The Research of PV MPPT based on RBF-BP Neural Network Optimized by GA

Transcription

1 nternational Conference on Logistics Engineering, Manageent and Coputer Science (LEMCS 215) The Research of PV MPPT based on RBF-BP Neural Network Optiized by GA Jian Liu iaolin u @qq.co * Corresponding Author Tong Li @qq.co Meiyan Cao li @qq.co Abstract- n order to track noinal output power of photovoltaic (PV) battery effectively and consider the feature of non-linear output, researchers present a Radial Basis Function and Back Propagation (RBF-BP) cobination neural network based on genetic algorith (GA) optiization to use in PV axiu-power-point-tracking (MPPT).First, cobination of double hidden layer RBF- BP neural network is presented by researching the output feature of PV battery. n order to predict the axiupower-point of PV battery ore accurately, GA is used to optiize cobination neural network. lluination and teperature which is the ain factors influencing the output of PV battery are treated as input to construct the prediction odel, and siulate the odel through MATLAB. Siulation shows that the syste has advantages which increase the accuracy and efficiency of tracking the output axiu-power-point-tracking of PV battery effectively of high tracking accuracy, high speed rate and little iteration. Keyword- Photovoltaic (PV) battery; Maxiu-powerpoint-tracking (MPPT); genetic algorith (GA); Radial Basis Function and Back Propagation (RBF-BP); MATLAB siulation. NTRODUCTON As traditional energy consuption brought ore serious environental probles coupled with the developent of odern society, power quality requireents are high [1]. Solar energy as a new, environentally friendly, renewable energy has got ore and ore attention [2]. PV power generation which is directly converted solar energy into electrical energy has quickly becoe a renewable energy technology, which has the potential for large-scale developent in the future and it ay becoe an iportant alternative energy source [3]. n order to iprove the efficiency of PV power generation, axiu power point tracking for PV power generation is particularly necessary. n recent years, doestic and foreign scholars have proposed a variety of axiu power point tracking algorith, such as perturbation and observation control [4], conductance increent ethod [5] and neural network ethod [5,6] and so on. Though having good results in a faster rate of change of the light situation, perturbation and observation control require higher precision of the sensor and long running tie. The disadvantage of conductance increent ethod is the step length is fixing. When the step length is too sall, the prediction result of PV is easy to reain in the low power output region. When the step length is too large, the easureent results will significantly oscillate. n contrast artificial neural network ethod not only has a strong ability of nonlinear function approxiation, but also does not require the physical paraeters of the PV battery in the establishent of the odel. BP [6] and RBF [7] are widely used around the existing artificial neural network. BP neural network has strong self-learning ability, but the network ay isleading by partial saple resulting in the deviation of learning outcoes; RBF neural network has faster convergence rate, but it is lack of generalization ability. This paper proposed a cobination of double hidden layer RBF- BP neural network with both advantages, and optiized this neural network by PV. Siulation and analysis through the MATLAB, cobination neural network based on PV optiization increase the speed and accuracy of prediction is verified.. PV CELL MODELS AND CHARACTERSTCS A. The atheatical odel of PV cells According to the theory of electronics, equivalent circuit of solar battery is shown as Figure 1 [8]. ph Rsh Figure 1. Solar battery equivalent circuit R s V 215. The authors - Published by Atlantis Press 1376

2 The equivalent atheatical description of PV cells as follows: q( V Rs ) V Rs ph exp AKT 1 Rsh () Where, is the battery output current and V is the output ph voltage; is Photo production current and is reverse saturation current inside the battery equivalent diode p-n 19 junction ;q is electronic charge which is C ;A is ideal factor of diode PN junction; K is The boltzann 23 constant which is J / K and T is working teperature; R is series resistance and resistance. R sh is parallel B. The output characteristics of PV cells Since the output of PV battery is affected by ultiple factors such as teperature, light intensity, and its output characteristic presents obvious nonlinear. This paper ainly focuses on considering the influence of the teperature and light intensity of PV panels for PV cells output. According to the atheatical odel of the PV cells, this paper selects Solarex MS6 6 w battery teplate MATLAB siulation of electrical paraeters ( V OC =21V, sc =3.74A, V =17.1V, =3.5A, P =59.9W) then researchers got the P - V curves of the PV cells as Figure 2 and Figure 3 P/W S=1W/? T=5 T=25 T= U/V Figure 2. P - V curves when T change and S is constant P/W S=1W/? S=85W/? S=7W/? T= U/V Figure 3. P - V curves when S change and T is constant Through analysis Figure 2 shows the output of the PV cells as the typical nonlinear. Under the condition of light intensity is constant, the teperature is higher, the peak value of the output power is saller. When under the condition of constant teperature, light intensity is greater, the peak of the output power is greater. The siulation results shows that the output of the PV cell is a typical nonlinear, and under different working conditions, the output power with the increase of the output voltage to the change trend of nonlinear first becoes big and then sall [9]. Therefore, in order to iprove the output efficiency of PV cells, it is particularly necessary for accurate fast tracking of axiu power point of PV cells.. THE RESEARCH OF MPPT METHODS BASED ON RBF-BP NEURAL NETWORK OPTMZED BY GA A. RBF-BP neural network This paper present a RBF-BP neural network which is cobined with advantages of strong learning ability, high fitness, converging fast of RBF neural network and high group classification perforance,building a cobination of double hidden layer RBF- BP neural network cobined with high fitness, converging fast and High precision. RBF-BP neural network structure is shown as Figure N1 Wi i N2 W j N3 Wjk Figure 4. RBF-BP network structure As shown in Figure 4, Saple data is preferentially trained by RBF neural network and the training result as input of the BP neural subnet. Network s initial weights and threshold value are rando nubers between (, 1).When the training accuracy did not reach the goal, the neural network can reverse change network weights and threshold until the training results eet the accuracy requireents.when the precision of the training results reach the expected value, end the training. B. RBF-BP neural network with GA optiized Due to the cobination of RBF-BP neural network training data when the initial weights and thresholds for the rando nuber between (, 1), in order to eet the requireents of the accuracy of the anticipated target, RBF-BP neural network to reverse changes according to the actual situation of the study result weights and thresholds, the process will undoubtedly reduce the efficiency of the whole network prediction. n order to solve the initial weights and threshold of uncertainty on the efficiency of the cobination of RBF- BP neural network prediction, this paper present using the cobination of GA for RBF-BP neural network for further optiization as Figure 5. k N4 y1 yk 1377

