CN106896724B - Tracking system and tracking method for solar tracker - Google Patents

Abstract

The invention provides a tracking system and a tracking method for a sun tracker, wherein an individual optimization is carried out according to sample data through a GA sample data preprocessing module, an individual with the maximum fitness is obtained and is transmitted to a BP network adjusting module to be used as an optimal initial value for training; on the basis of the optimal initial value, reversely transmitting through a BP network adjusting module, adjusting weights and thresholds of a hidden layer and an output layer to enable an obtained error evaluation function to be smaller than a set error threshold, and obtaining adjusted P, I, D parameters; the BP network adjusting module transmits the adjusted P, I, D parameter to the PID controller module, and the PID controller module controls according to the two degrees of freedom, so that the tracking performance and the anti-interference performance are respectively optimal. The invention combines the genetic algorithm and the neural network, fully utilizes the advantages of the genetic algorithm and the neural network, and ensures that the control system not only has the learning function, robustness and generalization capability of the neural network, but also has the global search optimization capability of the genetic algorithm.

Description

Tracking system and tracking method for solar tracker

technical field

The invention relates to the technical field of solar trackers, in particular to a tracking system and a tracking method for solar trackers.

Background technique

In order to accurately track the sun, solar trackers based on solar-occulation flux (SOF) telemetry for solar power generation and mobile vehicle-mounted measurements emerge as the times require. The solar tracker used for solar power generation is fixed-point monitoring. At present, dual-axis solar trackers are mostly used. According to the latitude and longitude information, the real-time azimuth and altitude angles of the sun can be calculated. state to achieve sun tracking. The solar tracker used in the vehicle SOF currently uses the PSD as the feedback element, and the position of the light spot on the PSD is used as the feedback signal to adjust the position of the main reflector, so that the spectrometer at the back end can obtain the maximum light intensity. In the control algorithm, different adjustment step sizes are set according to the size of the PSD output signal.

Among them, x is the adjustment step size, a, b are constants, and a>b, e is the error signal between the actual position of the sun and the center point output by the PSD, and e ₀ is the set threshold constant. The current control algorithm has slow tracking speed, poor accuracy and poor robustness. When the vehicle is bumped or there are interference signals caused by other factors, the tracking effect will be poor or even unable to track. In order to improve the performance of the mobile solar tracker, the GA algorithm is used to optimize the two-degree-of-freedom PID control of the BP network.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the problem that the control algorithm of the sun tracker in the prior art has slow tracking speed, poor accuracy and poor robustness. When the vehicle is bumped or there are interference signals caused by other factors, the tracking effect will be poor or even unable to Tracking defects, a tracking system and a tracking method for a solar tracker are provided to solve the above problems.

In order to achieve the above object, technical scheme of the present invention is as follows:

Tracking system for solar tracker, including GA sample data preprocessing module, BP network adjustment module, PID controller module;

The GA sample data preprocessing module performs individual optimization according to the sample data, obtains the individual with the largest fitness, and transmits it to the BP network adjustment module as the optimal initial value for training;

On the basis of the optimal initial value, reverse transmission through the BP network adjustment module, adjust the weights and thresholds of the hidden layer and the output layer, so that the obtained error evaluation function is smaller than the set error threshold, and the adjusted P, I, D parameters;

The BP network adjustment module transfers the adjusted P, I, D parameters to the PID controller module, and the PID controller module controls the controlled object according to the two degrees of freedom, so that the PID controller has the best tracking performance and anti-interference performance.

Preferably, the tracking system further includes a feedback system module; the feedback system obtains the output signal of the controlled object and the input information of the PID controller, the feedback system obtains the error signal according to the given input signal and the output signal of the controlled object, and Adjust according to the error signal to form a closed-loop control.

Preferably, the GA sample data preprocessing module includes a selection, crossover, and mutation operator operation unit; the selection, crossover, and mutation operator operation unit performs adaptive adjustment on the crossover rate p _c and the mutation rate p _m to obtain the population The individual with the best fitness in the middle, and the individual with the best fitness will directly enter the next generation population, that is, the new population, without any crossover and mutation operations.

Preferably, the GA sample data preprocessing module further includes a catastrophe judgment unit; the catastrophe judgment unit calculates the fitness of the new population, if the catastrophe stops or reaches the set number of catastrophes, the GA algorithm preprocessing is completed, and the The obtained optimal initial weights and thresholds are passed to the BP network.

