A fan MPPT control method based on wind speed prediction and PID-DHDP

Abstract

The application relates to the technical field of fan MPPT control, in particular to a fan MPPT control method based on wind speed prediction and PID-DHDP, and discloses a fan MPPT control method based on wind speed prediction and PID-DHDP, which comprises the following steps: acquiring wind speed, temperature, humidity and other data in real time and performing normalization processing on the data; and adopting a wolf swarm algorithm to optimize GRU (Gated Recurrent Unit) hyperparameters. The fan MPPT control method based on wind speed prediction and PID-DHDP predicts the wind speed through a GRU network, obtains a wind speed prediction sequence value, can conveniently calculate a fan rotating speed reference value through the wind speed prediction value and an optimal tip speed ratio, takes a deviation signal between the fan rotating speed reference value and an actual fan rotating speed value as a PID-DHDP controller input signal, constitutes a closed-loop rotating speed control system, realizes that the fan rotating speed is stably at an optimal rotating speed, the fan outputs maximum power, and thus the fan MPPT control is realized.

Description

Technical Field

[0001] This invention relates to the field of wind turbine MPPT control technology, specifically a wind turbine MPPT control method based on wind speed prediction and PID-DHDP. Background Technology

[0002] With the severe situation of global warming and energy shortage, countries have begun to focus on the new energy industry. At the same time, various new materials, technologies and processes are being used to manufacture wind turbines to improve their performance. People, businesses and governments are actively responding to low-carbon living and low-carbon economy, which has greatly promoted the popularization of wind power generation.

[0003] By adjusting the rotor speed of wind turbines, the tip speed ratio of the turbine blades can be maintained near the theoretically optimal tip speed ratio, achieving maximum power point tracking (MPPT). This allows for full utilization of wind power and improves the economic and social benefits for wind farm companies. Given the random, intermittent, and fluctuating nature of wind speed, the core issue is how to enhance the wind turbine's ability to respond to rapid changes in wind speed and maintain the turbine in MPPT mode as much as possible. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a wind turbine MPPT control method based on wind speed prediction and PID-DHDP. It has the advantages of wind turbine MPPT control and solves the problem of how to improve the wind turbine's response and tracking ability to rapid changes in wind speed, and keep the wind turbine operating in MPPT state as much as possible, given the randomness, intermittency and fluctuation of wind speed.

[0006] (II) Technical Solution

[0007] To achieve the aforementioned objective of wind turbine MPPT control, this invention provides the following technical solution: a wind turbine MPPT control method based on wind speed prediction and PID-DHDP, comprising the following steps:

[0008] 1) Acquire real-time data such as wind speed, temperature, and humidity, and normalize this data;

[0009] 2) The Wolf Pack Algorithm (WPA) is used to optimize the hyperparameters of the gated recurrent network (GRU).

[0010] 3) The normalized data is used as the input data for the GRU network, and the GRU network outputs the wind speed prediction sequence values;

[0011] 4) Calculate the reference value of the fan speed by using the predicted wind speed and the optimal tip speed ratio. Use the deviation signal between the reference value of the fan speed and the actual fan speed as the input signal of the PID (Proportional-Integral-Derivative) control-DHDP direct heuristic dynamic programming controller to form a closed-loop speed control system. This system can stabilize the fan speed at the optimal speed and output the maximum power of the fan, thereby realizing MPPT control of the fan.

[0012] Preferably, the hidden layer in the GRU contains only reset gates and update gates. These gates control the degree to which information is transferred, and the input to both gates is the current input x. t and the hidden state h from the previous moment t-1 .

[0013] Preferably, the reset gate is responsible for determining how much memory information needs to be retained, and the input to the reset gate is the input at the current moment. The hidden layer state h at the previous time step t-1 It is composed of concatenation, and its output is a number in the interval (0, 1), representing the state h of the hidden layer at the previous time step. t-1 The proportion to be retained, and the output formula for resetting the door are:

[0014] (1)

[0015] After discarding and remembering the input information through two gate structures, GRU calculates the candidate hidden state values. Its calculation expression is shown in equation (2) below:

[0016] (2)

[0017] in Let x be the tanh activation function. t h represents the input at time t. t-1 The vector represents the hidden state at the previous time step, and [ ] indicates that the two vectors are connected. h w r Let represent the weight matrix, and σ() represent the sigmoid function.

