摘 要:
局部遮荫下光伏阵列输出功率呈现多峰值,传统最大功率点跟踪算法常失效。为了提高发电的效率和稳定性,提出一种回归算法与粒子群算法融合的最大功率跟踪方法,该方法利用线性回归的预测性、泛化能力强的特点,改进粒子群算法初始化中的粒子随机性、易陷入局部极值的问题,并应用于局部遮荫下光伏阵列(单、多晶硅)最大功率点跟踪。通过实验与理论研究,首先发现随着训练集的增加,该方法跟踪性能呈现先上升后平缓趋势下降,且训练集比例选取总数据的55%~75%之间,跟踪性能最佳;其次,发现单(多)晶硅电池最佳跟踪精度分别为99.13%(99.16%),对比遮荫对电池最大功率影响,说明该方法具有普适性;最后,与粒子群算法和遗传算法相比,跟踪精度都提高了,且方差为零,表明该算法的精度高和鲁棒性强。结果表明,该方法具有普适性、精度高和鲁棒性强等优点。
关键词:光伏阵列;光伏发电;局部遮荫;多峰值;最大功率点跟踪;粒子群算法;回归算法;控制算法;
Research on PV system multi-peak MPPT based on fusion algorithm of regression
algorithm and particle swarm optimization algorithm
YE Guoming XIAO Wenbo WU Huaming
Key Laboratory of Nondestructive Testing,Ministry of Education,Nanchang Hangkong University
Jiangxi Engineering Laboratory for Optoelectronics Testing Technology
Abstract:
Under partial shading,the output power of PV array presents multiple peaks,so the traditional maximum power point tracking(MPPT) algorithm often fails. Therefore,an MPPT method fusing regression algorithm and particle swarm optimization(PSO)algorithm is proposed to improve the efficiency and stability of power generation. In the proposed method,the predictability and powerful generalization ability of linear regression is used to improve the particle randomness in the PSO initialization and eliminate the fact that the PSO algorithm is prone to falling into local extreme value. And then,it is applied to the MPPT of PV array(monocrystal silicon and polycrystalline silicon)under partial shading. By experimental and theoretical research,it is found firstly that,with the increase of training sets,the tracking performance of this method increases at first and then decreases gently,and the tracking performance is the best when the proportion of training sets is within 55%~75% of the total data;secondly,the optimal tracking accuracy of monocrystal silicon(polycrystalline silicon)PV cell is 99.13%(99.16%)respectively,which indicates that this method is suitable for MPPT of different types of solar PV cells as compared with the influence of the degree of shading on both cells′ maximum power;finally,the tracking accuracy of the proposed method is improved in comparison with those of PSO algorithm and genetic algorithm,and its variance is zero,which shows that the method has high accuracy and strong robustness. To sum up,the proposed method has advantages of universality,high precision and strong robustness.