Preferably, the BP network adjustment module includes a BP network training unit; the BP network training unit obtains the optimized weights and thresholds as initial values, adjusts the initial weights and thresholds according to the momentum-adaptive learning rate, and adopts Levenberg -Marquardt algorithm optimization algorithm to find the optimal PID parameter value.

The present invention also provides a tracking method applied to any of the above-mentioned tracking systems for solar trackers. The method firstly adjusts the PID parameters with the best tracking performance as the control target, and then adjusts the PID parameters with the best anti-interference performance as the control target. The parameters of the control target are adjusted, and the specific adjustment steps are as follows:

1) Sample data preprocessing

First, the sample data is optimized by the GA algorithm, and the individual with the largest fitness is obtained and passed to the BP network as the optimal initial value for training;

2) BP network training

On the basis of the optimal initial value, reverse transmission through the BP network, adjust the weights and thresholds of the hidden layer and the output layer, so that the obtained error evaluation function is less than the set error threshold, and the adjusted P, I, D parameter;

3) PID control

The BP network adjustment module transfers the adjusted P, I, D parameters to the PID controller module, and the PID controller module controls the controlled object according to the two degrees of freedom, so that the PID controller has the best tracking performance and anti-interference performance.

Preferably, the tracking method further includes a signal feedback adjustment process; specifically:

The feedback system obtains the output signal of the controlled object and the input information of the PID controller, the feedback system obtains the error signal according to the input signal of the PID controller and the output signal of the controlled object, and adjusts according to the error signal, forming a closed-loop control .

Preferably, in the step 1), the individual with the best fitness in the population is obtained by adaptively adjusting the crossover rate p _c and the mutation rate p _m through the selection, crossover, and mutation operator operation units, and adapting the The individual with the best degree of degree directly enters the next generation population without any crossover and mutation operations, that is, the new population; the calculation formulas of the adaptive crossover rate p _c and mutation rate p _m are as follows:

Among them, α ₁ , α ₂ are two constants greater than 0, p _c1 , p _c2 , p _m1 , p _m2 are constants obtained by experience, which are 0.85, 0.65, 0.1, 0.001 respectively; f _avg , f _max , f' are the average fitness of the population, the total fitness, and the fitness of the individual.

Preferably, after the new population is obtained, catastrophic judgment is performed on the fitness of the new population. If the catastrophic stops or the set number of catastrophes is reached, the preprocessing of the GA algorithm is completed, and the obtained optimal initial weights and thresholds passed to the BP network.

Preferably, in the step 2), the BP network training unit obtains the optimized weights and thresholds as initial values, adjusts the initial weights and thresholds according to the momentum-adaptive learning rate, and uses the Levenberg-Marquardt algorithm to optimize The algorithm finds the optimal PID parameter value; among them, the calculation formula of momentum-adaptive learning rate adjustment is:

β(k+1)=τ*3 ^λ *β(k)

Among them, β is the learning rate factor, λ is the gradient direction, τ is the learning error coefficient, and w is the weight function;

The calculation formula of the Levenberg-Marquardt algorithm is:

Δw=(J ^T J+μI) ^-1 J ^T e

Among them, e is the error amount, J is the Jacobian matrix of the network error degree to the weight, I is the identity matrix, μ is the scale coefficient, and Δw is the weight increment.

Compared with the prior art, the present invention has the following beneficial effects:

The two-degree-of-freedom ID adopts the best tracking performance and the best anti-interference ability as the control parameters, and adjusts the PID parameters respectively. The combination of genetic algorithm and neural network makes full use of their advantages, so that the control system not only has the learning function, robustness and generalization ability of neural network, but also has the global search and optimization ability of genetic algorithm.

In this patent, the GA algorithm is used to optimize the BP network to complete the tuning of the PID parameters according to the control parameters, so as to obtain the most accurate P, I, D parameters in the shortest time, so as to input more accurate control signals. As far as the instrument is concerned, it is to achieve real-time accurate tracking of the sun.

The momentum term in the heuristic learning rule is combined with the adaptive learning rate and the numerical optimization learning method. Each improvement method is complementary, which effectively improves the slow convergence speed of the BP network and has fallen into a minimum value. Greatly improved the control precision and speed in the later stage.

This algorithm is used to control the closed-loop sun tracking system, and has the characteristics of high tracking accuracy, short adjustment time and strong anti-interference ability.