[0018] Preferably, the input to the update gate is the input at the current time. The hidden layer state h at the previous time step t−1 It is constructed by concatenation, and its output is a number in the interval (0, 1). The output formula of the update gate is shown in formula (3):

[0019] (3)

[0020] After obtaining the updated state information through the update gate, the tanh activation function creates a vector of all possible values ​​based on the new input and calculates the candidate hidden state values. Then, the final state h at the current moment is calculated through the network. t The calculation formula is shown in equation (4):

[0021] (4)

[0022] According to the calculation formula above, GRU stores and filters information through two gates, retains important features through the gate function, and captures dependencies through learning, thereby obtaining the optimal output value.

[0023] Preferably, the detailed process of the wolf pack algorithm includes the following steps:

[0024] 1) Wolf Pack Initialization

[0025] The artificial wolf initializes its position in the solution space using random values. Assuming the iteration number k=0, X... i Let X represent the i-th artificial wolf. min X represents the lower bound of the variable's value space. max This represents the upper bound of the variable's value space.

[0026] (5)

[0027] Where rand∈(0,1) represents a random number between 0 and 1, assuming it exists in a D-dimensional space (D is the dimension of the space of the variable to be optimized), then X i It can be represented as:

[0028] (6)

[0029] 2) Rules for the emergence of the alpha wolf

[0030] The prey odor concentration at the location of each artificial wolf is calculated, and the artificial wolf with the highest prey odor concentration is selected as the alpha wolf. After each iteration, the wolf pack compares the best artificial wolf after the iteration with the alpha wolf and selects the wolf with the higher odor concentration as the alpha wolf.

[0031] 3) Wandering behavior

[0032] Within the solution space, select the T with the best fitness, excluding the alpha wolf. n One artificial wolf acts as a scout, responsible for searching for prey, among which T n The value is a random integer between [S / (α+1), S / α], where S is the total number of artificial wolves in the wolf pack, and α is the proportion of scout wolves among the artificial wolves. During the prey search process, the scout wolves first calculate the prey odor concentration Y at the location of each scout wolf.i If the odor concentration Y i >Y leader Then we can obtain Y leader =Y i That is, the scout wolf i takes the place of the alpha wolf and initiates a summoning action to all the wolves; if Y i <Y leader The wolf will move in the direction of h with a fixed step size s. a Swim forward one step, and take the current position of Detective Wolf i as the starting point. Then Detective Wolf i will move along... The position will be updated after the direction is moved:

[0033] (7)

[0034] At this moment, the scent concentration of the prey perceived by the wolf i is Y. ip After making a decision, the wolf detects the direction along the path with the highest odor concentration (Y) that is greater than the current odor concentration. i Move forward one step in the direction, while simultaneously adjusting its position X i Update. Repeat the above steps until the wolf i detects the concentration Y of the prey's scent. i >Y leader Or, the number of times the wolf roams (T) reaches the maximum number of roams (T). max ;

[0035] 4) Summoning behavior

[0036] After receiving information about prey from the scout wolves, the alpha wolf in the pack howls to summon the surrounding wolves. In the algorithm, the wolves with the lowest fitness scores are chosen as the alpha wolves, and their number is M. n =ST n -1. After receiving the summoning command, the wolf runs towards the prey's location with a large stride. Therefore, after k1 iterations, the position of wolf j in d-dimensional space can be represented as:

[0037] (8)

[0038] in, This represents the position of the alpha wolf in the d-th dimension after the k-th iteration. This indicates the current spatial location of the wolf j. This indicates that as the wolf gradually approaches the alpha wolf, during a pack attack, if the concentration of the prey's scent perceived by wolf j is greater than that perceived by the alpha wolf, that is, Y... j >Y leader Wolf j will take the position of the alpha wolf; otherwise, wolf j will continue its running maneuver until it reaches a spatial distance d from the alpha wolf's location. jm Less than the distance criterion d nearAt this point, the wolf begins to switch from a running attack to a surrounding attack, and the distance determination factor is calculated using the formula (9):

[0039] (9)

[0040] Here, ω is the distance decision factor, and its magnitude determines the convergence speed of the wolf pack algorithm. Under the same conditions, the larger the value of the decision factor, the faster the convergence speed of the wolf pack algorithm. However, if ω is too large, it will affect the wolves' siege behavior, causing them to fail to enter the siege state. d and max d Let represent the minimum and maximum values ​​in the d-th dimension space to be optimized, respectively;

[0041] 5) Besieging behavior

[0042] In a siege, the position of the alpha wolf closest to the prey is represented as the prey's position. For the k-th generation of artificial wolves, the prey's position in the d-th dimension can be represented as... Then, during the kth iteration, the position of the wolf pack after the siege can be represented as shown in equation (10):

[0043] (10)