Keyword:
PV array; PV power generation; partial shading; multi-peak value; MPPT; PSO algorithm; regression algorithm; control algorithm;
0 引言
最大功率点跟踪(Maximum Power Point Tracking,MPPT)是提高光伏发电效率的一项重要技术[1⁃6]。实际上,局部遮荫下阵列输出功率呈现多个峰值,此时传统的MPPT控制算法,如扰动观察法[7⁃8](Perturb and Observe,P&O)、电导增量法[9⁃10](Incremental Conductance,INC)等,容易陷入局部极值而跟踪失效。当前,人工智能算法,如粒子群算法、蚁群算法等,由于规则简单、可调参数少等优点,被广泛应用于局部遮荫下发电量跟踪[11]。为了提高跟踪精度,文献[12]提出了非线性动态惯性权重。为了提高跟踪稳定性,有研究提出机器学习与人工智能算法结合的方法[13]。实际上,跟踪的精度不仅受电池种类与特性影响[14],而且智能算法中训练过程也有影响[15⁃16]。因此,局部遮荫下光伏发电跟踪精度不仅有待提高,而且其中的随机性和波动性有待进一步研究。
针对上述问题,本文提出一种回归算法与粒子群算法的融合方法(RAPSO)跟踪光伏发电最大功率点,并研究了该方法的适用性,以及训练过程对该方法性能的影响规律,并与粒子群算法、遗传算法进行了对比。
1 实验测量设备、样品、测试内容及过程
实验装置主要由光强探测器(由宏诚科教公司出产型号HT⁃855)、卤钨灯太阳光模拟器(由佛山电器照明公司出产的管型卤钨灯)、Keithley 2400(由吉时利公司出产2400系列)、LabVIEW软件及PC机组成,具体如图1a)所示。测量样品是由艾科电子有限公司生产的9块54 mm×54 mm单晶硅、9块54 mm×54 mm多晶硅串并联,构成带有旁路二极管的阵列(3×3单晶硅阵列、3×3多晶硅阵列),具体如图1b)所示。动态遮荫采用黑卡片对第二列第二块电池进行遮挡。
测试内容包括电池阵列的输出电流(I)和电压(V)。测试过程如图1c)所示,卤钨灯太阳光模拟器垂直入射样品表面,用照度计测量正入射光强并记录;启动软件测量有无遮荫电流⁃电压曲线;记录有无遮荫下(光强变化在0~240 W/m2)输出电流⁃电压(I⁃U)和功率⁃电压(P⁃U)。所有实验都是在室温下进行。
2 融合方法建立
2.1 RA
回归算法(Regression Algorithm,RA)[17⁃18]是通过已获取自变量k与因变量y的基础上,建立回归方程,由新自变量得出因变量。多项式回归方程是通过增加变量自由度来捕获数据中非线性的变化。多项式回归方程为:
多项式回归方程不仅能够对非线性的数据迅速建模,而且可以在小的范围内任意逼近。
2.2 PSO算法
粒子群优化(Particle Swarm Optimization,PSO)[19⁃21]算法是模拟鸟群觅食行为而发展起来的一种多极值函数全局优化法,主要用来求解非线性多模态的约束和非约束优化问题。在PSO中,每个候选解都被看作一个粒子,是D维空间中的一个位置,粒子根据自身的个体经验和群体中最好的经验决定后面的运动。粒子会按式(2)、式(3)更新其速度和位置:
式中:Vit+1表示t+1时刻第i个粒子的速度;Vit表示t时刻第i个粒子的速度;Xit表示t时刻第i个粒子的位置;ω为惯性权重;c1和c2为学习因子;r1和r2为(0,1)之间均匀分布的随机数;Pbest表示个体最优值;Gbest表示全局最优值。
2.3 RAPSO
回归算法与粒子群算法的融合方法(RAPSO)执行流程主要是多项式回归训练历史采样数据得到预测模型,根据预测模型的趋势性确定峰值功率的大致位置,由粒子群搜索峰值位置。
具体步骤包括:
1)采集光伏阵列输出电压和电流数据。
2)划分训练集和测试集。
3)根据多项式回归方程(1)训练数据得到预测模型。
4)验证模型有效性。判别预测模型的误差,满足精度要求则转到下一步,否则返回步骤2)。
5)初始化粒子群。设置种群规模N为30,学习因子c1和c2为2,最大迭代次数M为100,搜索空间维数D为2,粒子最大速度Vmax为2。根据预测模型的趋势性确定粒子群算法搜索范围以及调整惯性权重。
6)计算粒子适应度值,寻找个体极值和群体极值。
7)根据式(2)、式(3)更新粒子速度和位置。
8)判断是否满足迭代终止条件。如果满足,则输出群体最优值;否则,跳转到步骤6),直到跟踪到群体最优值为止。
