A GRNN based algorithm for output power prediction

5 downloads 0 Views 437KB Size Report
KEY WORDS. Neural Networks, GRNN, Power Supervision, small PV .... M ×. Figure 1. GRNN architecture (MATLAB Neural Networks toolbox) .... Mathworks, Inc. Matlab Documentation Center, Neural Network Toolbox, User's Guide; 2013. [5].
2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

IC-AIRES2017

A GRNN based algorithm for output power prediction of a PV Panel K. KERBOUCHE *1, S. Haddad *2, A. Rabhi *3, A. Mellit *2, M. Hassan Ali *3, A. El Hajjaji*3

(1), Energetic Systems Modeling

Laboratory University of Biskra Biskra, Algeria

(2),Renewable Energy

(3),Modeling, Information and

Laboratory Department of Electronics University of Jijel Jijel, Algeria

Systems Laboratory UPJV Amiens, France

ABSTRACT In this paper we investigated the reliability of a GRNN algorithm for the power prediction of a PV panel in order to minimize the effect of fast changing of the meteorological conditions. An experimental database of meteorological data (irradiance and module temperature) as input and electrical measure (power delivered by PV Panel) as output has been used. A database composed of two sets 97 values each one is used for training and validating the proposed GRNN model. The data used to develop the proposed algorithm are attained during two separated days from a PV panel within the MIS-Lab of UPJV, France. According to the gained results the algorithm can help to predict real instantaneous power even during temporary change in meteorological data. KEY WORDS Neural Networks, GRNN, Power Supervision, small PV I.

INTRODUCTION The increasing energy demands combined with rising conventional fuel costs and environmental awareness have contributed to the emergence of renewable energy sources during the last decade. Photovoltaic (PV) systems have sustained a remarkable annual growth rate, driven by several factors including technological innovation, improved cost effectiveness and government incentives. From 2000 to 2011, the International Energy Agency (IEA) reports that global PV installed capacity increased from 1 GW to 67 GW. Nevertheless, significant constraints still hinder the large-scale integration of PV in the electricity mix. In particular, the unpredictability and variability of the solar energy cause major problems to the reliability and stability of existing grid-connected power systems. The intermittent nature of solar energy poses many challenges to renewable energy system operators in terms of operational planning and scheduling. PV output forecasting is therefore essential for utility companies to plan the operations of power plants properly so as to ensure the stability, reliability and cost effectiveness of the system. Predicting the output of photovoltaic systems is therefore essential for managing the operation and assessing the economic performance of power systems [1]. Artificial intelligence techniques are becoming useful as alternate approaches to conventional techniques or as components of integrated systems. They have been used to solve complicated practical problems in various areas and are becoming more popular nowadays. They can learn from examples, are fault tolerant in the sense that they are able to handle noisy and incomplete data, are able to deal with nonlinear problems and once trained can perform prediction and generalization at high speed. AI-based systems are being developed and deployed worldwide in a wide variety of applications, mainly because of their symbolic

2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

IC-AIRES2017

reasoning, flexibility and explanation capabilities. Artificial intelligence has been used in different sectors, such as engineering, economics, medicine, military, marine, etc. They have also been applied for modeling, identification, optimization, prediction, forecasting and control of complex systems [2]. We aim in this work to develop an algorithm that can predict the output power of a solar panel using meteorological data only. Thus, we installed the system with sensors, collected data and used them to train our algorithm so it can predict the power once we give it the module temperature and the solar irradiance. II.

GENERALIZED REGRESSION NEURAL NETWORK MODEL

A GRNN is a variation of the radial basis neural networks (RBFNs), which is based on kernel regression networks. A GRNN does not require an iterative training procedure as back propagation networks. It approximates any arbitrary function between input and output vectors, drawing the function estimate directly from the training data. In addition, it is consequent that as the training set size becomes large, the estimation error approaches zero, with only mild restrictions on the function [3]. Typically, the purpose of training is to make estimations for future cases in which only the inputs to the network are known. The result of conventional network training is a single set of weights that can be used to make such estimations. The GRNN is one of the simplest neural networks in term of network architecture and learning algorithm. The advantage is that the learning is instantaneous. GRNN is based on one-pass learning algorithm; it is a highly parallel network of radial basis. GRNN is composed of input layer, radial basis hidden layer and linear output layer; its architecture is shown in Fig.1 [4]. It is similar to the radial basis network, but has a slightly different second layer. In Fig. 1, the number of the units in the hidden layer is equal to the training sample size M, and the weight function of this level is the Euclidean distance measuring function ||dist||, its function is to calculate the distance between the input of the network and the hidden layer weight matrix IW1, 1, b1 is the threshold of the hidden layer. Symbol (.*) in Fig. 1, indicates the product element per element of the output of IW1,1. The result of the product n1 is the net input of the transfer function. Hidden layer transfer function is the Radial Basis Function (RBF), Gaussian function is often used. The weight function of the output layer is the standardization of the right point multiplication function (with nprod denoted). n2 indicates the vector of the network; each element of it can be valued first by making dot-product between a1 and the element in each line of the matrix LW2,1 then doing the division between the result and the sum of every element of a1. Finally, the results of n2 is offered to the linear transfer function a2=purlin(n2) to calculate the network output. Fig. 2 shows the linear transfer function, purelin(n) and radial basis transfer function, radbas(n). The estimation model takes the following form:

