CN104484833A

CN104484833A - Photovoltaic power generation output power tracking algorithm based on genetics algorithm improved RBF-BP neural network

Info

Publication number: CN104484833A
Application number: CN201410720814.2A
Authority: CN
Inventors: 朱正伟; 周谢益; 郭枫; 张丹; 张南; 钱露; 宋文浩; 黄晓竹
Original assignee: Changzhou University
Current assignee: Changzhou University
Priority date: 2014-12-02
Filing date: 2014-12-02
Publication date: 2015-04-01

Abstract

The invention discloses a photovoltaic power generation output power tracking algorithm based on an improved RBF-BP neural network based on a genetic algorithm. By establishing an RBF-BP neural network, the absolute value of the error between the photovoltaic power generation output power prediction output and the expected output is used as an adaptive Then use the genetic algorithm to select, cross and mutate the data collected by the photovoltaic power generation equipment to find the individual corresponding to the optimal fitness. Combining the advantages of fast convergence speed and group classification performance of RBF neural network and self-learning and strong self-adaptive ability of BP neural network, it has the characteristics of better generalization performance, faster convergence speed and higher prediction accuracy.

Description

Photovoltaic power generation output power tracking algorithm based on improved RBF-BP neural network based on genetic algorithm

技术领域technical field

本发明涉及一种光伏发电输出功率的追踪算法，尤其涉及一种基于遗传学算法改进的RBF-BP神经网络的光伏发电输出功率追踪算法。The invention relates to a tracking algorithm of output power of photovoltaic power generation, in particular to a tracking algorithm of output power of photovoltaic power generation based on an improved RBF-BP neural network based on a genetic algorithm.

背景技术Background technique

随着传统能源消耗造成的环境污染问题日益突出，可再生能源的利用引起广泛的重视。光伏发电作为一种新兴崛起的可再生能源形式，具有广泛的开发前景和商业价值。因而得到了越来越多的关注。大规模的光伏并网发电是目前光伏发电系统的主流趋势，目前大规模的光伏并网系统已得到应用。As the problem of environmental pollution caused by traditional energy consumption has become increasingly prominent, the utilization of renewable energy has attracted widespread attention. Photovoltaic power generation, as an emerging form of renewable energy, has broad development prospects and commercial value. Therefore, it has received more and more attention. Large-scale photovoltaic grid-connected power generation is the mainstream trend of photovoltaic power generation systems, and large-scale photovoltaic grid-connected systems have been applied.

光伏发电与风力发电一样，均属于波动性和间歇性电源。光伏发电受环境因素影响较大，特别是太阳辐射强度、环境温度等，因为其输出功率具有不确定性。其并入大电网后使得大电网的整体负荷预测准确性降低，也必然引起整个系统的电压、频率的波动，增加了传统发电，控制和运行计划的难度，不利于整个电网系统的调度。Photovoltaic power generation, like wind power generation, is a fluctuating and intermittent power source. Photovoltaic power generation is greatly affected by environmental factors, especially solar radiation intensity, ambient temperature, etc., because its output power is uncertain. After its integration into the large power grid, the accuracy of the overall load forecast of the large power grid will be reduced, and it will inevitably cause fluctuations in the voltage and frequency of the entire system, which will increase the difficulty of traditional power generation, control and operation planning, and is not conducive to the scheduling of the entire power grid system.