[0044] Where λ is a random constant between [-1, 1], Let be the stride length of a wolf when it launches an attack on its prey. When wolf i is attacking, if the scent concentration of the prey it senses is greater than the scent concentration at its original position, then wolf i will update its position. Otherwise, the position of wolf i will not change. There is a certain relationship between the stride lengths of the three intelligent behaviors in d-dimensional space, as shown in Equation (11):

[0045] (11)

[0046] Where C represents the step size factor of intelligent behavior, the magnitude of which affects the level of refinement in searching for prey;

[0047] 6) Wolf Pack Renewal Rules

[0048] The wolf pack is updated according to the law of "survival of the fittest". In the wolf pack, the relatively weak R wolves will be eliminated due to the competition mechanism. At the same time, R artificial wolves will be randomly generated. The value of R is in the range of [n / 2β, n / β], where β is the wolf pack update ratio factor.

[0049] Preferably, the PID-DHDP consists of three networks: an Actor network (ANN), a Critic network (CNN), and a Reinforcement network (RNN).

[0050] The ANN network module receives the input variable X(t) and outputs the control variable U(t), which is then returned to the controlled system for control.

[0051] The RNN network module receives the input variable X(t) of the controlled system and the control variable U(t) of the network module, and the RNN network module obtains the internal reinforcement learning function R(t).

[0052] The CNN network module receives the input variable X(t) of the controlled system, the control variable U(t) of the ANN network module, and the internal reinforcement learning function R(t) of the RNN network module. The CNN network module obtains the performance index function J(t).

[0053] The performance index function J(t) is combined with the target expectation function Uc(t) to obtain the first time-series differential signal. , ;

[0054] The performance index function J(t) is combined with the performance index function J(t-1) and the external reinforcement learning function r(t) through the multiplier α to obtain the second time-series difference signal. , ;

[0055] The performance index function J(t) is combined with the performance index function J(t-1) and the internal reinforcement learning function R(t) through multiplier α to obtain the third time-series differential signal. , ;

[0056] The ANN network, RNN network, and CNN network update their respective weight coefficients online using the corresponding first, second, and third time-series difference signals.

[0057] Preferably, the input X of the CNN network c (t) is:

[0058] (12)

[0059] The error function of the CNN network is defined as shown in equation (13):

[0060] (13)

[0061] Where λ is the discount factor, 0 < λ < 1,

[0062] The transfer function of the hidden layer neurons in the CNN network adopts the bipolar sigmoid function, as shown in Equation (14):

[0063] (14)

[0064] The output is the performance index function J(t). Its hidden layer uses the sigmoid activation function, while the output layer uses the linear activation function. It is easy to see that the input and output of the hidden and output neurons of the CNN network are shown in equations (15), (16), (17) and (18):

[0065] (15)

[0066] (16)

[0067] (17)

[0068] (18)

[0069] Where n is the dimension of the system input variable X(t), N co Let N be the dimension of the control vector U(t). c To evaluate the number of neurons in the hidden layers of a network, q k p k and p ck Let ω be the input value, intermediate state value, and output value of the k-th neuron in the hidden layer, respectively. c (1) ω cr (1)和ωcu(1) Let ω represent the weights of the system input variables X(t), R(t), and U(t) from the input layer to the hidden layer, respectively. c(2) The weights from the hidden layer to the output layer;

[0070] CNN network weight update calculation formula:

[0071] (19)

[0072] Where, η c (t) is the learning rate of the CNN network;

[0073] According to the inverse gradient descent rule, the formula for calculating the gradient from the hidden layer to the output layer can be obtained as follows:

[0074] (20)

[0075] The formula for calculating the gradient from the input layer to the hidden layer is as follows:

[0076] (twenty one)

[0077] (twenty two)

[0078] (twenty three)

[0079] in, .

[0080] Preferably, the ANN network adjusts the network weights to make the performance index function J(t) approximate the target expectation function U. c The value of (t) is given, and the error function of the execution network is defined as shown in equation (24):

[0081] (twenty four)

[0082] The input for executing the network is:

[0083] (25)

[0084] The inputs and outputs of neurons in the hidden and output layers of the ANN network are shown in equations (26), (27), (28), and (29).

[0085] (26)

[0086] (27)

[0087] (28)

[0088] (29)

[0089] Where N co The number of control signals output by the execution network is N, where N is the number of neurons in the hidden layer of the execution network, and h is the number of control signals output by the execution network. i g i and g ai Let ω be the input, intermediate state value, and output of the i-th neuron in the hidden layer, respectively. a (1) and ω a (2) These represent the weights from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. The ANN network weight update formula is shown in equation (30):

[0090] (30)

[0091] η a (t) is the learning rate of the ANN network.

[0092] The gradient calculation formulas from the hidden layer to the output layer and from the input layer to the hidden layer are shown in equations (31) and (32).

