Application of Neural Networks Technique in Renewable Energy Systems

Transcription

1 2014 First International Conference on Systems Informatics, Modelling and Simulation Application of Neural Networks Technique in Renewable Energy Systems Lamine Thiaw, Gustave Sow, Salif Fall Renewable Energy Laboratory, Ecole Superieure Polytechnique / Universite Cheikh Anta Diop Dakar, Senegal lamine.thiaw.sn@ieee.org gustave.sow@ucad.edu.sn salif.fall@ucad.edu.sn Abstract Artificial Intelligence (AI) techniques are increasingly used in various area due to their capability of handling complex systems specificities. Among the techniques of AI, Artificial Neural Networks (ANN) technique plays an important role. This technique is used in this work to perform important tasks encountered in Photovoltaic systems and in Wind Energy Systems: a) Maximum Power Point Tracking (MPPT) of Photovoltaic Generators; b) and wind energy resource assessment. It is shown how a neural network technique can be used to design an MPPT controller for photovoltaic generators, enabling to improve their efficiency, and how it is possible to assess the available and recoverable wind energy potential of a site, by means of finding an adequate distribution law of the wind speeds based on a neural model. The proposed methods are illustrated by simulation results which exhibit the advantages of using ANN techniques in Renewable Energy Systems. Keywords Photovoltaic Systems; Wind Energy Systems; Neural networks I. INTRODUCTION Energy is an unavoidable vector of development for all countries. In Senegal, fossil energy consumption accounts for 38% of the total consumption of energy, which corresponds to about 54% of the country export incomes. However a great deal of efforts are made today for the development of renewable energies such as photovoltaic solar energy, wind energy, biomass, bio-fuels etc. In the field of electrical energy output, photovoltaic solar energy and wind energy are increasingly used. Wind potential is not high enough in Senegal (an average wind speed of 4 m/s to 5 m/s along the northeast coastal region from Dakar to St Louis, the windiest area) and is unequally distributed throughout the national territory. Concerning the solar energy potential, it s estimated to an average of 5.4 kw h/m 2 /d over all the national territory, with an average duration of 8 h/d, corresponding to a daily average irradiance of 675 W/m 2. The development of renewable energies requires the use of sophisticated techniques for an accurate estimation of the available energy potential and an effective control of systems operation. In the last years, Artificial Intelligence (AI) techniques are increasingly used in various area( [1] and [2]). They enable to study complex systems without any knowledge of the exact relations governing their operation. They are able to handle noisy and incomplete data, and once trained, allow performing as complex tasks as prediction, modeling, identification, optimization, forecasting and control. Among the various techniques of AI, Artificial Neural Networks (ANN) are frequently used. An ANN can be defined as a complex network composed of interconnected elementary processing units (neurons). Neurons are organized into layers and can be connected in different ways. The topology of connections between neurons determines the network architecture and is related to the the problem to be solved (non linear regression, classification, optimization, etc.). The network comprises parameters which are determined through a leaning process. There are various types of ANN. Multi Layer Perceptron (MLP) are feedforward neural networks commonly used in problems of nonlinear regression. An MLP network comprises an input layer, one or more hidden layers and an output layer. The neurons of a hidden layer receive information from the neurons of the previous hidden layer or from the inputs, and are connected to the neurons of the next layer or to neurons of the output layer. There is no connection between the neurons of the same layer. Each neuron of the output layer performs a nonlinear function of the inputs of the network (see [3], [4], [5] and [6]). An application of using neural networks in wind energy systems is illustrated in [7] where a hybrid neural network approach, comprising a Self Organizing Map (SOM) and a Radial Basis Function (RBF) neural network, is used to predict wind speed automatically. The approach enables wind speed prediction with less errors. In [8], a recurrent neural network is used for the control of a battery energy storage system accounting state of charge (SOC) and terminal voltage. In the present work, a simple method of designing an ANN based controller for maximum power point tracking of photovoltaic generators is presented. The ANN technique is also used to more accurately determine the wind speed distribution law of a site, enabling to better assess wind energy potential and wind generators performances. The paper is structured as follows: section II presents the design method of neural controller for Maximum Power Point Tracking of PV Generators; section III describes the implementation of ANN technique for wind energy ressource assessment; the last part presents the results and discussion. II. IMPLEMENTATION OF NEURAL NETWORK TECHNIQUE FOR MAXIMUM POWER POINT TRACKING OF PV GENERATORS PV generators efficiency is known to be low and its maximum value depends on operating conditions (irradiation, /14 $ IEEE DOI /SIMS

2 (a) Fig. 1. Models of PV generator: (a) Neural model; (b) Neural model which gives the maximum power point. (b) Fig. 2. Principle of duty cycle generation using the PV generator I-V characteristic. temperature and load values). To extract the maximum available energy, PV generators must operate at their maximum power point at any time, which requires a system control for tracking that point. Various techniques of Maximum Power Point Tracking (MPPT) have been investigated. Among these techniques, neural network based controllers are very often referred in the literature. A. Maximum Power Point Tracking Principles A Photovoltaic Generator can operate at various points depending on its output voltage or delivered current. For a given solar irradiation and ambiant temperature, there is one operating point for which a maximum power can be extracted from the generator. This is the optimal operating point defined by the optimal voltage and current values, V opt and I opt respectively. The principle of the maximum Power Point Tracking (MPPT) is to control the output voltage and the current delivered by the photovoltaic generator so that at any time, they are as close as possible to their optimal values. An MPPT algorithm is used to generate the appropriate control signal for a converter connected to the PV generator. Many algorithms exist for MPPT: Perturb and Observe Method (P&O), Incremental Conductance (IncCond), Hill Climpbing Method, Neural Networks based approach (see [9], [10], and [11]). B. Neural Model of a PV generator The neural model of a PV generator can be represented as shown in figure 1-a where S designates the solar radiation, T the PV cells temperature, V pv and I pv the PV generator output voltage and delivered current respectively. The main advantage of the neural model is that there is no need to know the exact relation between inputs and outputs. Additionally, in the case of PV generator modeling, input data such as solar radiation and cell temperature may be replaced by their equivalent electrical values obtained from measuring devices. It is known that for PV cells, solar radiation mainly influences the short-circuit current while ambient temperature (or cell temperature) influences the open circuit-voltage. That s why, instead of using expensive devices to sense real values of solar radiation and cell temperature, two PV cells (or two small PV panels) can be used. The first one will be shortcircuited and will give information about the solar radiation (I sc(m) ). The second one is in open circuit mode and will give information about cell temperature (V oc(m) ). For various load values, the I V characteristics of the PV generator can be obtained for each set of I sc(m) and V oc(m) of monitoring cells. Fig. 3. A stand-alone PV system using the neural network based controller for MPPT A neural model of the PV generator can be developed to enable finding the optimal operating point starting from I sc(m) and V oc(m). For a given set of these parameters, it is easy to compute the point corresponding to the maximum generated power from the I-V curve. The neural model is then trained with I sc(m) and V oc(m) as inputs and optimal values of PV generator voltage and current (V pv(opt) and I pv(opt) ) as outputs (see figure 1-b). C. Design of a Neural Network Controller for MPPT The purpose of the neural controller is to generate for the converter connected to the PV generator a control signal which enables the input voltage to be set as close as possible to the PV generator optimal voltage, starting from I sc(m) and V oc(m) values obtained from monitoring cells. For each set of I sc(m) and V oc(m), the optimal voltage value V pv(opt) is computed using the I V characteristic