Description of drawings

1 is a block diagram of a module structure of Embodiment 1 of the present invention;

2 is a system principle block diagram of Embodiment 1 of the present invention;

Fig. 3 is a 4:7:3BP network topology structure diagram in Embodiment 2 of the present invention;

Fig. 4 is the algorithm flow chart of Embodiment 2 of the present invention;

5 is a block diagram of an algorithm function module according to Embodiment 2 of the present invention;

FIG. 6 is a control loop diagram of the solar tracker in Embodiment 3 of the present invention.

Detailed ways

In order to have a further understanding and understanding of the structural features of the present invention and the effects achieved, the preferred embodiments and accompanying drawings are used in conjunction with detailed descriptions, and the descriptions are as follows:

Concept description: The sample in the following is a combination of all individuals, and the population is a combination of a certain number of individuals, which is determined by the set population size. The individual is the initial value of the neural network or randomly generated training data.

Example 1

As shown in Figure 1 and Figure 2, the tracking system for solar trackers includes a GA sample data preprocessing module, a BP network adjustment module, a PID controller module, and a feedback system module.

The GA sample data preprocessing module performs individual optimization according to the sample data, obtains the individual with the largest fitness and transmits it to the BP network adjustment module as the optimal initial value for training;

On the basis of the optimal initial value, reverse transmission through the BP network adjustment module, adjust the weights and thresholds of the hidden layer and the output layer, so that the obtained error evaluation function is smaller than the set error threshold, and the adjusted P, Parameters of I and D;

The BP network adjustment module transfers the adjusted parameters of P, I and D to the PID controller module, and the PID controller module controls according to the two degrees of freedom, so as to achieve the best tracking performance and anti-interference performance respectively.

The output signal of the controlled object is transmitted to the input end of the PID controller through the feedback system module. The feedback system obtains the error signal according to the input signal of the PID controller and the output signal of the controlled object, and adjusts according to the error signal to form a closed-loop control. .

Among them, the GA sample data preprocessing module includes a unit for setting the number of iterations and initial values of parameters, a coding unit, a fitness calculation unit, a selection/crossover/mutation operator operation unit, and a disaster judgment unit; the BP network adjustment module includes a network structure determination unit, Initialize the network unit, obtain the initial value unit, BP network training unit, and accuracy judgment unit.

为基础，经过不断地仿真调试，确定隐含层神经元的个数为7个，即BP网络的拓扑结构为4:7:3，如图3所示。初始化网络单元对BP网络的权值和阈值初始值以及误差评价函数和误差阈值，并计算在初始值的情况下误差信号是否小于误差阈值，若是，则BP网络无需训练，直接进入测试阶段，否则，进入训练阶段。First, initialize the iteration number n, crossover rate _pc , mutation rate pm, and population size _m of the GA algorithm by setting the number of iterations and the parameter initial value unit; then determine the topology structure of the BP network through the network structure determination unit. : input layer, hidden layer, output layer; the input layer has 4 neurons for the specified input signal r _in (k), the actual output signal y _out (k), the error amount e (k) and the constant a; output The three neurons of the layer correspond to the three parameters k _p , k _i , k _d of the PID control respectively; and the number of neurons in the hidden layer is based on the empirical formula

Based on continuous simulation and debugging, it is determined that the number of neurons in the hidden layer is 7, that is, the topology of the BP network is 4:7:3, as shown in Figure 3. Initialize the weights and thresholds of the network unit to the initial value of the BP network, as well as the error evaluation function and error threshold, and calculate whether the error signal is less than the error threshold under the initial value. , into the training phase.

If the training stage is entered, the population containing the initial value of the BP network is encoded with real numbers by the coding unit, and the fitness function is defined as the difference between the expected data and the actual output data, and then the fitness calculation unit is used to calculate the fitness, specifically: The GA operation includes selection operator, crossover operator, and mutation operator; in which the crossover rate p _c and the mutation rate p _m use the adaptive method to obtain the individual with the best fitness in the population, and then the individual with the best fitness is not selected. After any crossover and mutation operations, it directly enters the next generation population, that is, the new population, and the probability of other individuals entering the next population is proportional to their relative fitness in the entire population; the adaptive crossover rate p _c and mutation rate p _m are used. The calculation formula is:

Among them, α ₁ and α ₂ are two constants greater than 0, p _c1 , p _c2 , p _m1 , and p _m2 are constants obtained according to experience, which are 0.85, 0.65, 0.1, and 0.001 respectively; f _avg , f _max , f' is the average fitness of the population, the total fitness, and the fitness of the individual, and finally the total fitness of the new population is calculated by the catastrophe judgment unit. If there is no better individual for n consecutive generations, this means that The GA algorithm has a precocious thread and falls into a local minimum. At this time, a catastrophe needs to be carried out to kill all the outstanding individuals of the current population and enter the next generation. If after several disasters, the fitness value is the same as before no disasters, then stop the disasters; if the disasters stop or reach the set number of disasters, the GA algorithm preprocessing is completed, and the obtained optimal initial weights and thresholds are passed to the acquisition. The initial value unit, the obtaining initial value unit transmits the obtained initial value to the BP network training unit. The BP network training unit adjusts the initial weights and thresholds according to the momentum-adaptive learning rate, and uses the Levenberg-Marquardt algorithm to find the optimal PID parameter values;

The momentum-adaptive learning rate adjustment is calculated as:

β(k+1)=τ*3 ^λ *β(k)

Among them, β is the learning rate factor, λ is the gradient direction, τ is the learning error coefficient, and w is the weight function;

The calculation formula of the Levenberg-Marquardt algorithm is:

Δw=(J ^T J+μI) ^-1 J ^T e

Among them, e is the error amount, J is the Jacobian matrix of the network error degree to the weight, I is the unit matrix, μ is the proportional coefficient, Δw is the weight increment; when μ is large, it is the Gauss-Newton algorithm, When μ is small, it is close to the gradient descent method. The accuracy judgment unit performs the final evaluation. When the error evaluation function of the BP network is less than the error threshold, the learning process of the entire BP network is completed, and then the learning effect and generalization ability of the BP network are tested with the test data.

Example 2

As shown in FIG. 4 and FIG. 5 , the tracking method for a solar tracker is applied to the tracking system provided in Embodiment 1. The algorithm firstly adjusts the PID parameters with the best tracking performance as the control goal, and then adjusts the parameters with the best anti-interference performance as the control goal. The specific adjustment steps are as follows:

Step 1. Randomly generate 2000 sets of data, of which 1500 sets of data are used to train the BP network, and the other 500 sets of data are used to test the BP network, and the data is normalized;

Step 2. Determine the BP network structure

为基础，经过不断地仿真调试，确定隐含层神经元的个数为7个，即BP网络的拓扑结构为4:7:3，如图3所示。First, determine that the network has three layers in total: input layer, hidden layer, and output layer; the input layer has 4 neurons, which are the specified input signal r _in (k), the actual output signal y _out (k), and the error amount e ( k) and constant a; the three neurons in the output layer correspond to the three parameters k _p , k _i , k _d of the PID control respectively; and the number of neurons in the hidden layer is based on the empirical formula

Based on continuous simulation and debugging, it is determined that the number of neurons in the hidden layer is 7, that is, the topology of the BP network is 4:7:3, as shown in Figure 3.

Step 3. Initialize the BP network

Set the initial value of the weights and thresholds of the BP network, the error evaluation function and the error threshold, and calculate whether the error signal is less than the error threshold under the initial value. If so, the BP network does not need to be trained, and directly enters the testing phase, otherwise, enter the step 4); where the initial value is the weight and the initial value of the threshold;

Step 4. Initialize the evolution times n, crossover rate _pc , mutation rate pm, and population size _m of the GA algorithm;

Step 5. Real-number coding is performed on the population containing the initial value of the BP network, and the fitness function is defined as the difference between the expected data and the actual output data;

Step 6. Perform GA operations

The GA operation includes selection operator, crossover operator, and mutation operator; in which the crossover rate p _c and the mutation rate p _m use the adaptive method to obtain the individual with the best fitness in the population, and then the individual with the best fitness is not selected. After any crossover and mutation operations, it directly enters the next generation population, that is, the new population, and the probability of other individuals entering the next population is proportional to their relative fitness in the entire population; the adaptive crossover rate p _c and mutation rate p _m are used. The calculation formula is:

Among them, α ₁ , α ₂ are two constants greater than 0, p _c1 , p _c2 , p _m1 , p _m2 are constants obtained by experience, which are 0.85, 0.65, 0.1, 0.001 respectively; f _avg , f _max , f' are the average fitness of the population, the total fitness, and the fitness of the individual;