[0093] (31)

[0094] (32)

[0095] in,

[0096] Preferably, the RNN network defines an error function E. r The formula for calculating (t) is shown below:

[0097] (33)

[0098] (III) Beneficial Effects

[0099] Compared with the prior art, the present invention provides a wind turbine MPPT control method based on wind speed prediction and PID-DHDP, which has the following beneficial effects:

[0100] This wind turbine MPPT control method based on wind speed prediction and PID-DHDP predicts wind speed through a GRU network to obtain a wind speed prediction sequence. The wind speed prediction value and the optimal tip speed ratio can be used to easily calculate the wind turbine speed reference value. The deviation signal between the wind turbine speed reference value and the actual wind turbine speed value is used as the input signal of the PID-DHDP controller to form a closed-loop speed control system, which realizes the wind turbine speed is stabilized at the optimal speed and the wind turbine outputs maximum power, thereby achieving wind turbine MPPT control. Attached Figure Description

[0101] Figure 1 The wind turbine MPPT control flow diagram is a wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0102] Figure 2 The diagram shows the GRU network model of the wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0103] Figure 3 This invention presents the hyperparameter optimization process of the GRU using a wolf pack algorithm based on wind speed prediction and PID-DHDP-based wind turbine MPPT control method.

[0104] Figure 4 The diagram shows the PID-DHDP principle of the wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0105] Figure 5This is a schematic diagram of the CNN network structure for a wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0106] Figure 6 This is a schematic diagram of the ANN network structure for a wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0107] Figure 7 This is a schematic diagram of the RNN network structure of a wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0108] Figure 8 This is a schematic diagram of the optimal tip speed ratio control principle of the wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention;

[0109] Figure 9 This invention presents a wind turbine MPPT control method based on wind speed prediction and PID-DHDP, which uses curves to show the relationship between wind turbine power and tip speed ratio under different wind speeds.

[0110] Figure 10 This is a schematic diagram of a 4-unit-11-bus wind power generation system based on wind speed prediction and PID-DHDP-based MPPT control method proposed in this invention.

[0111] Figure 11 The diagram shows the branch electrical parameters of a 4-unit-11-node wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0112] Figure 12 The wind turbine speed response curve is shown in the figure for the wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention.

[0113] Figure 13 The wind energy utilization coefficient diagram of the wind turbine proposed in this invention is based on wind speed prediction and PID-DHDP MPPT control method.

[0114] Figure 14 The diagram shows the power output curve of a wind turbine according to the wind turbine MPPT control method based on wind speed prediction and PID-DHDP proposed in this invention. Detailed Implementation

[0115] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0116] A wind turbine MPPT control method based on wind speed prediction and PID-DHDP includes the following steps:

[0117] 1) Acquire real-time data such as wind speed, temperature, and humidity, and normalize this data;

[0118] 2) The Wolf Pack algorithm (WPA) is used to optimize the hyperparameters of the Gated Recurrent Unit (GRU).

[0119] 3) The normalized data is used as the input data for the GRU network, and the GRU network outputs the wind speed prediction sequence values;

[0120] 4) Calculate the reference value of the fan speed using the predicted wind speed and the optimal tip speed ratio. Use the deviation signal between the reference value and the actual fan speed as the input signal of the PID (Proportional-Integral-Derivative)-DHDP (Direct Heuristic Dynamic Programming) controller to form a closed-loop speed control system. This system stabilizes the fan speed at the optimal speed and maximizes the fan output power, thereby achieving MPPT control of the fan.

[0121] In this embodiment, the hidden layer of the GRU contains only reset gates and update gates. The reset gates and update gates are used to control the degree to which information is transferred. The input to both gates is the current input x. t and the hidden state h from the previous moment t-1 .

[0122] In this embodiment, the reset gate is responsible for determining how much memory information needs to be retained. The input to the reset gate is the current input 𝑥. 𝑡 The hidden layer state h at the previous time step t−1 It is composed of concatenation, and its output is a number in the interval (0, 1), representing the state h of the hidden layer at the previous time step. t−1 The proportion to be retained, and the output formula for resetting the door are:

[0123] (1)

[0124] After discarding and remembering the input information through two gate structures, GRU calculates the candidate hidden state values. Its calculation expression is shown in the following formula:

[0125] (2)

[0126] in The tanh activation function is used. h represents the input at time t. t-1 The vector represents the hidden state at the previous time step, and [ ] indicates that the two vectors are connected. h w r Let represent the weight matrix, and σ() represent the sigmoid function.