所以，高效准确的预测光伏发电输出功率对于能够安全高效的利用光伏发电就显得尤为必要。目前光伏发电输出功率的预测方法有扰动观察法、电导增量法和神经网络法等。无论是扰动观察法和还是电导增量法都存在着一个步长固定的问题。如果步长过小，便会导致光伏阵列长时间地滞留在低功率输出区，而步长过大，就会导致系统振荡加剧。相比之下人工神经网络其追踪过程不需要光伏电池阵列的物理参数，且神经网络通过学习训练可逼近任意非线性特性曲线的能力。所以在非线性系统的建模与辨识的过程中运用神经网络模型，可以不受非线性模型的限制，适应性更好。神经网络法主要包括RBF和BP神经网络预测法，它们都有各自的优缺点。RBF神经网络预测的收敛速度快但是泛化能力比较差；BP神经网络的自学习能力强但是学习算法不能保证学习的结果达到均方误差的全局最小、训练结果容易受到不正确训练样本集的错误引导。并且两者的在构建神经网络的过程中阈值与权值需要反复训练修改，初始值具有随机性。Therefore, efficient and accurate prediction of the output power of photovoltaic power generation is particularly necessary for the safe and efficient use of photovoltaic power generation. At present, the output power prediction methods of photovoltaic power generation include disturbance and observation method, conductance increment method and neural network method, etc. Both the perturbation and observation method and the conductance incremental method have a problem of fixed step size. If the step size is too small, it will cause the photovoltaic array to stay in the low power output area for a long time, and if the step size is too large, it will lead to aggravated system oscillation. In contrast, the tracking process of the artificial neural network does not require the physical parameters of the photovoltaic cell array, and the neural network can approach the ability of any nonlinear characteristic curve through learning and training. Therefore, the use of neural network models in the process of modeling and identification of nonlinear systems can not be limited by nonlinear models and has better adaptability. Neural network methods mainly include RBF and BP neural network prediction methods, both of which have their own advantages and disadvantages. RBF neural network predicts fast convergence but poor generalization ability; BP neural network has strong self-learning ability but the learning algorithm cannot guarantee that the learning result reaches the global minimum of the mean square error, and the training result is vulnerable to errors caused by incorrect training sample sets guide. And the thresholds and weights of the two need to be trained and modified repeatedly during the process of building the neural network, and the initial values are random.

发明内容Contents of the invention

为了克服RBF泛化能力差和BP神经网络训练结果容易受到不正确训练样本集的错误引导的缺陷，本发明提出了一种能够高效准确的预测光伏发电输出功率的基于遗传学算法改进的RBF-BP神经网络的光伏发电输出功率追踪算法。In order to overcome the defects of poor generalization ability of RBF and BP neural network training results are easily misguided by incorrect training sample sets, the present invention proposes an improved RBF-based genetic algorithm that can efficiently and accurately predict the output power of photovoltaic power generation. BP neural network output power tracking algorithm for photovoltaic power generation.

本发明提出的基于遗传学算法改进的RBF-BP神经网络的光伏发电输出功率追踪算法是通过建立一个RBF-BP神经网络，将光伏发电输出功率预测输出和期望输出之间的误差绝对值作为适应度，再运用遗传学算法对光伏发电设备所采集到的数据进行选择、交叉和变异操作找到最优适应度对应的个体。RBF-BP神经网络预测用遗传算法得到最优个体对网络初始权值和阈值赋值，网络训练后对光法发电输出功率进行预测。The photovoltaic power generation output power tracking algorithm of the improved RBF-BP neural network based on the genetic algorithm proposed by the present invention is to establish an RBF-BP neural network, and use the absolute value of the error between the predicted output of the photovoltaic power generation output and the expected output as an adaptive Then use the genetic algorithm to select, cross and mutate the data collected by the photovoltaic power generation equipment to find the individual corresponding to the optimal fitness. The RBF-BP neural network prediction uses the genetic algorithm to obtain the optimal individual to assign the initial weight and threshold of the network. After the network is trained, the output power of the optical power generation is predicted.

所述的基于遗传学算法改进的RBF-BP神经网络的光伏发电输出功率预测方法，包括如下步骤：The method for predicting the output power of photovoltaic power generation based on the improved RBF-BP neural network of the genetic algorithm comprises the following steps:

(1)根据光伏发电输出特性，选取日常天气状况下统计并采集一天之中各个时段影响光伏发电电池板发电的因素，本发明中侧重于光伏电池板工作温度和光伏发电设备工作时的光照强度和光伏发电输出功率作为RBF-BP神经网络训练的输入输出。另选取同等条件下各个时段的样本数据作为RBF-BP神经网络的测试数据。(1) According to the output characteristics of photovoltaic power generation, the factors that affect the power generation of photovoltaic power generation panels in various periods of the day are selected under daily weather conditions and collected. The present invention focuses on the operating temperature of photovoltaic panels and the intensity of light when photovoltaic power generation equipment works And the output power of photovoltaic power generation is used as the input and output of RBF-BP neural network training. In addition, the sample data of each time period under the same conditions are selected as the test data of the RBF-BP neural network.

