A Stepwise Inertia Control Method for Wind Power Frequency Regulation Based on Deep Learning

Abstract

The present invention discloses a wind power frequency modulation step-by-step inertia control method based on deep learning. This method mainly includes the following three steps: 1) Obtain the optimal wind power frequency modulation step-by-step inertia control parameters using the Golden Eagle optimization algorithm in time-domain simulation and generate a data set; 2) Extract features from the data set based on a stacked denoising autoencoder; 3) Generate an optimal wind power frequency modulation step-by-step inertia control scheme by learning data features based on a deep neural network. The present invention can quickly and accurately provide the optimal wind power frequency modulation step-by-step inertia control scheme under the current system when power imbalance occurs in the power system; it can prompt the wind farm to actively provide power support for the power system, effectively participate in the power system frequency control, timely suppress the system frequency drop, is beneficial to compensating for the system inertia, effectively improves the quality and efficiency of wind power frequency modulation, provides an important scheme for power system planning and decision-making work, and has great significance for the safety and stability of power system operation.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power frequency modulation, and in particular to a stepwise inertia control method for wind power frequency modulation based on deep learning. Background Art

[0002] In order to address various issues such as climate, environment, and energy, it has become a trend that the proportion of renewable energy in the power system is increasing continuously. As an important renewable energy source, wind power has achieved stable and continuous development. While wind power provides clean energy, it also causes a decrease in the inertia of the power system, weakening the system's ability to cope with active power imbalance, and posing severe challenges to the quality of system frequency control and frequency stability. Wind turbines generally adopt maximum power point tracking control and are connected to the grid through power electronic converters, decoupled from the system frequency, unable to respond to power deviations by releasing or absorbing energy, without inertia response characteristics, and unable to actively provide inertia support for the grid under active power disturbances. To improve the quality and efficiency of system frequency control, it is necessary to involve wind turbines in this control and compensate for the reduction of the overall system inertia in this way. Recently, new grid codes also require wind farms to contribute to power system frequency control. To meet the requirements, when there is an active power imbalance in the grid, stepwise inertia control can be used to participate in grid frequency control, which can quickly provide short-term power support.

[0003] During the frequency modulation process using stepwise inertia control, the wind turbine provides additional power to the grid for a certain period by releasing the rotor kinetic energy, and the rotor speed also decreases accordingly. If the rotor speed of the wind turbine drops to its defined lower limit, the wind turbine will disconnect from the power system. Therefore, the wind turbine can only provide a short-term power support to the grid by releasing limited rotor kinetic energy. Stepwise inertia control must also ensure that the wind turbine rotor terminates power over-issuance before reaching the lowest speed. However, suddenly terminating power over-issuance will cause a new drop in the grid frequency, also known as secondary frequency drop (SFD). Under stepwise inertia control, the quality of frequency modulation is complexly related to the magnitude of secondary frequency drop. To mitigate secondary frequency drop while ensuring good frequency modulation effects, different methods can be adopted.

[0004] At present, some people propose to design the wind power frequency modulation strategy as a piecewise function and use the particle swarm algorithm to obtain the piecewise function values in different wind speed intervals, which can effectively improve the SFD phenomenon. However, the use of the piecewise function also causes the grid frequency to drop multiple times, resulting in more frequency fluctuations and being unfavorable for the rapid recovery of the frequency. Some people also propose to use a high-voltage DC link to compensate for the power reduction at the terminal of the wind turbine. Due to the fast response characteristics of the high-voltage DC link, the power attenuation compensation technology shows good performance in alleviating the secondary frequency drop. However, this method requires a large amount of investment. The current improved method under stepwise inertial control is difficult to quickly, efficiently, and economically provide corresponding frequency control schemes and give full play to the full strength of stepwise inertial control when facing different load disturbance events; it is difficult to meet the requirements for excellent stepwise inertial control during wind power frequency modulation. Summary of the Invention

[0005] The present invention provides a stepwise inertial control method for wind power frequency modulation based on deep learning, which is characterized by including the following three steps:

[0006] Step 1: Use the golden eagle optimization algorithm in time-domain simulation to obtain the optimal stepwise inertial control parameters for wind power frequency modulation and generate a data set;

[0007] Step 2: Extract features from the data set based on the stacked denoising autoencoder;

[0008] Step 3: Generate an optimal stepwise inertial control scheme for wind power frequency modulation by learning data features based on a deep neural network.