P  f ([T , G],W )  

(1)

Where P is the vector of model outputs (PV power output ); [T,G] is the vector of model inputs (module temperature Radial Basis Layer

Special Linear Layer

M  N IW1,1

P dist

N 1

b1 N

a (i )  radbas a (i ) is the i

LW2,1

M 1

n1

*

a1

th



M 1

M 1

M 1

1 1

M M

nprod

n2

a2  Y

M 1

M 1

M

IW1,1 (i )  p b1 (i )



M

a  purelin(n 2 ) 2

1

element of a where IW1,1 (i) is the vector made of the i th row of IW1,1 (i )

Figure 1. GRNN architecture (MATLAB Neural Networks toolbox)

2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

IC-AIRES2017

and irradiance); W is the vector of model parameters (connection weights), f(•) are the functional relationship between model outputs, inputs and parameters, and ɛ is the vector of model errors.

radbas(n)

purelin(n)

1

1 0.8

0.9

0.6

0.8

0.4

radbas(n)

purelin(n)

0.2 0 -0.2 -0.4

0.7

0.6

0.5

-0.6

0.4 -0.8 -1 -1

-0.8

-0.6

-0.4

-0.2

0 n

0.2

0.4

0.6

0.8

1

-1

Figure 2. Linear transfer function, purelin(n)

-0.8

-0.6

-0.4

-0.2

0 n

0.2

0.4

0.6

0.8

Fig.3. Radial basis transfer function, radbas(n)

The ANN was simulated in MATLAB using ‘newgrnn’ function, which creates a new GRNN (two-layer network). As part of the ANN model development process, the available data are generally divided into training, testing and validation subsets. The training set is used to estimate the unknown connection weights, the testing set is used to decide when to stop the training process in order to avoid over-fitting and/or which network structure is optimal, and the validation set is used to assess the generalization ability of the trained network. III.

SYSTEM DESCRIPTION

In order to conduct a comprehensive study of the relationship between PV output power and the meteorological conditions (irradiance and temperature) a system has been installed in the MIS-Lab of UPJV, France. The PV system selected for the present study consists of five main components namely a Sanyo 215 W solar panel, meteorological measurement devices (thermocouple and pyrometer), Electrical current and voltage sensors, a communication board, a programmable electric load and a PC equipped with a Labview and MATLAB interfaces as shown in figure (4) and described in [5].

Figure 4. Experimental set [5]

1

2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

IC-AIRES2017

The data acquisition process was established during a day from 9:15 until 17:15 with a measurement interval of 5 min which effect a database of 97 values for each parameter (Temperature and irradiance) then a similar database of maximum power was calculated from the current and voltage measurements data. The experiment was established again in the most typical irradiation day (no clouds) in order to compare between the two sets of data results IV.

SYSTEM DESCRIPTION

In this section we will describe the steps and data used to train and validate the developed algorithm A.

THE TRAINING PROCESS

In this work we choose the General Regression Neural Networks (GRNN) because it does not require iterative training, so it is simple, stable with fast training speed and good description for the characteristics of dynamic. Moreover, the parameters need to be adjusted has only one which determines the network can maximally avoid the influence to the results caused by man-made subjective assumptions [6]. First the Algorithm has been trained with the typical day database as following: -

input parameters are the module temperature and the solar irradiance (97 values for each victor).

-

output parameter is the maximum extracted power calculated using current and voltage databases (97 values).

Then, the second database of another day is used to test the performances of the algorithm. B.

THE INPUTS

The figures 5 and 6 show the input parameters for the two days. The first day is a typical day in terms of meteorological conditions (irradiation and temperature)

Figure 5. First meteorological input (temperature)

2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

IC-AIRES2017

Figure 6. Second meteorological input (irradiation)

The Figure 7 shows the maximum produced power obtained from the calculated power based on the measured currents and voltages

Figure7. Measured PV panel output power

2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

V.