(2)建立RBF-BP神经网络用于训练归一化后的样本数据。本发明提出建立的RBF-BP组合神经网络是结合了RBF神经网络收敛速度快、群分类性能好和BP神经网络自学习、自适应能力强等优点建立的一个由RBF子网和BP子网两部分构成双隐层RBF-BP组合神经网络，其具有泛化性能更好、收敛速度更快、预测精度更高等特点。本发明提出的RBF-BP组合神经网络分为：输入层、隐含层和输出层，各层的节点数设计如下：(2) Establish the RBF-BP neural network for training the normalized sample data. The RBF-BP combined neural network proposed by the present invention is a combination of the RBF neural network with fast convergence speed, good group classification performance and BP neural network self-learning and strong self-adaptive ability. Part of it constitutes a double-hidden layer RBF-BP combination neural network, which has the characteristics of better generalization performance, faster convergence speed, and higher prediction accuracy. The RBF-BP combination neural network that the present invention proposes is divided into: input layer, hidden layer and output layer, and the number of nodes of each layer is designed as follows:

输入层：针对光伏发电最大功率的预测，在不考虑突变的天气情况和局部光照不均匀的情况下对光伏发电最大功率点影响的情况下，神经网络输入量的选取主要考虑两个部分，即光伏电池板工作温度和光伏发电设备工作时的光照强度。Input layer: For the prediction of the maximum power of photovoltaic power generation, without considering the impact of sudden weather conditions and uneven local illumination on the maximum power point of photovoltaic power generation, the selection of neural network input mainly considers two parts, namely The operating temperature of photovoltaic panels and the light intensity of photovoltaic power generation equipment.

隐含层：本发明涉及到的RBF-BP神经网络，相比于传统的BP神经网络，在隐含层加入了RBF神经网络作为子层。输入步骤(1)中的样本先经过RBF神经网络子网进行训练，再将训练结果作为BP子网的输入对其进行训练。RBF-BP子网的隐含层节点的传递函数设置为高斯函数，如式(1)所示：Hidden layer: the RBF-BP neural network involved in the present invention, compared with the traditional BP neural network, adds the RBF neural network as a sublayer in the hidden layer. The samples in the input step (1) are first trained through the RBF neural network subnet, and then the training result is used as the input of the BP subnet to train it. The transfer function of the hidden layer nodes of the RBF-BP subnetwork is set to a Gaussian function, as shown in formula (1):

${u u}_{i i} ((X x)) = = exp exp [[- - {| | | | X x - - {c c}_{i i} | | | |}^{22} / / {22 σ σ}_{i i}^{22}]] ((i i = = 1,2 1,2,, . . . . . .,, {N N}_{22})) - - - - - - ((11))$

其中，u_i(X)是第i个隐层节点的输出，X是步骤(1)输入样本，c_i是高斯函数的中心向量，σ_i为节点的基宽度参数，且为大于零的数。Among them, u _i (X) is the output of the i-th hidden layer node, X is the input sample of step (1), ci is the center vector of the Gaussian function, _{σ i} _is the base width parameter of the node, and is a number greater than zero .

将RBF子网的输出作为组合神经网络中的BP子网的输入。其隐含层节点的传递函数设计为Sigmoid型函数，如式(2)：The output of the RBF subnet is used as the input of the BP subnet in the combined neural network. The transfer function of its hidden layer nodes is designed as a Sigmoid function, such as formula (2):

f(x)＝1/(1+e^-x) (2)f(x)＝1/(1+e ^-x ) (2)

因而BP子网络的隐层节点的输出为式(3)：Therefore, the output of the hidden layer nodes of the BP subnetwork is formula (3):

${O o}_{j j} = = f f [[{Σ Σ}_{i i = = 11}^{{N N}_{22}} {W W}_{ij ij} {u u}_{i i} ((X x))]] - - - - - - ((33))$

其中，W_ij为连接RBF隐含子层第i个节点到BP隐含子层的第j个节点的权值，N₂为RBF隐含子层节点数。Among them, W _ij is the weight connecting the i-th node of the RBF hidden sub-layer to the j-th node of the BP hidden sub-layer, and N ₂ is the number of nodes in the RBF hidden sub-layer.