RAPSO算法流程图如图2所示。
3 实验结果与讨论
3.1 实验结果与分析
单晶硅、多晶硅的I⁃U和P⁃U特性曲线如图3所示,其中实线是没有遮挡情况,虚线是遮挡下。测得单晶硅为207组数据,多晶硅为206组数据。
从图3可以看出,无遮荫下,单晶硅、多晶硅阵列的I⁃U特性曲线呈现单阶梯,且P⁃U特性曲线呈现单峰值。遮荫下,I⁃U特性曲线呈现多阶梯状态,且P⁃U特性曲线呈现多峰值状态。其次,可看出在无遮荫下,单晶硅、多晶硅峰值功率分别为0.079 0 W,0.055 4 W,而动态遮荫下,其峰值功率分别为0.027 5 W,0.023 0 W。由此,单晶硅在遮荫下峰值功率降低了约65%,多晶硅降低了约58%。可见,动态遮荫对单晶硅功率影响程度较大,原因在于单晶硅电池优于多晶硅的材料性能,易感于外界环境。
3.2 影响RAPSO跟踪效果的训练因素
RAPSO训练过程影响跟踪性能,为此研究了训练集从总数据量的10%变化到90%时发电量跟踪精度,结果如图4所示。
由图4可以看出随着训练集的增加,单(多)晶硅跟踪精度总体呈现先上升后平缓下降趋势。其次,看出跟踪精度峰值都在55%~75%,原因在于合适的训练集不仅能提供尽可能多的训练信息,而且有利于提高对数据的拟合能力,上述结果与文献[22⁃23]研究结果符合。最后,注意到单晶硅功率跟踪精度从91.21%变化到97.73%,最高达到99.13%;多晶硅功率跟踪精度从91.33%变化到98.33%,最高达到99.16%。最高和最低单(多)晶硅跟踪精度差分别为7.92%(7.83%),说明尽管动态遮荫对单晶硅功率影响程度较大,但是该方法对单(多)晶硅的跟踪精度不仅基本一致,而且变化小。说明该方法具有普适性。
3.3 RAPSO与PSO、GA对比
将上述单(多)晶硅电池阵列采集的数据应用于粒子群算法[24]、遗传算法[25]最大功率点跟踪,并与RAPSO结果对比。对比时选取总数据集的70%用于训练、30%用于验证,10次平均跟踪结果如表1所示。
由表1可知,RAPSO跟踪单(多)晶硅遮荫下精度达99.13%(99.08%);相比于PSO,单(多)晶硅跟踪精度分别提高了5.43%(2.2%);相比于GA,单(多)晶硅跟踪精度分别提高了1.56%(1.28%)。原因在于RAPSO利用RA的预测性改善了PSO的粒子随机性,使PSO缩小搜索范围,跳出局部极值。从方差来看,RAPSO为零,说明该方法鲁棒性强、稳定性高,进一步验证了上述结论。
4 结论
综上所述,RAPSO跟踪遮荫下单(多)晶硅电池最大功率点的研究结论如下:
1)训练集在总数据量的55%~75%之间,RAPSO跟踪性能最佳。
2)RAPSO对动态遮荫下不同类型的电池跟踪精度基本一样,具有适用性。
3)与粒子群算法和遗传算法对比发现,RAPSO跟踪精度高、稳定性强。
参考文献
[1] 戴增辉,李光布.应用于光伏发电系统的最大功率点跟踪控制器设计[J].现代电子技术,2019,42(5):131-134.
[2] DE ROCHA M V,SAMPAIO L P,DE SILVA S A O. Comparative analysis of MPPT algorithms based on Bat algorithm for PV systems under partial shading condition[J]. Sustainable energy technologies and assessments,2020,40:1-14.
[3] YILMAZ U,TURKSOY O,TEKE A. Improved MPPT method to increase accuracy and speed in photovoltaic systems under variable atmospheric conditions[J]. International journal of electrical power&energy systems,2019,113:634-651.
[4] 罗驰,任一峰,安坤,等.基于Hermite插值多项式的光伏MPPT改进算法的研究[J].现代电子技术,2021,44(5):137-142.
[5] SIVAKUMAR P,KADER A A,KALIAVARADHAN Y,et al.Analysis and enhancement of PV efficiency with incremental conductance MPPT technique under non-linear loading conditions[J]. Renewable energy,2015,81:543-550.
[6] PATHAK D,SAGAR G,GAUR P. An application of intelligent non-linear discrete-PID controller for MPPT of PV system[J]. Procedia computer science,2020,167:1574-1583.