3 nput pretreatent GA for the initial value encoding Deterine the structure of RBF- BP neural network cobination of RBF-BP neural network to PV MPPT, avoid uncertainty due to network the initial weights and threshold of PV forecasting, this design uses GA to optiize the initial threshold and weight of RBF-BP neural network ethods: RBF-BP neural network training error as fitness value N Select Crossover utation Calculation of fitness value eet conditions end Y The initial of RBF-BP neural network length threshold and weights To obtain the optial weight and threshold Calculation error update weight and threshold eet conditions end Figure 5. RBF-BP network process with GA optiized RBF-BP neural network with the GA optiized is divided into RBF-BP neural network structural and GA optiization fro Figure 5.Each individual in the saple population contains the weight and threshold of the neural network..ga optiizing parts find a fitness which is the absolute value of the error between the output and the desired output as a benchark,using GA on the saple data individual species selection, crossover and utation to find the individual corresponding to the optial fitness, and deterine the optial initial threshold and weight is assigned to the neural network. C. The research of the neural network optiized of PV MPPT The axiu output power of PV battery is affected by ultiple factors such as teperature, sunshine intensity.according to the actual situation in this paper, researchers focus on considering the teperature of the PV panels,light intensity and the output voltage of axiu power point as the ain input and output paraeters of neural network training.rbf-bp neural cobined with input layer, hidden layer and output layer.nput layer: without considering the coplex weather conditions and power board under the condition of uneven illuination, for PV axiu power prediction of ain consideration PV cells easured when working teperature and light intensity as two input nodes.hidden layer:rbf neural network is adopted in cobination of RBF-BP neural network hidden layer and its output as the input of the BP neural network, thus set a RBF neural network node nuber is 9, BP neural network node nuber is 5.Output layer: in order to reflects the axiu power output of PV battery point effective, researchers select output voltage of axiu power point of PV cell to be only. n order to iprove the efficiency and accuracy of the Y End N 1) Researchers put part of the training saple data into RBF-BP neural network, and get the error between the actual output and the expected output for fitness value of the GA. 2) According to the nuber of input and output paraeters of RBF-BP neural network, researchers deterine the nuber of the threshold and the weights. Then researchers deterine the length of the individuals of GA 3) Optiization of genetic algorith for training data, the training saple of each individual contains the RBF-BP weight value and threshold value of BP neural network, individual coding ethod for real nuber encoding. Each individual is a real nuber. ndividual fitness value of the individual is calculated by GA fitness function calculation.then: (1) Select Researchers choose roulette wheel ethod,such as Forula (2) fn k / Fn fn pn (2) M fn i1 p., choice probability of every individual n in A n population; f n,the reciprocal of fitness of individual n; k, coefficient; M, Nuber of individual species (2) Crossover n this paper, researchers use the real crossing ethod, crossover of the first K chroosoe and A chroosoe 1 in j, as shown in Forula (3): kj kj (1 b) ljb (3) lj lj (1 b) kjb B A rando nuber between the [, 1]. (3) Mutation Researchers select the j genes of the individuals i to utation, such as Forula (4): ( ax )* f ( g) a.5 (4) ( in )* f ( g) a.5 ax,upper liit of gene ; in,lower liit of 2 gene ; f ( g) r(1 g / Gax ) ;r,a rando nuber; g, current iteration nuber; Gax axiu evolutionary ties; a, a rando nuber between the [, 1]. The initial weights and thresholds which are best adaptive individuals obtained by GA selection, crossover and utation of RBF-BP neural network are k 1378