[0127] In this embodiment, the input to the update gate is changed from the input at the current moment. The hidden layer state h at the previous time step t-1 It is constructed by concatenation, and its output is a number in the interval (0, 1). The output formula of the update gate is shown in formula (3). ：

[0128] (3)

[0129] After obtaining the updated state information through the update gate, the tanh activation function creates a vector of all possible values ​​based on the new input and calculates the candidate hidden state values. Then, the final state h at the current moment is calculated through the network. t The calculation formula is shown in equation (4):

[0130] (4)

[0131] According to the calculation formula above, GRU stores and filters information through two gates, retains important features through the gate function, and captures dependencies through learning, thereby obtaining the optimal output value.

[0132] In this embodiment, the detailed process of the wolf pack algorithm includes the following steps:

[0133] 1) Wolf Pack Initialization

[0134] The artificial wolf initializes its position in the solution space using random values. Assuming the iteration number k=0, X... i Let X represent the i-th artificial wolf. min X represents the lower bound of the variable's value space. max This represents the upper bound of the variable's value space.

[0135] (5)

[0136] Where rand∈(0,1) represents a random number between 0 and 1, assuming it exists in a D-dimensional space (D is the dimension of the space of the variable to be optimized), then X i It can be represented as:

[0137] (6)

[0138] 2) Rules for the emergence of the alpha wolf

[0139] The prey odor concentration at the location of each artificial wolf is calculated, and the artificial wolf with the highest prey odor concentration is selected as the alpha wolf. After each iteration, the wolf pack compares the best artificial wolf after the iteration with the alpha wolf and selects the wolf with the higher odor concentration as the alpha wolf.

[0140] 3) Wandering behavior

[0141] Within the solution space, select the T with the best fitness, excluding the alpha wolf. n One artificial wolf acts as a scout, responsible for searching for prey, among which T n The value is a random integer between [S / (α+1), S / α], where S is the total number of artificial wolves in the wolf pack, and α is the proportion of scout wolves among the artificial wolves. During the prey search process, the scout wolves first calculate the prey odor concentration Y at the location of each scout wolf. i If the odor concentration Y i >Y leader Then we can obtain Y leader =Y i That is, the scout wolf i takes the place of the alpha wolf and initiates a summoning action to all the wolves; if Y i <Y leader The wolf will move in the direction of h with a fixed step size s. a Swim forward one step, and take the current position of Detective Wolf i as the starting point. Then Detective Wolf i will move along... The position will be updated after the direction is moved:

[0142] (7)

[0143] At this moment, the scent concentration of the prey perceived by the wolf i is Y. ip After making a decision, the wolf detects the direction along the path with the highest odor concentration (Y) that is greater than the current odor concentration. i Move forward one step in the direction, while simultaneously adjusting its position X i Update. Repeat the above steps until the wolf i detects the concentration Y of the prey's scent. i >Y leader Or, the number of times the wolf roams (T) reaches the maximum number of roams (T). max ;

[0144] 4) Summoning behavior

[0145] After receiving information about prey from the scout wolves, the alpha wolf in the pack howls to summon the surrounding wolves. In the algorithm, the wolves with the lowest fitness scores are chosen as the alpha wolves, and their number is M. n =ST n -1. After receiving the summoning command, the wolf runs towards the prey's location with a large stride. Therefore, after k1 iterations, the position of wolf j in d-dimensional space can be represented as:

[0146] (8)

[0147] in, This represents the position of the alpha wolf in the d-th dimension after the k-th iteration. This indicates the current spatial location of the wolf j. This indicates that as the wolf gradually approaches the alpha wolf, during a pack attack, if the concentration of the prey's scent perceived by wolf j is greater than that perceived by the alpha wolf, that is, Y... j >Y leader Wolf j will take the position of the alpha wolf; otherwise, wolf j will continue its running maneuver until it reaches a spatial distance d from the alpha wolf's location. jm Less than the distance criterion d near At this point, the wolf begins to switch from a running attack to a surrounding attack, and the distance determination factor is calculated using the following formula (9):

[0148] (9)

[0149] Here, ω is the distance decision factor, and its magnitude determines the convergence speed of the wolf pack algorithm. Under the same conditions, the larger the value of the decision factor, the faster the convergence speed of the wolf pack algorithm. However, if ω is too large, it will affect the wolves' siege behavior, causing them to fail to enter the siege state. d and max d Let represent the minimum and maximum values ​​in the d-th dimension space to be optimized, respectively;

[0150] 5) Besieging behavior

[0151] In a siege, the position of the alpha wolf closest to the prey is represented as the prey's position. For the k-th generation of artificial wolves, the prey's position in the d-th dimension can be represented as... Then, during the kth iteration, the position of the wolf pack after the siege can be represented as shown in equation (10):