输出层：输出层的节点个数可根据情况来定。但为了简化神经网络的设计，本发明选择了一个输出节点，即最大功率点处的电压。计算输出层节点的输出值公式为式(4)：Output layer: The number of nodes in the output layer can be determined according to the situation. However, in order to simplify the design of the neural network, the present invention selects an output node, that is, the voltage at the maximum power point. The formula for calculating the output value of the output layer node is formula (4):

${y the y}_{k k} = = f f [[{Σ Σ}_{j j = = 11}^{{N N}_{33}} {W W}_{jk jk} {O o}_{j j}]] - - - - - - ((44))$

其中，W_jk为连接BP隐含子层第j个节点到输出层第k个节点的权值，N₃为BP隐含子层节点数。Among them, W _jk is the weight connecting the jth node of the BP hidden sublayer to the kth node of the output layer, and _N3 is the number of nodes in the BP hidden sublayer.

f_i根据实际输出与期望输出计算输出层的误差绝对值，如式(5)：f _i Calculate the absolute value of the error of the output layer according to the actual output and the expected output, as shown in formula (5):

$E E. = = {Σ Σ}_{k k = = 11}^{{N N}_{44}} | | {d d}_{k k} - - {y the y}_{k k} | | - - - - - - ((55))$

其中，d_k为期望输出。N₄为输出层节点数，E为误差绝对值。Among them, d _k is the desired output. N ₄ is the number of nodes in the output layer, and E is the absolute value of the error.

(3)将步骤(1)采集到的数据带入步骤(2)中建立的RBF-BP神经网络中，得到实际输出与期望输出的误差E。作为用于遗传算法的适应度值。(3) Bring the data collected in step (1) into the RBF-BP neural network established in step (2) to obtain the error E between the actual output and the expected output. as the fitness value for the genetic algorithm.

(4)根据步骤(2)中构建的RBF-BP神经网络的拟合函数输入输出参数个数确定RBF-BP神经网络的网络结构，确定RBF-BP神经网络中的阈值和权值的个数，进而确定遗传算法的个体的长度。(4) determine the network structure of RBF-BP neural network according to the fitting function input and output parameter number of the RBF-BP neural network constructed in step (2), determine the number of threshold and weight in RBF-BP neural network , and then determine the length of the individual of the genetic algorithm.

(5)将步骤(1)采集到的数据用遗传学算法进行优化，具体实施步骤：(5) Optimizing the data collected in step (1) with a genetic algorithm, specific implementation steps:

A.种群初始化A. Population initialization

种群中的每个个体都包含整个RBF-BP神经网络的所有权值和阈值，个体编码方法为实数编码，每个个体均为一个实数串，个体通过遗传算法适应度函数计算个体的适应度值。Each individual in the population contains the ownership value and threshold of the entire RBF-BP neural network. The individual encoding method is real number encoding, and each individual is a real number string. The individual fitness value is calculated through the fitness function of the genetic algorithm.

B.适应度函数B. Fitness function

根据步骤(1)的个体得到RBF-BP神经网络的初始权值和阈值，用步骤(3)中得到的绝对误差值E作为个体适应度值F。According to the individual in step (1), the initial weight and threshold of the RBF-BP neural network are obtained, and the absolute error value E obtained in step (3) is used as the individual fitness value F.

C.选择操作C. Select operation

遗传算法选择操作有轮盘赌法、锦标赛等多种方法，本发明选择轮盘赌法，每个个体i的选择概率p_i为公式(6)：Genetic algorithm selection operation has multiple methods such as roulette method, tournament, the present invention selects roulette method, the selection probability _p of each individual i is formula (6):

$\{\begin{matrix} {f f}_{i i} = = k k / / {F f}_{i i} \\ {p p}_{i i} = = \frac{{f f}_{i i}}{{Σ Σ}_{j j = = 11}^{N N} {f f}_{i i}} \end{matrix} - - - - - - ((66))$

式中F_i为个体i的适应度值，由于适应度值越小越好，所以在个体选择前对适应度值求倒数；k为系数；N为种群个体数目。In the formula, F _i is the fitness value of individual i. Since the smaller the fitness value is, the better, so the reciprocal of the fitness value is calculated before individual selection; k is the coefficient; N is the number of individuals in the population.

D.交叉操作D. Cross operation

由于个体采用实数编码，所以交叉操作方法也采用相应的实数交叉法，第k个染色体a_k和l个染色体a_l在j位的交叉操作方法如公式(7)：Since the individual is coded with a real number, the crossover operation method also adopts the corresponding real number crossover method. The crossover operation method of the kth chromosome a _k and the l chromosome a _l at position j is shown in formula (7):

$\begin{matrix} {a a}_{kj kj} = = {a a}_{kj kj} ((11 - - b b)) + + {a a}_{lj lj} b b \\ {a a}_{lj lj} = = {a a}_{lj lj} ((11 - - b b)) + + {a a}_{kj kj} b b \end{matrix}\} - - - - - - ((77))$

其中b为[0，1]间的随机数。Where b is a random number between [0, 1].