IC-AIRES2017

RESULTS AND DISCUSSION

After training the developed algorithm using the first day database. We used the second day’s database to test its efficiency then the MSE error between the two graphs is calculated. The figure 8 shows both measured power and predicted power using the developed algorithm. The RMS calculated between the is: 435.4299 The above results show the efficiency of this Algorithm in predicting the output power of PV panel based on meteorological data only

Figure 8. Measured and predicted output power

VI.

CONCLUSION

In this work, A GRNN algorithm have been successfully developed for predicting the output power of a 215W PV panel installed on the MIS-Lab of UPJV, France. It has been demonstrated that the algorithm is able to predict the output power of the Panel based on available solar irradiance and temperature with good accurate results. The advantage of the proposed GRNN algorithm is that it does not require much calculations and use only meteorological data. Further works can use the advantages to use this algorithm to implement it for real time power supervision for diagnosis aims

ACKNOWLEDGMENT The authors would like to thank MIS-Lab of UPJV, France, for providing the facilities and databases to conduct this research.

2017 International Conference on Artificial Intelligence in Renewable Energetic Systems

IC-AIRES2017

REFERENCES [1]. Pamela Ramsami, Vishwamitra Oree " A hybrid method for forecasting the energy output of photovoltaic systems", Energy Conversion and Management Journal, Vol. 95, 2015, pp. 574–632. [2]. Adel Mellit, Soteris A. Kalogirou, " Artificial intelligence techniques for photovoltaic applications : A review", Progress in Energy and Combustion Science Journal, Vol. 34, 2008, pp. 406–413. [3]. Firat M, Gungor M, " Generalized Regression Neural Networks and Feed Forward Neural Networks", Advances in Engineering Software Journal, Vol. 40, 2009, pp. 731–737. [4]. Mathworks, Inc. Matlab Documentation Center, Neural Network Toolbox, User’s Guide; 2013 [5]. N. Boutana, A. Mellit, S. Haddad, A. Rabhi, A.Massi Pavan "An Explicit I-V model for photovoltaique module technologies", Energy Conversion and management Journal, Vol. 138, 2017, pp. 400-412. [6]. Xianyan Kuang, Lunhui Xu, Yanguo Huang, Fenglei Liu "Real-time Forecasting for Short-term Traffic Flow Based on General Regression Neural Network", Proceedings of the 8th World Congress on Intelligent Control and Automation, July 6-9 2010, Jinan, China [7]. W. Chine, A. Mellit, V. Lughi, A. Malek, G. Sulligoi, A.Massi Pavan " A novel fault diagnosis technique for photovoltaic systems based on artificial neural networks", Renewable Energy Journal, Vol. 90, 2016, pp. 501-512. [8]. Wei Zhou, Hongxing Yang, Zhaohong Fang, " A novel model for photovoltaic array performance prediction", Applied Energy Journal, Vol. 84, 2007, pp. 1187–1198. [9]. Amit Kumar Yadav, S.S. Chandel, " Solar radiation prediction using Arti fi cial Neural Network techniques: A review", Renewable and Sustainable Energy Reviews Journal, In press article. [10]. Atika Qazi, H. Fayaz, A. Wadi, Ram Gopal Raj, N.A. Rahim " The artificial neural network for solar radiation prediction and designing solar systems: a systematic literature review", Cleaner Production Journal, doi: 10.1016/j.jclepro.2015.04.041. [11]. Adel Mellit, Alessandro Massi Pavan, " Performance prediction of 20 kW p grid-connected photovoltaic plant at Trieste (Italy) using artificial neural network", Energy Conversion and Management Journal, Vol. 51, 2010, pp. 2431–2441. [12]. Kerim Karabacak, Numan Cetin, " Artificial neural networks for controlling wind–PV power systems: A review", Renewable and Sustainable Energy Reviews Journal, Vol. 29, 2014, pp. 804–827 [13]. Soteris A. Kalogirou, " Artificial neural networks in renewable energy systems applications: a review", Renewable and Sustainable Energy Reviews Journal, Vol. 05, 2001, pp. 373–401 [14]. Sandhya Samarasinghe, "Neural networks for applied sciences and engineering ", Auerbach pub, 2nd edition 2009, ISBN 978-0-8493-3375-X. [15]. Kerim Karabacak, Numan Cetin, " Artificial neural networks for controlling wind–PV power systems: A review", Renewable and Sustainable Energy Reviews Journal, Vol. 29, 2014, pp. 804–827 [16]. William J. Palm III, introduction to Matlab 7 for engineers. New York: MC Craw Hill; 2000. [17]. Mathworks, Inc. Matlab Documentation Center, Neural Network Toolbox, User’s Guide; 2013

Suggest Documents