[0152] (10)

[0153] Where λ is a random constant between [-1, 1], Let be the stride length of a wolf when it launches an attack on its prey. When wolf i is attacking, if the concentration of the prey's scent it senses is greater than the scent concentration at its original location, then wolf i will update its position. Otherwise, the position of wolf i will not change. There is a certain relationship between the stride lengths of the three intelligent behaviors in d-dimensional space, as shown in the following formula (11):

[0154] (11)

[0155] Where C represents the step size factor of intelligent behavior, the magnitude of which affects the level of refinement in searching for prey;

[0156] 6) Wolf Pack Renewal Rules

[0157] The wolf pack is updated according to the law of "survival of the fittest". In the wolf pack, the relatively weak R wolves will be eliminated due to the competition mechanism. At the same time, R artificial wolves will be randomly generated. The value of R is in the range of [n / 2β, n / β], where β is the wolf pack update ratio factor.

[0158] It's important to note that each GRU unit learns how to acquire dependency information at different time scales. Units with higher update port values ​​will acquire longer dependency information, while units with higher reset port values ​​will acquire shorter dependency information. This leads to drawbacks such as high parameter requirements and large memory demands for large-scale data computation. Therefore, to address the optimization issues of large-scale data and model parameters, appropriate optimization algorithms are needed to optimize the model, enabling automatic learning and updating of model parameters.

[0159] Theoretically, the complexity of a network increases with the number of hidden layers and neurons within them, and the prediction accuracy of the neural network model also improves. However, when building a model, an excessive number of hidden layers and neurons increases the model's complexity, easily leading to overfitting and thus reducing prediction accuracy. At the same time, overly simple networks are prone to poor training results due to insufficient data, failing to meet expected requirements.

[0160] The parameters of the hidden layer, learning rate, time step, and maximum number of iterations have a great impact on the training effect of the GRU neural network. Therefore, the wolf pack algorithm is used to optimize these hyperparameters.

[0161] In this embodiment, PID-DHDP consists of three networks: an Actor network (ANN), a Critic network (CNN), and a Reinforcement network (RNN).

[0162] The ANN network module receives the input variable X(t) and outputs the control variable U(t). The control variable U(t) is returned to the controlled system to control the controlled system.

[0163] The RNN network module receives the input variable X(t) of the controlled system and the control variable U(t) of the network module, and the RNN network module obtains the internal reinforcement learning function R(t).

[0164] The CNN network module receives the input variable X(t) of the controlled system, the control variable U(t) of the ANN network module, and the internal reinforcement learning function R(t) of the RNN network module. The CNN network module then obtains the performance index function J(t).

[0165] The first time-series differential signal is obtained by combining the performance index function J(t) with the target expectation function Uc(t). , ;

[0166] The performance index function J(t) is multiplied by multiplier α and combined with the performance index function J(t-1) and the external reinforcement learning function r(t) to obtain the second time-series difference signal. , ;

[0167] The performance index function J(t) is combined with the performance index function J(t-1) and the internal reinforcement learning function R(t) through multiplier α to obtain the third time-series difference signal. , ;

[0168] The ANN, RNN, and CNN networks update their respective weight coefficients online using the first, second, and third time-series difference signals, respectively.

[0169] In this embodiment, the input X of the CNN network c (t) is:

[0170] (12)

[0171] The error function of the CNN network is defined as shown in equation (13):

[0172] (13)

[0173] Where λ is the discount factor, 0 < λ < 1,

[0174] The transfer function of the hidden layer neurons in the CNN network adopts the bipolar sigmoid function, as shown in equation (14) below:

[0175] (14)

[0176] The output is the performance index function J(t). Its hidden layer uses the sigmoid activation function, while the output layer uses the linear activation function. It is easy to see that the input and output of the hidden and output neurons of the CNN network are shown in equations (15), (16), (17) and (18):

[0177] (15)

[0178] (16)

[0179] (17)

[0180] (18)

[0181] Where n is the dimension of the system input variable X(t), N co Let N be the dimension of the control vector U(t). c To evaluate the number of neurons in the hidden layers of a network, q k p k and p ck Let ω be the input value, intermediate state value, and output value of the k-th neuron in the hidden layer, respectively. c (1) ω cr (1)和ωcu(1) Let ω represent the weights of the system input variables X(t), R(t), and U(t) from the input layer to the hidden layer, respectively. c(2) The weights from the hidden layer to the output layer;

[0182] CNN network weight update calculation formula:

[0183] (19)

[0184] Where, η c (t) is the learning rate of the CNN network;

[0185] According to the inverse gradient descent rule, the formula for calculating the gradient from the hidden layer to the output layer can be obtained as follows:

[0186] (20)

[0187] The formula for calculating the gradient from the input layer to the hidden layer is as follows:

[0188] (twenty one)

[0189] (twenty two)

[0190] (twenty three)

[0191] in, .