Step 7. Catastrophe judgment, calculate the total fitness of the new population generated, if there is no better individual for n consecutive generations, it means that the GA algorithm has a premature thread and falls into a local minimum value. All outstanding individuals of the current population will be killed and will enter the next generation. If after several disasters, the fitness value is the same as before the disaster, then stop the disaster; if the disaster stops or reaches the set number of disasters, the GA algorithm is optimized, and the obtained optimal initial weights and thresholds are passed to the BP network ;

Step 8. BP network training

According to the received weights and thresholds optimized by the GA algorithm as the initial values, adjust the initial weights and thresholds according to the momentum-adaptive learning rate, and use the Levenberg-Marquardt algorithm to find the optimal PID parameter values;

The momentum-adaptive learning rate adjustment is calculated as:

β(k+1)=τ*3 ^λ *β(k)

Among them, β is the learning rate factor, λ is the gradient direction, τ is the learning error coefficient, and w is the weight function;

The calculation formula of the Levenberg-Marquardt algorithm is:

Δw=(J ^T J+μI) ^-1 J ^T e

Among them, e is the error amount, J is the Jacobian matrix of the network error degree to the weight, I is the unit matrix, μ is the proportional coefficient, Δw is the weight increment; when μ is large, it is the Gauss-Newton algorithm, When μ is small, it is close to the gradient descent method.

Step 9. When the error evaluation function of the BP network is less than the error threshold, the learning process of the entire BP network is completed, and then the learning effect and generalization ability of the BP network are tested with the test data.

Example 3

This embodiment will introduce the evolution process from the prior art to the algorithm provided in Embodiment 2.

two degrees of freedom

The main control indicators of the PID control system are: external disturbance suppression and target tracking characteristics. When one degree of freedom control is used, the two show opposite trends, and they achieve optimal performance at the same time. Two degrees of freedom PID (Two degree of freedom PID, 2DOF PID) control is to set the PID parameters with the optimal target tracking characteristics and the optimal external disturbance suppression characteristics respectively, so as to achieve the best performance of the entire control system.

When using 2DOF PID control, it needs to meet the requirements of easy to understand, simple structure, good combination with traditional technology and can inherit its technical achievements. In the control loop for the solar tracker, the target value filter type 2DOFPID control is used. The control loop is shown in Figure 6.

The target value filter H(s) can be expressed as:

Among them, α is the 2-DOF coefficient of the proportional gain (generally 0≤α≤1), β is the 2-DOF coefficient of the integral time (generally 0≤β<1), and γ is the 2-DOF coefficient of the differential time (generally 0≤γ<2), 1/η is the differential gain (generally 0.1≤γ≤1).

The PID control of the solar tracker is carried out by adopting the method with variable two-degree-of-freedom coefficient. The adjustment steps are: adding external disturbance, adjusting K _p , K _i , and K _d to make the external disturbance suppression characteristic the best; according to Chian-Hrone- The Reswick (CHR) adjustment method initially obtains the values of α and γ. According to engineering experience, β = 0.15. According to the change of the target value, fine-tune the two-degree-of-freedom coefficients α, β, and γ near the set value to make the target track. Features are the best.

Two degrees of freedom can not only improve the control characteristics, but also easy to adjust, only need to increase the target value filter, in the case of ensuring the simple structure, can achieve the best control performance.

BP network

The artificial neural network appeared in the 1940s. It is formed by connecting a large number of neurons with adjustable connection weights. It has achieved good application results in the field of intelligent control. As the most widely used artificial neural network model, the error reverse Propagation neural network (BPNN) has the ability to realize any complex nonlinear mapping, and can approximate any nonlinear continuous function with any precision; parallel distributed processing mode; has self-learning ability; the ability to apply this set of weights to general situations; the ability of data fusion, which can process quantitative and qualitative information at the same time; the ability to be used in multivariate systems and online learning.

The basic idea of BP algorithm is the least square method, which uses gradient search technology to minimize the mean square error between the actual output value and the expected output value of the network. The learning process of the algorithm consists of forward propagation of information and back propagation of error. The input information is transmitted from the input layer to the output layer after being processed layer by layer through the hidden layer, and the state of each layer of neuron nodes only affects the state of the next layer of neurons. When the output layer cannot get the expected output, it turns to backpropagation, returns the error signal along the original connection path, and modifies the connection weights and closed values of neurons in each layer to make the error function follow the direction of the negative gradient. decrease, and finally the error between the actual output value and the expected output value is minimized.

