in [21] to provide the reference voltage for the MPPT controller and in [22] to justify the proposition for a new sizing procedure of stand alone PV system. Combining with neuro-fuzzy regulator, the RBF function has been used to increase the efficiency of PV system [23]. The ANFIS model has been used to train individually on each component of PV power system under variable climate conditions [24]. In associated with PV system technology, the researchers are still concentrating on modeling and simulating the c-Si based technology [25], while the solar cell technology is rapidly developing where the non c-Si solar cells are entering the PV market. The modeling of spectral effects on the short-circuit current of a-Si solar cell can be found in [26].

In this paper, three different structures of neural network; RBF neural network, ANFIS network and TFFN network are tested with non c-Si PV modules in order to investigate the performance of ANN structures. All ANN structures were trained using the voltage at maximum power point (VMPP) as the target. The obtained best network configurations are validated for four different scenarios. The results show that the flexibility of training process, the simplicity of network structure and the accuracy of validation error are important factors of selecting a suitable network for a PV module type.

II. DESCRIPTION OF THE MODEL

A. Modeling of PV Modules

Crystalline Si solar cells are the current mainstream in the market, but sales have been significantly expanded for non c-Si solar cells due to the development of nano-material technology. The main reason of this trend is to cut the manufacturing cost of conventional c-Si technology.

TABLE ISPECIFICATION OF PV MODULES UNDER 1000W/M2 AND 25OC

PVmodules

ISC

(A) VOC

(V) IMPP

(A) VMPP

(V) PMPP

(W) MST-43MV

US-32 ST-40 FS-50

0.787 2.4

2.59 1.0

101.0 23.8 22.2 90

0.616 1.94 2.41 0.77

71 16.5 16.6 65

43.74 32.01 40.00 50.05

The PV modules of specifications under 1000W/m2 and 25oC are presented in Table I. The Solarex MST-43MV and USSC UniSolar US-32 PV modules are the tandem junction and triple junction thin-film module amorphous silicon cells, respectively. These modules were designed for the purpose of effectively using the conversion of sunlight spectrum. In the tandem junction technology, solar cells are developed by depositing semiconductor alloys in thin layers on glass. The tandem-junction structure stacks two or three solar cells vertically, with each cell tuned for optimum conversion of

different segments of the spectrum. On the other hand, Siemens ST-40 and First Solar FS-50 are CIS and thin film CdTe PV modules, respectively. The ST-40 module is composed of a monolithic structure of series-connected Copper Indium Diselenide (CIS) based solar cells. These multiple-layer cells are characterized by exceptional spectral response and long-term performance integrity. The FS-50 uses very thin layers of compound semiconductor material with low temperature coefficients which provides for cost effective and greater energy production.

These PV modules are modeled following the I-V curve characteristic model developed by Sandia National Laboratory [27, 28]. The characteristics of short-circuit current and open circuit voltage are almost similar for all semiconductor types of solar cells. However, there might be different characteristics at maximum power point voltage. The VMPPcharacteristic is shown in Fig. 1. It can be seen from this figure that the correlation between the optimum voltage and irradiance/cell temperature for each module is non-linear. In this respect, tracking the optimum voltage using intelligent technique by means the ANN can give very good promising solution.

B. ANN Structures based Optimum Voltage

In this section, three different ANN structures are introduced. The configurations have two input signals; irradiance level (E) and cell temperature Tc and a single output signal; optimum voltage (Vop). The hidden nodes are determined during the training process. There are 228 numbers of training data set, between 100-1000W/m2 and 10-65oC, to cover the entire domain problems as VMPP=f(E,Tc).

RBF neural network is a typical neural network structure using local mapping instead of global mapping as in multi layer perceptron (MLP) [29]. In MLP method, all inputs cause an output, while in RBF method; only inputs near a receptive field produce activation function. The hidden layer is locally tuned neurons centered over receptive fields. Receptive fields are located in the input space areas where input vectors exist. If an input vector lies near the center of a receptive field, then that hidden layer will be activated. The training process using RBF network is very simple. Once the goal error is set, the training is stop and the number of hidden nodes is confirmed.