E.变异操作E. Mutation operation

选取第i个个体的第j个基因a_ij进行变异，变异操作的个公式如(8)：Select the j-th gene a _ij of the i-th individual to mutate. The formula for the mutation operation is as follows (8):

${a a}_{ij ij} = = \{\begin{matrix} {a a}_{ij ij} = = {a a}_{ij ij} + + (({a a}_{ij ij} - - {a a}_{max max})) * * f f ((g g)) & r r > > 0.5 0.5 \\ {a a}_{ij ij} = = {a a}_{ij ij} + + (({a a}_{min min} - - {a a}_{ij ij})) * * f f ((g g)) & r r \leq \leq 0.5 0.5 \end{matrix} - - - - - - ((88))$

公式(8)中，a_max为基因a_ij的上界；a_min为基因a_ij的下界；f(g)＝r₂(1-g/G_max)²；r₂为一个随机数；g为当前迭代次数；G_max是最大进化次数；r为[0，1]间的随机数。In formula (8), a _max is the upper bound of gene a _ij ; a _min is the lower bound of gene a _ij ; f(g)=r ₂ (1-g/G _max ) ² ; r ₂ is a random number; g is the current number of iterations; G _max is the maximum number of evolutions; r is a random number between [0, 1].

进一步，还包括：Further, it also includes:

将步骤(1)的训练数据通过步骤(4)的遗传算法通过选择、交叉和变异操作找到最优适应度值对应个体。RBF-BP神经网络用遗传算法得到的最优个体对网路初始权值和阈值赋值。将步骤(1)采集的测试数据测试用于遗传算法改进后的RBF-BP神经网络，当神经网络训练误差小于目标误差时，网络收敛；当网络训练次数等于最大迭代次数而训练误差仍大于目标误差时，网络不收敛。此时再通过RBF-BP神经网络的反向学习能力，反向修改神经网络的权值和阈值，调整公式如(9)、(10)：The training data in step (1) is passed through the genetic algorithm in step (4) to find the individual corresponding to the optimal fitness value through selection, crossover and mutation operations. RBF-BP neural network assigns the initial weight and threshold of the network to the optimal individual obtained by genetic algorithm. Test the test data collected in step (1) for the RBF-BP neural network improved by the genetic algorithm. When the neural network training error is less than the target error, the network converges; when the network training times are equal to the maximum number of iterations and the training error is still greater than the target error, the network does not converge. At this time, through the reverse learning ability of the RBF-BP neural network, reversely modify the weight and threshold of the neural network, and adjust the formulas such as (9) and (10):

W_jk＝W_jk+λy_k(1-y_k)(d_k-y_k)O_j (9)W _jk ＝W _jk +λy _k (1-y _k )(d _k -y _k )O _j (9)

θ_k＝θ_k+λy_k(1-y_k)(d_k-y_k) (10)θ _k ＝θ _k +λy _k (1-y _k )(d _k -y _k ) (10)

λ为学习速率。训练完毕后的RBF-BP神经网络可用于光伏发电最大输出功率进行追踪。λ is the learning rate. The trained RBF-BP neural network can be used to track the maximum output power of photovoltaic power generation.

本发明的优点是对于光伏发电输出功率的追踪：The advantage of the present invention is the tracking of the output power of photovoltaic power generation:

(1)建立了一个由RBF子网和BP子网两部分构成双隐层RBF-BP组合神经网络，结合了RBF神经网络收敛速度快、群分类性能好和BP神经网络自学习、自适应能力强等优点，具有泛化性能更好、收敛速度更快、预测精度更高等特点。(1) A double-hidden layer RBF-BP combined neural network composed of RBF subnet and BP subnet was established, combining the fast convergence speed of RBF neural network, good group classification performance and self-learning and self-adaptive ability of BP neural network It has the advantages of better generalization performance, faster convergence speed and higher prediction accuracy.

(2)运用遗传学算法对样本数据进行选择、交叉和变异操作得到最优适应度个体，用于进一步优化RBF-BP神经网络。(2) Use the genetic algorithm to select, cross and mutate the sample data to obtain the individual with the best fitness, which is used to further optimize the RBF-BP neural network.