[7] SHER H A,MURTAZA A F,NOMAN A,et al. A new sensorless hybrid MPPT algorithm based on fractional short-circuit current measurement and P&O MPPT[J]. IEEE transactions on sustainable energy,2015,6(4):1426-1434.
[8] 梁国壮,赵博,田涵雷,等.扰动-模糊结合的双模式MPPT算法研究[J].现代电子技术,2020,43(16):131-134.
[9] 马永翔,王一君,闫群民,等.局部阴影下光伏阵列的MPPT综合优化[J].现代电子技术,2019,42(11):140-143.
[10] TEY K S,MEKHILEF S. Modified incremental conductance MPPT algorithm to mitigate inaccurate responses under fastchanging solar irradiation level[J]. Solar energy,2014,101:333-342.
[11] LIAO T J,STÜTZLE T,DE OCA M A M,et al. A unified ant colony optimization algorithm for continuous optimization[J]. European journal of operational research,2014,234(3):597-609.
[12] WANG Y L, NAN B. Research of MPPT control method based on PSO algorithm[C]//International Conference on Computer Science&Network Technology. Harbin, China:IEEE,2016:698-701.
[13] TEO T T,LOGENTHIRAN T,WOO W L,et al. Forecasting of photovoltaic power using regularized ensemble Extreme Learning Machine[C]//2016 IEEE Region 10 Conference(TENCON). Singapore:IEEE,2017:455-458.
[14] XIAO W B,NAZARIO G,WU H M,et al. A neural network based computational model to predict the output power of different types of photovoltaic cells[J]. Plos one,2017,12(9):e0184561.
[15] ALBERTO D,FRANCESCO G,SONIA L,et al. Comparison of training approaches for photovoltaic forecasts by means of machine learning[J]. Applied sciences,2018,8(2):228.
[16] ABDELWAHAB O,BAHGAT M,LOWRANCE C J,et al.Effect of training set size on SVM and Naive Bayes for Twitter sentiment analysis[C]//2015 IEEE International Symposium on Signal Processing and Information Technology(ISSPIT).Abu Dhabi,United Arab Emirates:IEEE,2015:46-51.
[17] CHEN J,DE HOOGH K,GULLIVER J,et al. A comparison of linear regression,regularization,and machine learning algorithms to develop Europe-wide spatial models of fine particles and nitrogen dioxide[J]. Environment international,2019,130:104934.
[18] WEICHENTHAL S,VAN RYSWYK K,GOLDSTEIN A,et al.A land use regression model for ambient ultrafine particles in Montreal,Canada:a comparison of linear regression and a machine learning approach[J]. Environmental research,2016,146:65-72.
[19] ELTAMALY A M,AL-SAUD M S,ABOKHALIL A G. Photovoltaic maximum power point tracking under dynamic partial shading changes by novel adaptive particle swarm optimization strategy[J]. Transactions of the institute of measurement and control,2019,42(1):104-115.
[20] RAM J P,RAJASEKAR N. A new robust,mutated and fast tracking LPSO method for solar PV maximum power point tracking under partial shaded conditions[J]. Applied energy,2017,201:45-59.
[21] VEERAPEN S,WEN H Q,LI X S. A novel global maximum power point tracking algorithm for photovoltaic system with variable perturbation frequency and zero oscillation[J]. Solar energy,2019,181:345-356.
[22] SHAHIN M A,MAIER H R,JAKSA M B. Data division for developing neural networks applied to geotechnical engineering[J]. Journal of computing in civil engineering,2004,18(2):105-114.
[23] THEOCHARIDES S,MAKRIDES G,GEORGHIOU G E,et al.Machine learning algorithms for photovoltaic system power output prediction[C]//2018 IEEE International Energy Conference(ENERGYCON). Limassol,Cyprus:IEEE,2018:1-6.
[24] KENNEDY J,EBERHART R. Particle swarm optimization[C]//Proceedings of IEEE International Conference on Neural Networks. Perth,Australia:IEEE,1995:1942-1948.
[25] JONES E A,JOINES W T. Design of Yagi-Uda antennas using genetic algorithms[J]. IEEE transactionS on antennas and propagation, 2002,45(9):1386-1392.