The BP neural network has an input layer, a hidden layer and an output layer, and each layer has a different number of neurons. According to the structure and complexity of the research object, the topology of the neural network is also different. In this system, an input layer, a hidden layer, and an output layer are adopted, and the topology structure of the number of neurons is 3:5:3.

The three input signals are the specified input signal r _in (k), the actual output signal y _out (k), and the error amount e (k). The three output signals correspond to the three parameters k _p , k _i , and k _d of the PID control respectively.

The learning process of BP network consists of two parts: forward propagation and back propagation. The first is the forward propagation process of the signal. The information of the input layer is propagated forward to the nodes of the hidden layer, after the activation function operation of each unit, and the information of the hidden layer is transmitted to the nodes of the output layer, and then passes through the output layer. The operation output of the activation function of each node above. During forward propagation, the neuron state of each layer only affects the next layer of neuron network. If the error between the actual output and the expected output value is greater than the set error function threshold, turn to the backpropagation process, backpropagate the error signal, modify the weights of neurons in each layer successively, and then get the correction through forward propagation After the output, the two processes are applied iteratively to minimize the error signal. When the actual error signal is less than the set error threshold, the learning process of the network ends.

In the forward propagation process, the input and output of the hidden layer nodes are:

The input and output of the output layer are:

分别为隐含层和输出层的输出值。f(k),g(k)为激活函数，用来加入非线性因素，从而弥补线性模型表达力不够的缺点，激活函数的选择需满足单调递增有界且一阶可微的特点，经过仿真，隐含层的激活函数双极S形函数：w _ij , w _jk are the connection weights between the input layer and the hidden layer, and the hidden layer and the output layer, respectively.

are the output values of the hidden layer and the output layer, respectively. f(k), g(k) are activation functions, which are used to add nonlinear factors to make up for the lack of expressiveness of linear models. The selection of activation functions should meet the characteristics of monotonically increasing bounded and first-order differentiable. After simulation , the activation function of the hidden layer bipolar sigmoid function:

The error evaluation function is expressed as:

令

则

According to the error evaluation function, the weight of each node is adjusted through the back-propagation process, and the correction of the weight is proportional to the negative gradient direction of the error, namely:

make

but

The weight adjustment between the hidden layer and the input layer is:

有Δw_ij＝ηδ_jx_i make

have Δw _ij = ηδ _j x _i

Among them, η is the learning step size, that is, the learning rate. The weight between the hidden layer and any node of the output layer in the next iteration, and the weight between the input layer and the hidden layer are:

w _jk (k)=Δw _jk +w _jk (k)

w _ij (k)=Δw _ij +w _ij (k)

However, this standard BP neural network is actually a gradient descent search method, which has the shortcomings of slow convergence and easy to fall into local minima, and these shortcomings can be corrected by using the improved BP algorithm. Algorithms can be divided into two categories, one is using heuristic learning rules, such as adding additional momentum terms, using adaptive learning rate, etc.; the other is learning methods based on numerical optimization, such as conjugate gradient method, quasi-Newton method and Levenberg-Marquardt method method (L-M method for short),

1. Add momentum term

In order to speed up the convergence speed, add the last correction weight coefficient to the weight correction, and use it as one of the basis for this correction, namely:

Δw _ij (k+1)=ηδ _j x _i +αΔw _ij (k)

where α is the momentum factor.

2. Momentum-Adaptive Learning Rate Adjustment Algorithm

In the self-adaptive adjustment of the learning rate, the basic idea is to increase η to shorten the learning time in the case of learning convergence; when η is too large to converge, decrease η in time until convergence.

β(k+1)=τ*3 ^λ *β(k)

Among them, β is the learning rate factor, λ is the gradient direction, τ is the learning error coefficient, and w is the weight function.

3. Using the Levenberg-Marquardt algorithm

The L-M method is essentially a combination of the gradient descent method and the Newton method, and has a high convergence speed when the network weights are small. The weight adjustment rate of the L-M optimization algorithm is:

Δw=(J ^T J+μI) ^-1 J ^T e

Among them, e is the error amount, J is the Jacobian matrix of the network error degree to the weight, I is the unit matrix, μ is the scale coefficient, Δw is the weight increment. When μ is large, it is Gauss-Newton algorithm, and when μ is small, it is close to the gradient descent method.