4 assigned. Then researchers predict the axiu output power of PV cells by training RBF-BP neural network. V. SMULATON RESULTS AND ANALYSS Researchers achieve the training and the iterative process of RBF-BP neural network optiized with GA by MATLAB R29a. There is a total of 144 sets of experiental data of the 5in of the PV panels, the teperature and the intensity of the light intensity in the day during 6h to 18h fro the reference [1]. Test data is 24 groups of the power plate teperature and light intensity collected of every 3in acquisition fro 6h to 18h in next day. Prograing and setting the cobination of RBF-BP neural network with GA optiized through MATLAB, cobination of RBF-BP neural network training uses trainl function, hidden layer of RBF and BP are adopted radbas function and Tansig function as the transfer function. Transfer function of the output layer uses purelin function. The evolution iteration nuber of GA is 1; the crossover probability is.4.the utation probability is.1.the axiu nuber of training ties of the network is set to 1 ties. The expected error is set to.1 and the learning rate is set to.1. The process of generating the individual which suit optial fitness fro the saple data through GA selection, crossover and utation as Figure 6 fitness The average fitness Evolution algebra Figure 6. The process of generating the individual which suit optial fitness The siulation results of Figure 7 (a),(b) are training error curve of the RBF-BP neural network with GA optiized and cobination of RBF-BP neural network.obviously, though lack of training data and saple data, under the sae accuracy requireent, cobination of RBF-BP neural network need training 44 epochs and cobination of RBF-BP neural network with GA optiized which iprove the efficiency need only 39 epochs. Mean Souared Error(se) Epochs (a) Training error curve of RBF-BP neural network with GA optiized Mean Souared Error(se) Epochs (b) Training error curve of RBF-BP neural network Figure 7. Training error curve of Neural network Square error of cobination of RBF-BP neural network with GA optiized and RBF-BP neural network were and fro the MATLAB MSE function of Figure 7(a) and (b). By contrast, prediction accuracy and efficiency of cobination of RBF-BP with GA optiized neural network for axiu power output of PV battery point has iproved obviously. n order to reflect the accuracy of the cobination of RBF-BP neural network with GA optiized intuitively, Figure 7 shows 24 set of the actual error distribution of test saple point. The error of the axiu power point voltage of values by network predicted is between.2 ~.2 fro Figure8. t is showed that the cobination of RBF-BP neural network with GA optiized is used to track the axiu power point of P V cells with high accuracy. 1379

5 Voltage error of axiu power output The actual error Test saple point Figure 8. Test error Figure V. CONCLUSONS PV array operating teperature and optical light intensity are selected as input of RBF-BP cobined neural network without considering the coplex weather conditions in prediction odel of this paper and ake optiization on cobined neural network using GA at the sae tie. Therefore, tracking speed and accuracy of MPPT of PV cells are iproved. Results of siulation also show that optiized RBF-BP cobined neural network by GA is better than no optiized cobined neural network both on the nuber of iterations and the accuracy of prediction. Therefore, the reliability of the RBF-BP cobined neural network using GA optiized is proved and the proble of MPPT of PV array is effectively solved. V. REFERENCE [1] Fang liang-fei.research on the control strategy of photovoltaic energy storage [D]. HeFei University of Technology,21. [2] Zhao zheng-ing,liu jian-zheng,sun xiao-ying.solar photovoltaic power generation and its application[m].being:science Press,25. [3] Zhu yan-wei,shi xin-chun,dan yang-qing.et al.application of the particle swar optiization algorith in ulti peak axiu power point tracking of photovoltaic array [J][J]. Chinese Journal of electrical engineering,212,32(4): [4] Feia N, Petrone G, Spagnuolo G, et al. Optiization of perturb and observe axiu power point tracking ethod[j].eee Transactions on Power Electronics,25,2(4): [5] Yusof Y, Sayuti S H,A bdul Latif M, et al. Modeling and siulation of axiu power point tracker for PV syste[c].proceedings of National Power and Energy Conference. Kuala Lupur, Malaysia,24: [6] Liang xing-yan,zhang wei.research of MPPT based on BP neural network for photovoltaic power generation syste [J]. anufacturing autoation,212, (24): ,13. [7] Li tan,dong hai-ying,yang lei.photovoltaic power generation technology of RBF neural network based on ant colony algorith MPPT[J]. neural network,214, 38(9): [8] Zhang jian-po,zhang hong-yan,wang tao,et al.siulation Research on the axiu power tracking algorith in photovoltaic syste [J]. coputer siulation,21,27(1): [9] Fu wang,luo shi-wu,qing zhi-ing.research on the atheatical odel of PV array under the condition of partial shading [J]. coputer siulation 213,3(7): [1] Guo liang.maxiu power point tracking of photovoltaic syste based on particle swar optiization BP neural network [D]. Southwest Jiao Tong University,