of the PV generator. This value of V pv(opt) is then used with the converter output voltage value (V out ) to compute the control signal. For a buck DC-DC converter functioning in a continuous conduction mode, the duty cycle of the control signal is expressed by the relation (1). D opt = V out V pv(opt) (1) The calculation of the duty cycle can be refined if it is necessary to take into account the imperfections of the system or if another converter is used. To generate a convenient database for the neural model training, for each value of V pv(opt) in equation (1), various duty cycle values are calculated using values of V out randomly chosen between desired minimal and maximal values of the output voltage. Figure 2 shows the principle of generation the duty cycle values needed for the neural controller training. The trained neural controller is used as indicated in figure 3 for a stand alone PV system. 7

3 III. IMPLEMENTATION OF A NEURAL NETWORKS TECHNIQUE FOR WIND ENERGY RESSOURCE ASSESSMENT The interest towards renewable energies implies a good assessment of the available wind energy potential. The wind potential assessment of a site requires the knowledge of the distribution law of the wind speed measured on the site. The statistical treatment of these measurements makes it possible to have a discrete distribution law. However, a more accurate analysis of the wind potential needs obtaining a continuous distribution law. The Weibull or Rayleigh models are often used. The approach consists in assimilating the distribution law to one of these models and to determine the model parameters so that it gets closest to the discrete law achieved by the statistical treatment of the wind speed measurements. Determining a distribution law for the speeds can be considered as a non linear regression problem, in which the distribution law chosen (Weibull or Rayleigh) is identified so as to get nearer the discrete law ( [12], [13], [14]). As regards function approximation, however, the techniques based on the artificial neural networks approach have shown that very good performances can be obtained ( [5], [15], [16]). A. Wind Resource and Wind Generators Production Assessment Principles To study the distribution of wind speeds, the two-parameter Weibull model defined by equation (2) is used very often: f(v) = k ( v ) [ k 1 ( v ) ] k exp (2) c c c where f( ) is distribution law or probability density ; v is the wind speed (m/s); c is the scale factor (m/s); k is the shape factor (characterizes the distribution dissymmetry). The average energy available for a duration period of T hours (in kw h/m 2 ) is assessed, for a continuous distribution law by the relation 3): E mt = T ρ 0 f(v) v 3 dv (3) where ρ is the air density. A wind generator is generally characterized by its power curve function of the speed P = P (v), its rated power P r, its starting V s, rated V r, maximum V c speeds. These characteristics make it possible to assess the available energy amount the wind generator can recover on a given site. For a duration period of T hours, this energy amount is assessed by the equation 4): E gt = T ρs V r V s f(v) v 3 3 dv + V r V c V r f(v) dv (4) The amount of energy a wind generator would produce on a given site during the period T can be assessed from the characteristics P = P (v) and the wind speed distribution law on the site by the relation (5): E gpt = T V c 1000 f(v) P (v) dv (5) Relations (3)-(5) enable the amounts of energy available, recoverable or produced monthly (T = h) or yearly (T = h). B. Assessing Wind Energy by Neural Networks Technique In this research work, an MLP neural network is used to determine the wind speed distribution law. The statistical study of wind speed measurements makes it possible to determine the frequencies of the speed contained in each speed interval I i =]V i 1,V i ], i =1...M, M being the index corresponding to the maximum speed recorded. For each interval I i, frequency f i is calculated by the expression: f i = N i (6) N in which N i is the number of speed values contained in I i. An MLP neural networks is then worked out to determine the speed distribution law f(v). The learning database consists of the couples of points (V i,f i ), where the V i are the inputs and the f i, the outputs. Determining the relation f(v) enables the amounts of energy available, recoverable or produced (cf equations (3), (4) and (5)) to be determined. The assessment of a wind generator production is made starting from the machine power curve, provided by the constructor in the form of couples of points (v i,p i ). The relation