In the training process, the hidden nodes is obtained by adding one by one neuron in time until the sum-squared error of the network falls beneath an error goal. The equation of the sum of the squared errors (SSE) is described as:

( )2

1=−=

N

kopMPP VVSSE (1)

where N is the total number of training patterns. In this study, the parameter of training process: the mean squared error goal (GOAL), spread of radial basis functions (SPREAD), maximum number of neurons (MN) and the number of neurons to add between displays (DF) are 0.003, 1.0, 228, 1, respectively.

a. Solarex MST-43MV (2j a-Si) b. USSC Unisolar US-32 (3j a-Si)

c. Siemens ST-40 (CIS) d. First Solar FS-50 (thin film CdTe)

Fig. 1 Voltage at maximum power point (VMPP) for irradiance (100-1000W/m2) and cell temperature (10-60oC)

ANFIS network is especially designed for single output, called, Sugeno type fuzzy inference systems (FIS). This method is considered as hybrid learning algorithm because it combines the least-squares and back propagation gradient descent methods for training FIS membership function parameters. The training process using ANFIS method is very fast and the network structure is also directly confirmed. During the training process, once the number of epoch is reached, then the training is stop. In this study, the number of epoch set in the simulation is 20. The important structure of ANFIS network is the type and number of membership function for each input signal. In this study, the generalized bell membership function ‘gbellmf' is set for each input. More accuracy can be reached using this network by adding the number of membership function for each input, but the simulation progress becomes very slow.

The remained ANN structure is TFFN method. This method also uses back propagation algorithm and descent gradient method for adjusting the weights in order to reduce the learning error [30]. Once the error gradient is calculated, the weights are adjusted. There are many possibilities to construct this network. The selection of network structure is

based on the intuitive thinking of trainer. In this respect, the training process takes much time.

In this study, the training process of TFFN method is following the chart depicted in Fig.2. The ‘logsig’ function is utilized an activation function between layers. Other important factors during the training of TFFN network are the learning and momentum rates. Learning rate means how much the weight is changed at each step. If this rate is too small, the output training will be very precise, but the algorithm will take long time to converge. On the other hand, if too large, the outputs will be bouncing around and algorithm may diverge. The momentum rate is related to the time when the weights are updated. This rate is commonly determined by trial and error. In this study, the learning and momentum rate are specified to 0.2 and 0.85, respectively [31].

The training results in terms of number of hidden nodes (nh) and the sum of the squared errors (SSE) for each proposed ANN structure is presented in Table II. This result also confirms that the number of hidden nodes depend on the complexity of output target to the input signals.

Fig. 2. Flow chart of training process of TFFN structure

TABLE IIRESULTS OF ANN TRAINING PROCESS

PV modules RBF ANFIS TFFN

nh SSE nh SSE nh SSE

MST-43MV 12 0.000375 21 0.004373 4 0.001429 US-32 4 0.000608 21 0.000146 4 0.000114

ST-40 5 0.000948 21 0.000914 4 0.000227 FS-50 11 0.000739 21 0.003154 7 0.001087

III. SIMULATION RESULTS

The last stage of using ANN method is the validation error using different input signals. The trained ANN structures in the previous section are validated with four different input signals. These are defined as follows:

a. Ramp signal: )(105.5)/(10090 2 CtTandmWtE o

c +=+= (2) where 0 < time (t) < 10sec.

b. Random number, where the mean of E and Tc are 800W/m2 and 50oC, respectively, with variance=1.0

c. Repeating sequence: )(155.27)/(100450 2 CtTandmWtE o

c +=+= (3) where the periodic time (t) is 2 sec with 10 sec of time simulation

d. Uniform random number; Emin=100W/m2 and Emax=1000W/m2 and Tcmin=10oC and Tcmax=65oC.

