Short-term prediction of photovoltaic power based on GWPA - BP neural network model

Transcription

1 Short-term preiction of photovoltaic power base on GWPA - BP neural networ moel Jian Di an Shanshan Meng School of orth China Electric Power University, Baoing. China Abstract In recent years, ue to China's energy crisis an environmental pollution an other increasingly serious, so in recent years to evelop renewable energy, solar energy has become one of China's major green energy, an solar photovoltaic power generation has become a solar energy use Main way. However, the photovoltaic power generation is becoming more an more ranom ue to the influence of meteorological factors, an the callability is wea, which maes the photovoltaic power generation forecast more important for the ogri control strategy, the gri power ispatching an the improvement of the gri power quality. In this paper, we improve the convergence spee an global convergence of wolves algorithm by improving the wolves algorithm (WPA), an propose a GWPA-BP photovoltaic power generation forecasting moel base on wolf algorithm optimization BP neural networ. Finally, the valiity of the propose metho is verifie by comparing the preicte ata with the measure ata. Keywors photovoltaic power preiction; wolves algorithm; neural networ. 1. Introuction Base on the numerical ata or the measure ata, the photovoltaic power preiction is base on the parameterization scheme of the geographical coorinates of the photovoltaic power plant an the specific regional characteristics. The preiction moel an algorithm are establishe to forecast the output power of the PV power plant in the future [1]. From the experience of foreign inustry research, photovoltaic power preiction from the forecast metho can be ivie into two ins of statistical methos an physical methos [2~3]. Statistical methos for statistical analysis of historical ata to fin out the inherent law an use to preict; physical methos base on solar raiation transfer equation, PV moules operating equations an other physical equations to preict the nee for etaile technical information of photovoltaic power plants an weather an solar raiation ata. At present, the methos use to preict the photovoltaic power are mainly use in neural networs, support vector machines an other statistical methos, which is the most wiely use BP neural networ. However, ue to the traitional BP neural networ in the training process nees to repeately moify the networ weights an threshols, often nee hunres or thousans of iterations it is possible to complete the training to achieve the target error, which le to a larger amount of calculation. In the practical application, if the initial weight of BP neural networ, the threshol is not properly selecte, then prone to BP neural networ convergence spee is slow, easy to fall into the local extreme value of the situation. Therefore, this paper proposes a preiction moel (WPA-BP) base on the wolves algorithm to optimize the BP neural networ. This metho can avoi the blinness of the original selection of networ weights an threshols in orer to achieve the efficiency an precision of short-term power preiction. However, ue to the basic wolves algorithm in the iterative upate process, will irectly eliminate the poor egree of poor iniviuals, while ranomly generating new iniviuals, this is not conucive to 86

2 the rapi implementation of the algorithm. Therefore, this paper proposes an improve wolf algorithm base on genetic algorithm operation, calle GWPA algorithm, an applies the propose algorithm to the training of BP neural networ, an finally through the optimize BP neural networ to carry out the photovoltaic power preiction. 2. Wolves algorithm an its improvement 2.1 Wolves algorithm principle The wolves algorithm is propose by simulating the process of prey on wolves. It consists of prey, wolf, wolf an wolf four basic parts, an abstracte out of three acts: wal, summon an siege, an "winner is ing" of the first wolf mechanism an "strong survival "Wolves upate mechanism of these two mechanisms [4]. The steps of the wolves algorithm are as follows: 1) Initialize. The population size of the wolves is, the initial evolutionary algebra t = 0, the iniviual imension of each wolves is the D imension variable, that is, the spatial imension of the search is D, L an U are the minimum maximum range of the wolves search space, the maximum number of iterations of the algorithm is M max, then the position of the i-th artificial wolf can be expresse as: ( (1), (2),... ( D)), i 1,2,..., Where, t i i i i i t i = L + ran ( U - L ), i 1,2,..., (1) is the ranom number in the i-th artificial wolf, ran (0, 1) in the t-th population. 2) Accoring to the size of the fitness value, generate the wolf. Through the concrete fitness function to calculate the fitness value of each artificial wolf, select the optimal fitness value of the artificial wolf as the wolf. 3) Wal behavior. In aition to the first wolf, select the best S _ num artificial wolf as a wolf an perform waling behavior, until a wolf etect the prey oor concentration greater than the wolf's lea prey oor concentration or reach the maximum number of wals i in the -imensional space is as follows: i T max i. The location of the wolf p sin 2 p / h step (2) i i a In the formula, h represents the number of irections in which the wanerer wals, stepa represents the waling step of the wolf, an represents the current irection ( p 1,2,..., h). 4) Rai behavior. The wolf through the howling initiate summons behavior, calle aroun the wolf quicly gathere to the hea of the wolf. Wolves accoring to type (3) the implementation of rais, if the way in the rai, the wolf perceive prey oor concentration i lea, then, the wolf i instea of the wolf into a new wolf, an re-initiate summon, otherwise, the wolf continue to rai until is < [5]. When the wolf in the +1 evolution, in the -imensional variable space in which the location as shown in equation (3): p step 1 i i b ( g ) / i g i lea = (3) Where step b is the length of the rai of the wolf, g for the generation of wolves in the (=1,2,..., D) imensional space position. Determine the istance from the formula (4) calculate: Where, w is the istance etermination factor. D 1 U L near D w 1 (4) i 87