[0192] In this embodiment, the ANN network adjusts the network weights to make the performance index function J(t) approximate the value of the target expectation function Uc(t). The error function of the execution network is defined as shown in the following equation (24):

[0193] (twenty four)

[0194] The input for executing the network is:

[0195] (25)

[0196] The hidden layer contains three types of nodes: proportional (P) nodes, integral (I) nodes, and differential (D) nodes. Figure 6 This is a schematic diagram of the multilayer neuron structure of an ANN network. From... Figure 6 The inputs and outputs of neurons in the hidden and output layers of the ANN network are shown in equations (26), (27), (28), and (29): (26)

[0197] (27)

[0198] (28)

[0199] (29)

[0200] Where N co The number of control signals output by the execution network is N, where N is the number of neurons in the hidden layer of the execution network, and h is the number of control signals output by the execution network. i g i and g ai Let ω be the input, intermediate state value, and output of the i-th neuron in the hidden layer, respectively. a (1) and ω a (2) These represent the weights from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. The ANN network weight update formula is shown in equation (30):

[0201] (30)

[0202] η a (t) is the learning rate of the ANN network.

[0203] The gradient calculation formulas from the hidden layer to the output layer and from the input layer to the hidden layer are shown in equations (31) and (32):

[0204] (31)

[0205] (32)

[0206] in, .

[0207] In this embodiment, the RNN network defines an error function E. r The formula for calculating (t) is shown below:

[0208] (33)

[0209] Where r(t) is the external reinforcement learning function, defined as: (34)

[0210] Where ε is a constant greater than 0.

[0211] The multi-layered neuron structure of an RNN network is as follows: Figure 7 As shown. Its input vector is x. r =[x1, x2, ..., x n The output is the internal PID-DHDP function R(t). The transfer function of the hidden and output layer neurons in the RNN network is the bipolar sigmoid function.

[0212] The input and output expressions for neurons in the output layer and hidden layer of an RNN network are as follows:

[0213] (35)

[0214] (36)

[0215] (37)

[0216] (38)

[0217] Where N r z is the number of neurons in the hidden layer. v y v and y fv Let ω be the input, intermediate state value, and output of the i-th neuron in the hidden layer, respectively. fv (1) and ω fv (2) These represent the weights from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. Gradient descent is used to optimize the error function E in equation (39). r(t) is minimized, and equation (39) is the formula for updating the weights of the RNN:

[0218] (39)

[0219] Where η g (t) represents the learning rate of the RNN network. According to the back propagation (BP) rule, the gradient from the hidden layer to the output layer is calculated as shown in equation (40).

[0220] (40)

[0221] The gradient from the input layer to the hidden layer is calculated as follows:

[0222] (41)

[0223] in, .

[0224] It should be noted that, according to aerodynamics, the characteristics of wind turbines in capturing wind energy can be represented by the following simplified expression of the wind turbine's output mechanical power:

[0225] (42)

[0226] in, The mechanical power output (W) of the wind turbine. Where is the blade radius (m). air density (kg / m³) 3 ), Equivalent wind speed (m / s);

[0227] The wind energy utilization coefficient, and and related, The ratio of leaf tip speed, The blade pitch angle is (°).

[0228] ① Tip speed ratio

[0229] Defined as the ratio of the linear velocity at the tip of a wind turbine blade to the wind speed:

[0230] (43)

[0231] in, This refers to the mechanical rotational speed of the wind turbine.

[0232] (44)

[0233] in, denoted as per-unit value (pu) for generator mechanical speed, p is the number of generator pole pairs, and GR is the gear ratio.

[0234] ② Blade pitch angle

[0235] The pitch angle refers to the angle between the wind turbine blades and the rotor plane. The smaller the pitch angle, the larger the windward surface of the blades, and therefore the greater the wind energy captured.

[0236] Considering the mechanical efficiency of the wind turbine gearbox, the actual output power of the wind turbine is:

[0237] (45)

[0238] in, For gearbox efficiency.

[0239] A wind power generation system mainly includes a pitch controller, a pitch servo system, blades, a generator, and sensors. The inputs are wind speed and the optimal tip speed ratio, the disturbance comes from changes in wind speed, and the output is the wind turbine rotational speed. The entire system control is shown in Figure 8. A PID-DHDP controller is used to control the wind turbine rotational speed ω. m Follow the reference rotational speed ω ref This allows the wind turbine to achieve the optimal tip speed ratio and the maximum wind energy conversion coefficient, thus realizing the purpose of the wind turbine MPPT.