[0009] Further, the optimal stepwise inertial control parameters in Step 1 are used when the stepwise inertial control strategy is used in wind power frequency modulation.

[0010] Stepwise inertial control (SIC) is a control strategy for wind power to participate in the frequency adjustment of the power system, including two main stages: a short-term over-generation stage and a rotational speed recovery stage. Figure 2 、 Figure 3 respectively represent the relationship between the rotor speed and the output power of the fan and the change of the output power during the stepwise inertial control process when the stepwise inertial control strategy is used under the conditions of a fixed wind power ratio and a constant wind speed. The specific process of participating in frequency modulation is as follows:

[0011] 1) The fan normally operates at point A on the MPPT curve, and the output electromagnetic power is , the rotor speed is , and the output power is . TAt time 0, a sudden power imbalance occurs in the system, resulting in a frequency drop. The wind turbine detects that the frequency drop exceeds the dead - band limit and then switches to the step - by - step inertia control mode, immediately increasing the output power. , the output power , the output power is increased from point A to point B, entering the short - term over - generation stage.

[0012] 2) During the short - term over - generation stage, the output power of the wind turbine remains unchanged for a period of time, that is, the output power is always , lasting from point B to point C. At this stage, the electromagnetic power output by the wind turbine is greater than the mechanical power, and the rotor then decelerates. At this time, the rotor motion equation of the wind turbine is:

[0013] (1)

[0014] In the formula: refers to the inertia constant of the wind turbine.

[0015] It should be noted that the short - term over - generation stage must be terminated before the rotor over - decelerates to avoid the rotor stall event of the wind turbine. Therefore, the duration of the short - term over - generation stage should be selected to ensure that the rotor speed does not reach the minimum speed at time , which is generally 0.7 p.u.

[0016] 3) At time , the short - term over - generation stage ends, and the output power of the wind turbine suddenly drops . At this time , the rotor speed is

[0017] 4) The wind turbine enters the speed recovery stage. At this time, the wind turbine operates in the maximum power point tracking mode, and the output power of the wind turbine is . The system power deficit generated is: . At this stage, the mechanical power of the wind turbine is greater than the electromagnetic power, and the rotor starts to accelerate. After the recovery time, at time, the rotor speed recovers from to . The electromagnetic power output by the wind turbine will also recover to as the rotor speed increases, and the output power slowly returns from point D to point A along the MPPT trajectory.

[0018] Furthermore, the optimal wind power frequency modulation step - by - step inertia control parameters obtained in step 1 are obtained using the Golden Eagle optimization algorithm in time - domain simulation.

[0019] The Golden Eagle Optimizer (GEO) is inspired by the wisdom of golden eagles and has advantages such as strong optimization ability and fast convergence speed. When golden eagles hunt, they adjust their speeds at different hovering trajectory stages. In the initial stage of hunting, they show more tendency to cruise and search for prey, and in the later stage, they show more tendency to attack. Golden eagles adjust these two parts to capture the best prey within the shortest time and feasible area. Such behaviors are written as a mathematical model and developed by methods that highlight exploration and global optimization.

[0020] Hovering action: The Golden Eagle Optimizer is based on the hovering motion of golden eagles. Each golden eagle remembers the best position it has reached so far. Golden eagles want to attack prey and cruise to find better food. In each iteration, each golden eagle randomly selects the prey of another golden eagle and hovers at the best position it has visited so far. Golden eagles can also choose to circle their own memories.