P = P (v) can be determined by using an MLP. The learning database consists of the points (v i, P i ) provided by the constructor. V s IV. RESULTS AND DISCUSSION A. Neural MPPT Controller for a PV Battery Charging System The stand-alone PV system represented in figure 3 is used for simulation. The system comprises a 1.33 kwp PV generator and a 24 V/150 Ah lead-acid battery as load. The neural model of the MPPT controller comprises three inputs (I sc(m), V oc(m) and V out ), one hidden layer with 4 neurons and one output neuron representing the optimal duty cycle D opt. Figure 4 shows variations of solar radiation, cell temperature and also the corresponding electrical values (short circuit current and open circuit voltage) obtained from monitoring cells. The tracking of the maximum power point is illustrated in figure 5 where actual and optimal values of the PV generator power are shown. The maximal value of the relative error of generated power, determined by equation 7 is less than 8

4 2% which confirm good tracking ability. Figure 6 shows the evolution of the battery voltage and current. err P = P opt P pv P opt 100% (7) Results show that for neural network based controllers, the optimal operating point of the PV generator is known in advance, starting from environmental conditions. The matter is to set the system to that operating point. This is not valuable for the other MPPT control techniques, where the optimal operating point is unknown and the problem is to search for the right direction to move, as environmental conditions vary. The latter approach sometimes results in oscillations around the optimal operating point or loosing it in the case of fast variations. The main drawback of neural network based controllers is the difficulty of taking into account the PV generator parameter change due to ageing. This leads to the necessity of retraining the neural model when the PV generator parameters are expected to have changed. Fig. 6. Evolution of the Battery voltage and current. B. Assessment of the Wind Energy Potential of the Dakar Site The wind energy assessment methodology proposed was applied on the site of Dakar, Senegal (latitude: N, longitude:17 27 W, altitude: 0 m). The available wind speed measurements, obtained at the Dakar Airport site, are instantaneous values recorded every three hours and cover a period of 10 years, from 1995 to These data are used here only for the purpose of the proposed method validation. More accurate wind energy assessment require more complete wind speed measurements. The speed variation slot has been divided into 18 intervals and in each interval, the frequencies have been determined by relation (6). Weibull model parameters for the site are: c =4.9 m/s and k =2.19. On the other hand, an MLP neural network with 1 input, 2 hidden layers including 3 neurons each and 1 output has been identified. Wind potential characteristics assessed by the use of these two models are then compared with those assessed using the discrete method. The discrete distribution law obtained by treating the measurements statistically as well as the continuous distribution laws corresponding to the Weibull model and to the identified MLP neural network are represented in figure 7. Fig. 4. Solar radiation, cell temperature and corresponding outputs of monitoring cells Fig. 5. Evolution of PV generator output power and its optimal value. Fig. 7. Wind speed distribution laws obtained by different methods. To verify that the relation obtained with the MLP network is a distribution law, the integral I v defined by relation (8) 9

5 has been calculated. The results gives I v =0.997, showing that the relation obtained can be considered as a distribution law (f must verify the condition f(v) dv =1). I v = V max 0 0 f(v) dv (8) The amount of energy available assessed with the relation (3) for a one-year period has been determined. Figure 8 shows the contribution of each speed to the amount of energy available annually. The assessments made with the MLP network are nearer the assessments obtained with the discrete law than those obtained with the Weibull law. The speed-duration curves have also been determined. Figure 9 shows that the results obtained with the MLP neural network are closer to the results obtained by means of the discrete law. The wind generator WES18 has been simulated with the site data. The characteristics of this wind generator. The powerspeed characteristics of the wind generator supplied by the constructor has been modeled by an MLP neural network. The characteristic supplied by the constructor and that obtained