[0240] The wind energy utilization factor is defined as:

[0241] (46)

[0242] in,

[0243] (47)

[0244] Among them, C p It is negatively correlated with β. When β=0, the wind energy utilization coefficient C p At its maximum, MPPT control of the fan can be achieved. Therefore, setting β=0, we can obtain from equation (47):

[0245] (48)

[0246] in,

[0247] (49)

[0248] The relationship between wind turbine power generation and speed at different wind speeds can be obtained from equations (48) and (49). The optimal wind energy utilization coefficient C is used. pmaxSubstituting the value of the optimal tip speed ratio into equation (48), we obtain the relationship curve between electric power and rotational speed at this time, which is the optimal electric power curve.

[0249] Experimental Example: Simulation Results of MPPT Control for Wind Turbines Based on PID-DHDP

[0250] The principle of a 4-unit-11-node power generation system including wind power is as follows: Figure 10 As shown, the 4-generator-11-node power generation system includes 4 generators, with generator No. 1 being an equivalent generator, 11 nodes, 4 transformers, and 12 lines. The system's base power is 100 MVA, and the base voltage is 345 kV.

[0251] Electrical parameters of the branch circuits of the 4-unit-11-node generator system are as follows: Figure 11 As shown, a transformer ratio of 0 indicates that the branch is a power transmission line, while a ratio other than 0 indicates a transformer branch.

[0252] At 0.5s, the wind speed increased from 2m / s to 4m / s. For comparison, PID control and PID-DHDP control were used respectively. Figure 12 It is the result of the fan speed control. Figure 13 It is the wind energy utilization coefficient of the wind turbine. Figure 14 This is the power output curve of the wind turbine.

[0253] The beneficial effects of this invention are as follows:

[0254] By adjusting the rotor speed of wind turbines, the tip speed ratio of the turbine blades can be maintained near the theoretically optimal tip speed ratio, achieving maximum power point tracking (MPPT). This allows for full utilization of wind power and improves the economic and social benefits for wind farm companies. Given the random, intermittent, and fluctuating nature of wind speed, the core issue is how to enhance the wind turbine's ability to respond to rapid changes in wind speed and maintain the turbine in MPPT mode as much as possible.

[0255] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind turbine MPPT control method based on wind speed prediction and PID-DHDP, characterized in that, Includes the following steps: 1) Acquire wind speed, temperature, and humidity data in real time, and normalize these data. 2) The Wolf Pack Algorithm (WPA) is used to optimize the hyperparameters of the gated recurrent network (GRU). 3) The normalized data is used as the input data for the GRU network, and the GRU network outputs the wind speed prediction sequence values; 4) The reference value of the fan speed is calculated by the wind speed prediction value and the optimal tip speed ratio. The deviation signal between the reference value of the fan speed and the actual fan speed is used as the controller input signal of PID-DHDP direct heuristic dynamic programming to form a closed-loop speed control system, so as to realize the fan speed is stabilized at the optimal speed and the fan outputs the maximum power, thereby realizing the MPPT control of the fan. The hidden layer of the GRU contains only reset gates and update gates, which control the degree to which information is transferred. The input to both gates is the current input x. t and the hidden state h from the previous moment t-1 ; The reset gate is responsible for determining how much memory information needs to be retained. The input to the reset gate is the current input x. t The hidden layer state h at the previous time step t-1 It is composed of concatenation, and its output is a number in the interval (0, 1), representing the state h of the hidden layer at the previous time step. t-1 The proportion to be retained, and the output formula for resetting the door are: （1） After discarding and remembering the input information through two gate structures, GRU calculates the candidate hidden state values. Its calculation expression is shown in equation (2) below: (2) Where f is the tanh activation function, x t h represents the input at time t. t-1 The vector represents the hidden state at the previous time step, and [ ] indicates that the two vectors are connected. h w r Represents the weight matrix. Represents the sigmoid function; The input to the update gate is the input x at the current time. t The hidden layer state h at the previous time step t-1 It is constructed by concatenation, and its output is a number in the interval (0, 1). The output formula of the update gate is shown in formula (3): (3) After obtaining the updated state information through the update gate, the tanh activation function creates a vector of all possible values ​​based on the new input and calculates the candidate hidden state values. Then, the final state h at the current moment is calculated through the network. t The calculation formula is shown in equation (4): (4) According to the calculation formula above, GRU stores and filters information through two gates, retains important features through the gate function, and captures dependencies through learning, thereby obtaining the optimal output value.

Images