3 5) Siege behavior. We loo at the position of the wolf as the prey of the mobile position, for the wolves of the generation, assuming that the prey of the mobile position, the wolves siege behavior can be expresse by (5) 1 i i c i G step G (5) 1,1 Where is a ranom number between ; is the attac step size; if the wolves are perceive after the siege of the prey oor concentration is greater than the original state of the state of the oor concentration, then upate the location of artificial wolves, otherwise, the location of the artificial wolf will remain unchange. Will be the greatest value of the artificial wolf selecte as the first wolf. Wal step size step a, rai step step b, attac step step c step c the relationship between the three as follows: U L step step / 2 2 step / S (6) a b c Where, S is the step factor. 6) Wolves upate mechanism. Wolves in accorance with the "survival of the fittest, from strong to wea," the principle of istribution of foo, that is, the worst egree of fitness out of the R row of wolves, an then ranomly generate artificial wolf, where is a ranom integer between n/ 2, n/ an is the population upate scale factor. Determine whether the result of the algorithm satisfies the accuracy requirement, or whether the maximum number of iterations is reache. If it is reache, the position of the wolf is the optimal solution of the problem, an the algorithm ens; otherwise, the process is repeate 2) to 5). M max 2.2 Improve Wolves Algorithm R Through the brief escription of the basic principle of the wolves algorithm, we can see that in the last step of the algorithm, the iterative upating mechanism of the wolves is to irectly eliminate the iniviuals with poor fitness value an increase the iniviuals ranomly. Although this can increase the iversity of wolves, but ignore the location of these iniviuals have reache, reucing the convergence rate of the algorithm [6]. Therefore, in this paper, we propose a new improve wolves algorithm, which integrates the selection, crossover an mutation of genetic algorithm into the last step of the wolves algorithm to improve the convergence spee of the algorithm. The specific operation is as follows: 1) When the wolves are allocate foo, the wolves are sorte accoring to the size of the fitness value. The wolves are ivie into three parts, each of which is one-thir of the size of the wolves, an the subgroups with the greatest fitness value are Part of the genetic algorithm use in the elite retention strategy, irectly retaine into the next iteration. 2) That the secon part of the fitness value, that is, the secon part of the genetic algorithm use in the cross operation. In orer to obtain high-precision weight, in this paper, the iniviual is use in real coing, so cross-metho operation using real cross metho. The cross-operation of the -th gene an the l-th gene l in the j-th position is as follows: j j 1b ljb (7) 1b b lj lj j Where b is a ranom number between 0,1. 3) The smallest part of the fitness value, that is, the thir part of the genetic algorithm use in the mutation operation. The mutation operation selects the j-th gene of the i-th iniviual for mutation operation, as follows: R 88

4 Where max 0,1 between an min ( ) f ( g), r0.5 ( ) ( ), 0.5 ij ij max ij ij min ij f g r (8) g f ( g) (1 ) r2 (9) M max are the upper an lower bouns of the factor ; r 2 is a ranom number; g is the current iteration number. ij ; r is the ranom number The improve wolf algorithm is calle GWPA algorithm, the specific algorithm flow shown in Figure 1. Initialize Siege behavior Wal behavior i or lea T T max The optimal subgroups retaine into the next iteration The secon subgroups is use in the cross operation The worst subgroups is use in the mutation operation Rai behavior Meet the stop conition i lea Output optimal solution is near 3. GWPA-BP power preiction moel Figure 1 GWPA algorithm flow chart Although the BP multi-layer sensor has a strong nonlinear learning ability an therefore has a wie range of applications in photovoltaic power preiction, there are still many efects. For example: the stanar BP algorithm convergence spee is slow an easy to fall into the local extreme, training time is too long an other shortcomings. In this paper, a GWPA algorithm is propose to optimize the parameters of BP neural networ, an the GWPA-BP power preiction moel is establishe to improve the accuracy of preiction [7]. Wolves algorithm optimization BP neural networ specific process is as follows: 1) Encoing. The improve Wolves algorithm (GWPA) is an integer coing metho, an the parameters of the BP neural networ are encoe as a whole parameter. Each iniviual contains all the weights an threshols of BP an can represent a BP neural networ structure. 2) Initialize the parameters. The position i of the iniviual wolves in the wolves, the population size, the maximum number of iterations M max, the maximum number of wals T max, the waling step length step a, the attac step size step b, the attac step size step c, the istance etermination factor W, the upate scale factor an the step size factor. 3) The fitness function is set to the reciprocal of the mean square error function, an the fitness value is an important inex to evaluate the pros an cons of the population. 4) GWPA algorithm to optimize the initial weights an threshols of BP neural networs. Accoring to the above steps 1) - 4), the iterative optimization of the optimal iniviual is performe. 5) To etermine whether to meet the termination conitions, if satisfie, the output of the esire solution of the optimal solution - the location of the wolf is the optimal solution to be optimize; otherwise the implementation of step S