The performance index (Jv) has been taken from [32] to measure the accuracy of proposed method. This index is calculated as follows:

−−===

N

iMPP

iMPP

N

i

iMPP

iopv VVVVJ

1

2

1

2 )()( (4)

where N is the number validation data set, i is the ith sample of data, VMPP is the ideal voltage at maximum power point, Vop is the estimated optimum voltage and MPPV is the average of VMPP.

The simulation results for VMPP estimation are presented in Fig. 3. The figures are only presented in the random number signal. This kind of signal is necessary to be considered following the randomly characteristic nature of the irradiance level. The results confirm that the RBF method is likely suitable to map between the optimum voltage and the input signals for 2j a-Si cell technology. On the other hand, ANFIS network is effectively used in the 3j a-Si technology. In other results, ANFIS and TFFN methods are giving similar responses in CIS technology; while RBF and TFFN methods are producing the same outcomes in thin-film CdTe technology. For random number signal, the deviation between the ideal and the optimum voltage is almost the same. From this result, a bias can be set in the ANN output to shift the optimum values to the ideal ones. For example, with the TFFN network for ST-40, the average deviation is 0.0709. This value can be used as the proposed bias. The overall performance index for all validation signals is shown in Table III.

In this study, three indicators are used to measure the selection of ANN structure in the application for non c-Si cells technology. They are the flexible of training process, the simplicity of network structure and the accuracy of validation error. To reach such kind of goals, the ANN network structures are denoted with ranks; high, modest and low. The overall result of this approach is shown in Table IV.

IV. CONCLUSION

This paper has presented different models of artificial neural network to deal with the VMPP of non crystalline Si PV modules. The trained configurations are verified using ramp, random, repeating sequence and uniform random signals of irradiance and cell temperature. The simulation results confirm that RBF and ANFIS methods have the flexible training process; while the TFFN method has simpler network structure than others. For the accuracy of validation error, RBF and ANFIS models are more suitable for 2j a-Si and 3j a-Si PV models, respectively. On the other hand, ANFIS and TFFN are effectively used in CIS technology. For thin film CdTe technology, the RBF and TFFN methods are the best option.

TABLE IIIPERFORMANNCE INDEX DURING VALIDATION PROCESS

PV modules ramp signal random signal repeating sequence signal uniform random signal

RBF ANFIS TFFN RBF ANFIS TFFN RBF ANFIS TFFN RBF ANFIS TFFN MST-43MV 0.035 0.096 0.065 0.258 0.982 0.992 0.035 0.113 0.070 0.035 0.092 0.065

US-32 0.143 0.007 0.059 2.474 0.036 0.418 0.128 0.007 0.059 0.146 0.008 0.060

ST-40 0.221 0.146 0.143 1.912 1.003 1.023 0.236 0.171 0.158 0.217 0.137 0.146 FS-50 0.063 0.115 0.081 0.205 2.009 0.868 0.071 0.133 0.064 0.067 0.112 0.080

TABLE IV EVALUATION PERFORMANCE OF ANN MODELS

ANN models

Flexibility Training process

Simplicity of network structure

Accuracy of validation error MST-43MV US-32 ST-40 FS-50

RBF high moderate high low moderate high ANFIS high low low high high moderate

TFFN moderate high moderate moderate high high

.a. Solarex MST-43MV (2j a-Si)

with RBF network b. USSC Unisolar US-32 (3j a-Si)

with ANFIS network

c. Siemens ST-40 (CIS) with TFFN network

d. First Solar FS-50 (thin film CdTe) with RBF network

Fig. 3. Validation results for ramp and random signals for all PV modules

V. REFERENCES

[1] A. Mellit, and S. A. Kalogirou, "Artificial intelligence techniques for photovoltaic applications: A review," Progress in Energy and Combustion Science, vol.34, no.5, pp.574-632, 2008.