(3)通过双隐层RBF-BP组合神经网络和遗传学算法的结合不但能够高效准确的预测光伏发电输出功率，安全高效的利用光伏发电，同时能够更好的应对光伏发电中的波动性和间歇性电源。(3) The combination of double-hidden layer RBF-BP combined neural network and genetic algorithm can not only efficiently and accurately predict the output power of photovoltaic power generation, but also make use of photovoltaic power generation safely and efficiently. intermittent power.

附图说明Description of drawings

图1：RBF-BP网络结构图；Figure 1: RBF-BP network structure diagram;

图2：遗传算法改进后的算法流程图。Figure 2: The algorithm flow chart of the improved genetic algorithm.

具体实施方式Detailed ways

如图1所示，RBF-BP神经网络的拥有两个隐含层，输入数据经由RBF神经网络子网训练后作为BP神经网络子网的输入进行训练，该网络具有误差反向学习的能力，当训练结果达不到精度要求时，能反向修改神经网络的权值和阈值直至训练结果达到精度要求。As shown in Figure 1, the RBF-BP neural network has two hidden layers. The input data is trained as the input of the BP neural network subnet after being trained by the RBF neural network subnetwork. The network has the ability of error reverse learning, When the training result does not meet the accuracy requirements, the weights and thresholds of the neural network can be reversely modified until the training results meet the accuracy requirements.

如图2所示，基于遗传学算法改进的RBF-BP神经网络算法是在确认网络结构的情况下去，确认输出权值阈值长度。将输出和期望输出之间的误差绝对值作为适应度，再运用遗传学算法对数据进行选择、交叉和变异操作找到最优适应度对应的个体，进而确认RBF-BP神经网络的权值和阈值。测试训练后网络的误差是否满足预测光伏发电功率的精度要求。As shown in Figure 2, the improved RBF-BP neural network algorithm based on the genetic algorithm is to confirm the output weight threshold length under the condition of confirming the network structure. The absolute value of the error between the output and the expected output is used as the fitness, and then the genetic algorithm is used to select, cross and mutate the data to find the individual corresponding to the optimal fitness, and then confirm the weight and threshold of the RBF-BP neural network . Test whether the error of the network after training meets the accuracy requirements for predicting photovoltaic power generation.

所述的基于遗传学算法改进的RBF-BP神经网络的光伏发电输出功率预测方法，其特征在于包括如下步骤：The method for predicting the output power of photovoltaic power generation based on the improved RBF-BP neural network of the genetic algorithm is characterized in that it comprises the following steps:

(1)根据光伏发电输出特性，选取日常天气状况下统计并采集一天之中各个时段影响光伏发电电池板发电的因素，本发明中侧重于光伏电池板工作温度和光伏发电设备工作时的光照强度和光伏发电输出功率作为RBF-BP神经网络训练的输入输出。另选取同等条件下各个时段的样本数据作为RBF-BP神经网络的测试数据。(1) According to the output characteristics of photovoltaic power generation, the factors that affect the power generation of photovoltaic power generation panels in various periods of the day are selected under the daily weather conditions and collected. The present invention focuses on the operating temperature of photovoltaic panels and the intensity of light when photovoltaic power generation equipment works And the output power of photovoltaic power generation is used as the input and output of RBF-BP neural network training. In addition, the sample data of each time period under the same conditions are selected as the test data of the RBF-BP neural network.

(2)建立RBF-BP神经网络用于训练归一化后的样本数据。本发明提出建立的RBF-BP组合神经网络是结合了RBF神经网络收敛速度快、群分类性能好和BP神经网络自学习、自适应能力强等优点建立的一个由RBF子网和BP子网两部分构成双隐层RBF-BP组合神经网络，其具有泛化性能更好、收敛速度更快、预测精度更高等特点。本发明提出的RBF-BP组合神经网络分为：输入层、隐含层和输出层，各层的节点数设计如下：(2) Establish the RBF-BP neural network for training the normalized sample data. The RBF-BP combined neural network proposed by the present invention is a combination of the RBF neural network with fast convergence speed, good group classification performance and BP neural network self-learning and strong self-adaptive ability. Part of it constitutes a double-hidden layer RBF-BP combination neural network, which has the characteristics of better generalization performance, faster convergence speed, and higher prediction accuracy. The RBF-BP combined neural network that the present invention proposes is divided into: input layer, hidden layer and output layer, and the number of nodes of each layer is designed as follows:

f(x)＝1/(1+e^-x) (2)f(x)＝1/(1+e ^-x ) (2)

A.种群初始化A. Population initialization

种群中的每个个体都包含整个RBF-BP神经网络的所有权值和阈值，个体编码方法为实数编码，每个个体均为一个实数串，个体通过遗传算法适应度函数计算个体的适应度值。Each individual in the population contains the ownership value and threshold of the entire RBF-BP neural network. The individual encoding method is real number encoding, and each individual is a real number string. The individual calculates the individual fitness value through the genetic algorithm fitness function.