The addition of the momentum term improves the convergence speed, but it is difficult to choose the learning rate and requires a high initial value; the adaptive learning rate greatly improves the disadvantage of being easily trapped in local minima, but the convergence speed is slow.

The Levenberg-Marquardt algorithm greatly improves the convergence speed, but it cannot improve the problem of being easily trapped in local minima. The momentum term in the heuristic learning rule is combined with the adaptive learning rate and the numerical optimization learning method. These improved methods complement each other, effectively improve the shortcomings of the BP network that the convergence speed is slow and has fallen into a minimum value, and greatly improves the control accuracy and speed in the later stage.

The improved BP algorithm reduces the probability of entering the local minimum value, but does not fundamentally realize fast global search. GA algorithm GA is a search technology that simulates the natural biological evolution mechanism and has the ability to search globally. Therefore, the GA algorithm is used. The mode combined with the BP network can better adjust the network weights and achieve fast and high-precision PID control.

GA algorithm:

The main idea of the GA-BPNN algorithm: firstly use the GA algorithm to quickly search for the vicinity of the optimal solution, locate a better search space in the solution space, and provide a better initial value for the BP network, and then use the BP algorithm The optimal solution is found in this small search space. The switching between the GA algorithm and the BP algorithm can be realized by the size of the error. If the error is greater than a certain value, the GA algorithm is used, and when the error is smaller than the value, the BP algorithm is used until the limited accuracy or the maximum number of steps is reached.

The simple GA algorithm uses the selection operator, the crossover operator, and the mutation operator. The GA process is as follows:

1. Determine the encoding method: realize the conversion from the solution space to the search space, and the encoding form is:

分别为输入层与隐含层，隐含层与输出层之间的连接权值。in,

are the connection weights between the input layer and the hidden layer, and the hidden layer and the output layer, respectively.

2. Determine the parameters: population size, crossover, mutation probability, number of termination iterations;

3. Initialization: randomly generate n individuals to provide the initial population P(0);

4. Evaluation: Decode to the solution space, calculate the candidate solutions, the fitness of the solution set, and the average fitness.

5. GA operation: including selection operator, crossover operator, mutation operator, operate on the population P(t) to generate the next generation

P(t+1).

6. Repeat 4 and 5 until the parameters converge or reach the predetermined target.

The simple GA algorithm is prone to precociousness, that is, premature convergence of the local optimal solution. Therefore, the crossover rate and mutation rate need to be adaptively adjusted. The calculation formulas of the improved adaptive _crossover rate _pc and mutation rate pm are:

α ₁ , α ₂ are two constants greater than 0, p _c1 , p _c2 , p _m1 , and p _m2 are constants obtained from experience, which are 0.9, 0.6, 0.1, and 0.001, respectively.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above-mentioned embodiments. The above-mentioned embodiments and descriptions describe only the principles of the present invention. Without departing from the spirit and scope of the present invention, there are various Variations and improvements are intended to fall within the scope of the claimed invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims (9)

1. for the tracking system of solar tracker, it is characterized in that: comprise GA sample data preprocessing module, BP network adjustment module, PID controller module; The GA sample data preprocessing module performs individual optimization according to the sample data, obtains the individual with the largest fitness, and transmits it to the BP network adjustment module as the optimal initial value for training; Specifically, the individual with the best fitness in the population is obtained by adaptively adjusting the crossover rate p _c and the mutation rate p _m through the selection, crossover, and mutation operator operation units, and the individual with the best fitness is selected without any crossover, The mutation operation directly enters the next generation population, that is, the new population; the calculation formulas of the adaptive crossover rate p _c and mutation rate p _m used are:

Among them, α ₁ , α ₂ are two constants greater than 0, p _c1 , p _c2 , p _m1 , p _m2 are constants obtained by experience, which are 0.85, 0.65, 0.1, 0.001 respectively; f _avg , f _max , f' are the average fitness of the population, the total fitness, and the fitness of the individual; On the basis of the optimal initial value, reverse transmission through the BP network adjustment module, adjust the weights and thresholds of the hidden layer and the output layer, so that the obtained error evaluation function is smaller than the set error threshold, and the adjusted P, I, D parameters; The BP network adjustment module transfers the adjusted P, I, D parameters to the PID controller module, and the PID controller module controls the controlled object according to the two degrees of freedom, so that the PID controller has the best tracking performance and anti-interference performance. 2. The tracking system for a solar tracker according to claim 1, wherein the tracking system further comprises a feedback system module; the feedback system module obtains the output signal of the controlled object and the input information of the PID controller, The error signal is obtained and adjusted according to the error signal to form a closed-loop control. 3. The tracking system for a solar tracker according to claim 1, wherein the GA sample data preprocessing module comprises a selection, crossover, and mutation operator operation unit; the selection, crossover, and mutation operators The operating unit adjusts the crossover rate p _c and the mutation rate p _m adaptively to obtain the individual with the best fitness in the population, and directly enters the next generation of the population without any crossover and mutation operation, that is, the new population. population. 4. The tracking system for a solar tracker according to claim 3, wherein the GA sample data preprocessing module further comprises a catastrophe judgment unit; the catastrophe judgment unit calculates the fitness of the new population, If the disaster stops or reaches the set number of disasters, the preprocessing of the GA algorithm is completed, and the obtained optimal initial weights and thresholds are passed to the BP network. 5. The tracking system for a solar tracker according to claim 4, wherein the BP network adjustment module comprises a BP network training unit; the BP network training unit obtains an optimized weight and a threshold as an initial The initial weight and threshold are adjusted according to the momentum-adaptive learning rate, and the Levenberg-Marquardt algorithm is used to find the optimal PID parameter value. 6. A tracking method applied to the tracking system of any one of claims 1 to 5 for a solar tracker, wherein the method firstly adjusts the PID parameters with the best tracking performance as the control target, and then adjusts the PID parameters. To adjust the parameters with the best anti-interference performance as the control goal, the specific adjustment steps are as follows: 1) Sample data preprocessing First, the sample data is optimized by the GA algorithm, and the individual with the largest fitness is obtained and passed to the BP network as the optimal initial value for training; In the step 1), specifically, the individual with the best fitness in the population is obtained by adaptively adjusting the crossover rate _{p c} and the mutation rate pm through the selection, crossover, and mutation operator operation units, and the optimal fitness is determined. The individuals of , directly enter the next generation population without any crossover and mutation operations, that is, the new population; the calculation formulas of the adaptive crossover rate p _c and mutation rate p _m are:

Among them, α ₁ , α ₂ are two constants greater than 0, p _c1 , p _c2 , p _m1 , p _m2 are constants obtained by experience, which are 0.85, 0.65, 0.1, 0.001 respectively; f _avg , f _max , f' are the average fitness of the population, the total fitness, and the fitness of the individual; 2) BP network training On the basis of the optimal initial value, reverse transmission through the BP network, adjust the weights and thresholds of the hidden layer and the output layer, so that the obtained error evaluation function is smaller than the set error threshold, and the adjusted P, I, D parameter; 3) PID control The BP network adjustment module transfers the adjusted P, I, D parameters to the PID controller module, and the PID controller module controls the controlled object according to the two degrees of freedom, so that the PID controller has the best tracking performance and anti-interference performance. 7. The tracking method for a tracking system of a solar tracker according to claim 6, wherein the tracking method further comprises a signal feedback adjustment process; specifically: The feedback system obtains the output signal of the controlled object and the input information of the PID controller. The feedback system obtains the error signal according to the given input signal and the output signal of the controlled object, and adjusts according to the error signal to form a closed-loop control. 8. The tracking method for a tracking system of a solar tracker according to claim 6, characterized in that: after obtaining the new population, the fitness of the new population is judged by catastrophe. If the number of catastrophes is determined, the preprocessing of the GA algorithm is completed, and the obtained optimal initial weights and thresholds are passed to the BP network. 9. the tracking method for the tracking system of solar tracker according to claim 8, is characterized in that: in described step 2), BP network training unit obtains the weight after optimization and threshold value as initial value, according to momentum - Adaptive learning rate adjusts the initial weights and thresholds, and uses the Levenberg-Marquardt algorithm optimization algorithm to find the optimal PID parameter value; among them, the momentum-adaptive learning rate adjustment formula is:

β(k+1)=τ*3 ^λ *β(k) Among them, β is the learning rate factor, λ is the gradient direction, τ is the learning error coefficient, and w is the weight function; The calculation formula of the Levenberg-Marquardt algorithm is: Δw=(J ^T J+μI) ^-1 J ^T e Among them, e is the error amount, J is the Jacobian matrix of the network error degree to the weight, I is the identity matrix, μ is the scale coefficient, and Δw is the weight increment.

Images