[2] A. Mellit, S. A. Kalogirou, L. Hontoria, and S. Shaari, “Artificial intelligence techniques for sizing photovoltaic systems: A review,” Renewable and Sustainable Energy Reviews, vol.13, no.2, pp.406-419, 2009.

[3] F. Giraud, and Z.M. Salameh, “Analysis of the effects of a passing cloud on a grid-interactive photovoltaic system with battery storage using neural networks,” IEEE Transactions on Energy Conversion, vol.14, no.4, pp. 1572-1577, 1999.

[4] A. Mellit, M. Benghanem, and S.A. Kalogirou, “An adaptive wavelet-network model for forecasting daily total solar-radiation,” Applied Energy, vol.83, no.7, pp.705-722, 2006.

[5] G. Tina, and G. Capizzi, “Improved lead-acid battery modeling for photovoltaic application by recurrent neural networks,” in Proc. 2008 International Symposium on Power Electronics, Electrical Drives, Automation and Motion, pp.1170-1174.

[6] X. Sun, W. Wu, X. Li, and Q. Zhao, “A research on photovoltaic energy controlling system with maximum power point tracking,” in Proc. 2002 Power Conversion Conf., vol.2, pp.822-826.

[7] Syafaruddin, E. Karatepe, and T. Hiyama, “Polar coordinated fuzzy controller based real-time maximum power point control of photovoltaic system,” Renewable Energy, doi:10.1016/j.renene.2009.04.022, 2009.

[8] Syafaruddin, E. Karatepe, and T. Hiyama, “Artificial neural network polar coordinated fuzzy controller based maximum power point tracking control under partially shaded conditions,” IET Renewable Power Generation, vol.3, no.2, pp.239-253, 2009.

[9] A. de Medeiros Torres, F.L.M. Antunes, and F.S. dos Reis, “An artificial neural network-based real time maximum power tracking controller for connecting a PV system to the grid,” in Proc. 1998 the 24th Annual Conf. of the IEEE on Industrial Electronics Society,Vol. 1, pp.554-558.

[10] A. Yona, T. Senjyu, A.Y Saber, T. Funabashi, H. Sekine, and K. Chul-Hwan,” Application of neural network to 24-hour-ahead generating power forecasting for PV system,” in Proc. 2008 Power and Energy Society General Meeting - Conversion and Delivery of Electrical Energy in the 21st Century, pp.1-6.

[11] D. Paul, S. N. Mandal, D. Mukherjee, and S. R. Bhadra Chaudhuri,” Artificial neural network modeling for effficient photovoltaic system design, in Proc. 2008 International Conference on Advanced Computer Theory and Engineering, pp.50-56.

[12] A. Mellit, H. Mekki, and S. Shaari,” FPGA-based neural network for simulation of photovoltaic array: application for estimating the output power generation,” in Proc. 2008 Photovoltaic Specialists Conference,pp.1-7.

[13] S. Zhou, L. Kang, G. Guo, Y. Zhang, J. Cao, and B. Cao, ”The application of combinatorial optimization by Genetic Algorithm and Neural Network,” in Proc. 2008 the 3rd Industrial Electronics and Applications, pp.227-231.

[14] S. Zhou, L. Kang, G. Guo, Y. Zhang and B. Cao, “The combinatorial optimization by Genetic Algorithm and Neural Network for energy storage system in Solar Energy Electric Vehicle,” in Proc. 2008 7th

World Congress on Intelligent Control and Automation, pp.2838-2842. [15] J.A. Momoh, Y.Wang, and F. Eddy-Posey, “Optimal power dispatch of

photovoltaic system with random load,” in Proc. 2004 IEEE Power Engineering Society General Meeting, vol.2, pp.1939-1945.

[16] S. Chakraborty, and M.G Simoes, “PV-microgrid operational cost minimization by neural forecasting and heuristic optimization,” in Proc. 2008 IEEE Industry Applications Society Annual Meeting, pp.1-8.