B.适应度函数B. Fitness function

C.选择操作C. Select operation

D.交叉操作D. Cross operation

其中b为[0，1]间的随机数。Where b is a random number between [0, 1].

E.变异操作E. Mutation operation

(6)步骤(1)的训练数据通过步骤(4)的遗传算法通过选择、交叉和变异操作找到最优适应度值对应个体。RBF-BP神经网络用遗传算法得到的最优个体对网路初始权值和阈值赋值。(6) The training data in step (1) finds the individual corresponding to the optimal fitness value through the genetic algorithm in step (4) through selection, crossover and mutation operations. RBF-BP neural network assigns the initial weight and threshold of the network to the optimal individual obtained by genetic algorithm.

(7)将步骤(1)采集的测试数据测试用于遗传算法改进后的RBF-BP神经网络，当神经网络训练误差小于目标误差时，网络收敛；当网络训练次数等于最大迭代次数而训练误差仍大于目标误差时，网络不收敛。此时再通过RBF-BP神经网络的反向学习能力，反向修改神经网络的权值和阈值，调整公式如(9)、(10)：(7) Test the test data collected in step (1) for the RBF-BP neural network improved by the genetic algorithm. When the neural network training error was less than the target error, the network converged; when the network training times equaled the maximum number of iterations and the training error When still greater than the target error, the network does not converge. At this time, through the reverse learning ability of the RBF-BP neural network, reversely modify the weight and threshold of the neural network, and adjust the formulas such as (9) and (10):

式中，λ为学习速率。训练完毕后的RBF-BP神经网络可用于光伏发电最大输出功率进行追踪。In the formula, λ is the learning rate. The trained RBF-BP neural network can be used to track the maximum output power of photovoltaic power generation.

以上所述仅为本发明的优选实施例而已，并不用于限制本发明，对于本领域的技术人员来说，本发明可以有各种更改和变化。凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A photovoltaic power generation output power tracking algorithm based on the improved RBF-BP neural network of genetic algorithm, it is characterized in that, comprises steps:

1) According to the output characteristics of photovoltaic power generation, the factors that affect the power generation of photovoltaic power generation panels in various periods of the day are selected and collected under daily weather conditions, and the operating temperature of photovoltaic panels and the light intensity of photovoltaic power generation equipment are used as the RBF-BP neural network. As the input of network training, the output power of photovoltaic power generation is used as the output of RBF-BP neural network training, and the sample data of each time period under the same conditions are selected as the test data of RBF-BP neural network;

2) Establishing the RBF-BP neural network for training the normalized sample data, the RBF-BP neural network is composed of two parts of the RBF subnet and the BP subnet to form a double hidden layer RBF-BP combined neural network, which is divided into : Input layer, hidden layer and output layer, the number of nodes in each layer is designed as follows:

Input layer: For the prediction of the maximum power of photovoltaic power generation, without considering the impact of sudden weather conditions and uneven local illumination on the maximum power point of photovoltaic power generation, the selection of neural network input mainly considers two parts, namely Operating temperature of photovoltaic panels and light intensity when photovoltaic power generation equipment is working;

Hidden layer: the RBF-BP neural network involved in the present invention adds the RBF neural network as a sublayer in the hidden layer, and the samples in the input step 1) are first trained through the RBF neural network subnet, and then the training results are used as BP The input of the subnet to train it;

Output layer: the number of nodes in the output layer can be determined according to the situation, but in order to simplify the design of the neural network, the present invention selects an output node, that is, the voltage at the maximum power point;

3) the data collected in step 1) is brought into the RBF-BP neural network set up in step 2), and the error E between the actual output and the expected output is obtained as the fitness value for the genetic algorithm;

4) determine the network structure of RBF-BP neural network according to the fitting function input and output parameters of the RBF-BP neural network built in step 2), determine the number of threshold and weight in RBF-BP neural network, and then Determine the length of the individual for the genetic algorithm;