5 6) Output the global optimal solution, that is, the location of the space where the wolf. The spatial position of the wolf is use as the initial weight an threshol of the BP neural networ an the training BP neural networ is use to preict. 4. Simulation experiment In this paper, the UQ Solar project's previous PV ata an the corresponing meteorological ata are use as experimental samples. The atmospheric temperature, light intensity an output power of the PV power station from April 2015 to June 2015 are selecte as the sample set, The ata from April 25th to June 2n is the training sample set, an finally the GWPA-BP moel is use to preict the PV output power at eleven hour points from 7:00 to 5:00 on June 3, an the unmoifie BP preiction moel. The preiction results for the current ay are compare with the actual measure ata as shown in Fig.2. Figure 2. Comparison of preicte power values for both moels Table 1 gives the classical BP moel, GWPA-BP moel of the respective power preiction results an relative error. Table 2 gives a comparative analysis of the relative error of the two moels. Table 1. Comparison of two preictive moel preictive results Time Actual power value (w) Preictive value (w) BP moel Relative error (%) GWPA-BP moel Preictive value (w) Relative error (%) 7: : : : : : : : : : : Table 2. Comparison of two preiction moel errors Forecasting moel Minimum relative error Maximum relative Average relative error (%) error (%) (%) BP moel GWPA-BP moel

6 It can be seen from the above figure an table that the average relative error of the classical BP networ moel is 8.17% an the average relative error of the GWPA-BP moel is 5.08%. Therefore, the improve wolf algorithm improves the BP neural networ, the preiction error is graually reuce an the preiction accuracy is improve accoringly. It is prove that the GWPA-BP moel can preict the PV output power more accurately. 5. Conclusion As the wolves algorithm has the shortcomings of slow convergence, this paper uses the genetic algorithm to improve the wolves algorithm, the improve algorithm is calle GWPA algorithm. An the GWPA algorithm is applie to optimize the parameters of BP neural networ, a new moel GPWA-BP is propose. The simulation results show that the GPWA-BP moel can effectively preict the PV output power at a certain time. Compare with the unpreictable classical BP preiction moel, the preiction accuracy is greatly improve, an the short - term PV power can be preicte well, an it is wiely use in the power system with high accuracy. References [1] ingfei Gong, Zongxiang Lu, ing Qiao, Qiang Wang. Photovoltaic power preiction technology [J]. Power System Automation, 2016, (04): [2] I. S. Kim, M. B. Kim an M. J. oun, "ew Maximum Power Point Tracer Using Sliing-Moe Observer for Estimation of Solar Array Current in the Gri-Connecte Photovoltaic System," in IEEE Transactions on Inustrial Electronics, vol. 53, no. 4, pp , June [3] W. iao, M. G. J. Lin, W. G. Dunfor an A. Capel, "Real-Time Ientification of Optimal Operating Points in Photovoltaic Power Systems," in IEEE Transactions on Inustrial Electronics, vol. 53, no. 4, pp , June [4] Guoliang Li. Wolves algorithm research an application [D]. East China University of Technology, [5] Dongwen Cui, Bo Jin. Application of WPA-BP eural etwor Moel to Preiction of Monthly Runoff in Dry Season [J]. People's Pearl, 2016, (01): [6] Bingbing Li, Guangjun He, iaoliang ou, Dewei Tian, Jun Wang. Target threat estimation base on GA-WPA optimization for BP neural networs [J]. Computer Measurement an Control, 2015, (12): [7] awi,.m., Rehman, M.Z.; Khan, A. WS-BP: An efficient wolf search base bac propagation algorithm[j].aip Conference Proceeings, v 1660, p (7 pp.), 2015 [8] Huang Min-min ; an Wen-jun.Photovoltaic power forecast base on neural networ optimize by genetic algorithm [J]. Journal of Mechanical & Electrical Engineering, v 32, n 11, p , ov [9] Qian Wang, Zhisheng Zhang, Shuai Wang, Dongliang Cao. Research on photovoltaic power generation preiction base on PSO-R[J]. Journal of Qiingao University (Engineering an Technology Eition), 2014, (04): [10] Jia Liangbao.Stuy on short-term preiction of photovoltaic power output power base on neural networ [D]. Zhongyuan Institute of Technology,