[17] S.I. Sulaiman, T.K. Abdul Rahman, and I. Musirin, “Optimizing one-hidden layer neural network design using evolutionary programming,”in Proc. 2009 5th International Colloquium on Signal Processing & Its Applications, pp.288-293.

[18] P. Petchjatuporn, W. Ngamkham, N. Khaehintung, P. Sirisuk, and W. Kiranon, “A Solar-powered battery charger with neural network maximum power point tracking implemented on a low-cost PIC-microcontroller,” in Proc. 2005 International Conf. on Power Electronics and Drives Systems, vol. 1, pp.507-510.

[19] Y.Narri, and V. Mummadi, “Adaptive controller for PV supplied buck-boost converter,” in Proc. 1999 IEEE International Conf. on Power Electronics and Drive Systems, pp.789-793.

[20] K. Jae-Sub, J. Byung-Jin, P. Ki-Tae, C. Chung-Hoon, and C. Dong-Hwa, “Maximum power point tracking control of PV system for DC motors drive with neural network,” in Proc. 2008 International Conference on Smart Manufacturing Application, pp.514-519.

[21] B. Cao, L. Chang, and H. Li, “Implementation of the RBF neural network on a SOPC for maximum power point tracking,” in Proc. 2008Canadian Conference on Electrical and Computer Engineering, pp.981-985.

[22] A. Mellit, M. Benghanem, and S.A. Kalogirou, “Modeling and simulation of a stand-alone photovoltaic system using an adaptive artificial neural network: Proposition for a new sizing procedure,” Renewable Energy, vol.32, no.2, pp.285-313, 2007.

[23] A. Varnham, A.M. Al-Ibrahim, G.S. Virk, and D. Azzi, “Soft-computing model-based controllers for increased photovoltaic plant efficiencies,” IEEE Transactions on Energy Conversion, vol.22, no.4, pp.873-880, 2007.

[24] A. Mellit, and S. A. Kalogirou, “Neuro-fuzzy based modeling for photovoltaic power supply system,” in Proc. 2006 First International Power and Energy Conf. pp.88-93.

[25] M.G. Villalva, J.R. Gazoli, E.R. Filho, “Comprehensive approach to modeling and simulation of photovoltaic arrays,” IEEE Transactions on Power Electronics, vol.24, no.5, pp.1198-1208, 2009.

[26] H.G. Beyer, R. Gonschalg, T.R. Betts, and D.G. Infield, “Modeling the realistic short circuit current and MPP power of a-Si single and multijunction devices,” in Proc. 2003 3rd World Conf. on Photovoltaic Energy Conversion, pp.2435-2438.

[27] D.L. King, “Photovoltaic module and array performance characterization methods for all system operating conditions,“ Proceedings of NREL/SNL Photovoltaics Program Review Meeting,pp.1-22, 1997.

[28] D.L. King, “Sandia’s PV module electrical performance model,” Sandia National Laboratories, September 2000.

[29] A. Mellit, M. Menghanem, and M. Bendekhis, “Artificial neural network model for prediction solar radiation data: application for sizing stand-alone photovoltaic power system,” in Proc. 2005 IEEE Power Engineering Society General Meeting, vol.1, pp.40-44.

[30] K. Hyun-Soo, B.G. Morris, H. Seung-Soo and G.S. May, “A comparison of genetic and particle swarm optimization for contact formation in high-performance silicon solar cells,” in Proc. 2008 IEEE International Joint Conf. on Neural Networks, pp.1531-1535.

[31] Syafaruddin, E. Karatepe and T. Hiyama, ANN based real-time estimation of power generation of different PV module types, IEEJ Transaction on Power and Energy, vol.129, no.6, pp.783-790, 2009.

[32] E. Karatepe, and M. Alci, “A new approach to fuzzy wavelet system modeling,” International Journal of Approximate Reasoning, vol. 40, no.3, pp.302-322, 2005.