5) The data collected in step 1) is optimized with a genetic algorithm, and the specific implementation steps are:

A. Population initialization

Each individual in the population contains the ownership value and threshold of the entire RBF-BP neural network. The individual encoding method is real number encoding, and each individual is a real number string. The individual calculates the individual fitness value through the fitness function of the genetic algorithm;

B. Fitness function

According to step 1) the individual obtains the initial weight and the threshold of RBF-BP neural network, with the absolute error value E obtained in step 3) as individual fitness value F;

C. Select operation

Genetic algorithm selection operation has multiple methods such as roulette method, tournament, the present invention selects roulette method, the selection probability _p of each individual i is formula (6):

In the formula, F _i is the fitness value of individual i. Since the smaller the fitness value is, the better, so the reciprocal of the fitness value is calculated before the individual selection; k is the coefficient; N is the number of individuals in the population;

D. Cross operation

Since the individual is coded with a real number, the crossover operation method also adopts the corresponding real number crossover method. The crossover operation method of the kth chromosome a _k and the l chromosome a _l at position j is shown in formula (7):

Where b is a random number between [0, 1];

E. Mutation operation

Select the j-th gene a _ij of the i-th individual to mutate. The formula for the mutation operation is as follows (8):

In formula (8), a _max is the upper bound of gene a _ij ; a _min is the lower bound of gene a _ij ; f(g)=r ₂ (1-g/G _max ) ² ; r ₂ is a random number; g is the current number of iterations; G _max is the maximum number of evolutions; r is a random number between [0, 1].

2. the photovoltaic power generation output power tracking algorithm based on the improved RBF-BP neural network of genetic algorithm according to claim 1, is characterized in that, described step 2) in the transfer of the hidden layer node of RBF-BP subnetwork The function is set to a Gaussian function, as shown in formula (1):

u _i (X)＝exp[-||Xc _i || ² /2σ _i ² ] (i=1, 2, ..., N ₂ ) (1)

Among them, u _i (X) is the output of the i-th hidden layer node, X is the input sample of step (1), ci is the center vector of the Gaussian function, _{σ i} _is the base width parameter of the node, and is a number greater than zero ;

The output of the RBF subnetwork is used as the input of the BP subnetwork in the combined neural network, and the transfer function of the hidden layer nodes is designed as a Sigmoid function, as shown in formula (2):

f(x)＝1/(1+e ^-x ) (2)

Therefore, the output of the hidden layer nodes of the BP subnetwork is formula (3):

Among them, W _ij is the weight connecting the i-th node of the RBF hidden sub-layer to the j-th node of the BP hidden sub-layer, and N ₂ is the number of nodes in the RBF hidden sub-layer.

3. the photovoltaic power generation output power tracking algorithm based on the improved RBF-BP neural network of genetic algorithm according to claim 1, is characterized in that, the output value formula of calculating output layer node in described step 2) is formula (4 ):

Among them, W _jk is the weight connecting the jth node of the BP hidden sublayer to the kth node of the output layer, and _N3 is the number of nodes of the BP hidden sublayer;

f _i Calculate the absolute value of the error of the output layer according to the actual output and the expected output, as shown in formula (5):

Among them, d _k is the desired output, N ₄ is the number of nodes in the output layer, and E is the absolute value of the error.

4. the photovoltaic power generation output power tracking algorithm based on the improved RBF-BP neural network of genetics algorithm according to claim 1, is characterized in that, also comprises the step:

a) Use the training data in step 1) to find the individual corresponding to the optimal fitness value through the genetic algorithm in step 4) through selection, crossover and mutation operations, and the RBF-BP neural network uses the genetic algorithm to obtain the optimal individual for the initial weight of the network value and threshold assignment;

b) test the test data collected in step 1) for the RBF-BP neural network improved by the genetic algorithm, when the neural network training error is less than the target error, the network converges; when the number of network training times is equal to the maximum number of iterations and the training error is still greater than When the target error occurs, the network does not converge; at this time, through the reverse learning ability of the RBF-BP neural network, the weight and threshold of the neural network are reversely modified, and the adjustment formulas are as (9) and (10):

W _jk ＝W _jk +λy _k (1-y _k )(d _k -y _k )O _j (9)

θ _k ＝θ _k +λy _k (1-y _k )(d _k -y _k ) (10)

In the formula, λ is the learning rate, and the trained RBF-BP neural network can be used to track the maximum output power of photovoltaic power generation.