by the MLP model are presented figure 10. TABLE I ANNUAL ENERGIES AVAILABLE, RECOVERABLE AND PRODUCED BY THE WIND GENERATOR WES18 AT 10 m ABOVE THE GROUND. Energies Discrete law Weibull MLP available (kw h/m 2 ) recoverable (MWh) produced (MWh) Fig. 10. The Power-speed characteristic of the wind generator WES18. Table I gives the total amounts of energies available as well as those recoverable and produced by the wind generator WES18 at 10 m above the ground. V. CONCLUSION Fig. 8. Yearly available energies assessed with different methods. Fig. 9. Speed-duration curves obtained with different methods. Through this work, importance of using Artificial Neural Networks in Renewable Energy Systems is exhibited. A method for designing efficient neural controller for maximum power point tracking of PV generators is presented and simulated. The controller enables to find the optimal value of the DC-DC converter s duty cycle, starting from the sensed load voltage, the output values of the short circuit current and the open circuit voltage of monitoring cells which reflect environmental conditions. While classical MPPT controllers search for the location of the maximum power point, neural based controllers know its exact location and try to move the system to that point. There is no oscillations around the maximum power point, even if fast variations of environmental conditions occur. Obtained result from a stand alone PV system simulation show fast tracking performance of the designed controller with a low error. A neural-based approach enabling to assess the wind energy potential has also been proposed in the paper. The proposed neural approach enables to accurately determine a site wind energy characteristics: wind speed frequencies, wind speedduration curves, energies available, recoverable and produced by a given wind generator. A comparative study with the method using the Weibull model shows that better results are obtained with the help of the neural model. 10

6 ACKNOWLEDGMENT The authors would like to thank the UEMOA (West African Monetary Union) Commission who supports their research work and has financed their participation in the SIMS2014 Conference. REFERENCES [1] A. Mellita and S. A. Kalogiroub. Artificial intelligence techniques for photovoltaic applications: A review. Progress in Energy and Combustion Science, 34:574632, [2] H.Silver. Neural networks in electrical engineering. In Proceedings of the ASEE New England Section 2006 Annual Conference, [3] Haykin Simon. Neural networks: A Comprehensive Foundation [4] R.C. Eberhart and R.W. Dobbins. Neural Network PC Tools: A practical Guide. Academic Press, [5] K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators. Neural Networks, 2(5): , [6] D.E. Rumelhart, G.E. Hinton, and R.J. Williams. Learning representations by back-propagating errors. Nature (London), 323: , [7] K. S. Gnana Sheela and S.N. Deepa. An efficient hybrid neural network model in renewable energy systems. In Proc IEEE International Conference on Advanced Communication Control and Computing Technologies (ICACCCT), pages , [8] G. Capizzi, F. Bonanno, and C. Napoli. Recurrent neural networkbased control strategy for battery energy storage in generation systems with intermittent renewable energy sources. In Proc International Conference on Clean Electrical Power (ICCEP), pages , [9] V. Salas, E. Olias, A. Barrado, and A. Lazaro. Review of the maximum power point tracking algorithms for stand-alone photovoltaic systems. Solar Energy Materials & Solar Cells, 90: , [10] T. Esram and P.L. Chapman. Comparison of photovoltaic array maximum power point tracking techniques. IEEE Transactions of Energy Conversion, 22(2): , [11] A. Al-Amoudi and L. Zhang. Optimal control of grid connected pv system for maximum power point tracking and unity power factor. In Proc. Seventh Int. Conf. Power Electron. Variable Speed Drives, pages 80 85, [12] F. Ben Amar, M. Elamouri, and R. Dhifaoui. Energy assessment of the first wind farm section of sidi daoud, tunisia. Renewable Energy, 33: , [13] F. A. L. Jowder. Wind power analysis and site matching of wind turbine generators in kingdom of bahrain. Applied Energy, 86: , [14] A. Ucar and F. Balo. Evaluation of wind energy potential and electricity generation at six locations in turkey. Applied Energy, 86: , [15] G. Cybenko. Approximation by superposition of a sigmoidal function. Mathematical Control Signals Systems, 2: , [16] L. Thiaw, G. Sow, S.S. Fall, M. Kasse, E. Sylla, and S. Thioye. A neural network based approach for wind resource and wind generators production assessment. Applied Energy, 87: ,