[0021] Prey selection: In each iteration, each golden eagle must select a prey to perform cruising and attacking operations. In the Golden Eagle Optimizer, the prey is modeled as the best solution found by the golden eagle group so far. Each golden eagle can remember the best solution found so far. In each iteration, each search agent selects the target prey from the memory of the entire group. Then, the attack and cruise vectors of each golden eagle for the selected prey are calculated. If the new position calculated by the attack and cruise vectors is better than the previous position in the memory, the memory is updated. The prey selection strategy plays an important role in the Golden Eagle Optimizer. The selection can be carried out in a basic way, where each golden eagle only selects prey from its own memory. To enable golden eagles to better explore the land, a random one-to-one mapping scheme is proposed, where each golden eagle randomly selects the prey in the current iteration from the memories of other group members. It should be noted that the selected prey is not necessarily the nearest or farthest prey. In this scheme, each prey in the memory is assigned or mapped to a unique golden eagle. Then each golden eagle attacks and cruises the selected prey.

[0022] Attack behavior: The attack behavior can be simulated by a vector that starts from the current position of the golden eagle and ends at the position of the prey in the golden eagle's memory. The attack vector of the golden eagle can be calculated by Equation (2):

[0023] (2)

[0024] In the formula: is the attack vector of the i-th golden eagle; is the golden eagle f the best location (prey) reached so far; is the current position of the i-th golden eagle.

[0025] Cruise (exploration) behavior: The cruise vector is calculated based on the attack vector. The cruise vector is the tangent vector of the circle and perpendicular to the attack vector. Cruise can also be considered as the linear velocity of the golden eagle relative to the prey. The n-dimensional cruise vector lies in the hyperplane tangent to the circle. Therefore, in order to calculate the cruise vector, the equation of the tangent hyperplane must be calculated first. The equation of the n-dimensional hyperplane can be determined by any point on the hyperplane and a vector perpendicular to the hyperplane, and this vector is called the normal vector of the hyperplane. The scalar form of the hyperplane equation in n dimensional space is shown in Equation (3):

[0026] (3)

[0027] In the formula: is the normal vector; is the variable vector; is any point on the hyperplane; . Regarding (attack vector) as the normal of the hyperplane, then (the cruise vector of the golden eagle in the iteration) can represent the hyperplane to which it belongs according to Equation (4).

[0028] (4)

[0029] In the formula: is the attack vector; is the decision / design variable vector; is the position of the selected prey.

[0030] In order to find a random vector on the cruise hyperplane, we must first find a random destination point C point on this hyperplane, rather than the point we already have (the current position of the golden eagle). Note that the starting point of its cruise vector is the current position of the golden eagle. Since the hyperplane is one dimension smaller than their environmental space, a simple random 1×n point cannot be generated. A simple random point in n-dimensional space cannot guarantee to be located on the cruise hyperplane. A new point located on the n-dimensional cruise hyperplane has n-1 degrees of freedom, which means that n-1 dimensions can be freely selected, but the hyperplane equation determines the last 1 dimension, as shown in Equation (3). The last dimension must be selected to satisfy the hyperplane equation. Therefore, there are n-1 free variables and 1 fixed variable. Use the following steps to find a random n dimensional destination point C on the cruise hyperplane of the golden eagle, and the specific steps are as follows:

[0031] Step 1) Randomly select one variable from n variables as the fixed variable;

[0032] Step 2) Assign random values to all variables except the kth variable since the kth variable is fixed.

[0033] Step 3) Use equation (5) to find the value of the fixed variable.

[0034] (5)

[0035] Where: Target point C The kth element of ; Attack vector The jth element of ; k is the number of the fixed variable. Find a random target point on the cruise hyperplane. The general representation of the target point on the cruise hyperplane is shown in formula (6):

[0036] (6)

[0037] Transfer to a new position: The displacement of the Golden Eagle is composed of the attack and vectors. We define the step vector of the Golden Eagle iteration as Equations (7) and (8).

[0038] (7)

[0039] (8)

[0040] Where: is the attack coefficient; is the cruise coefficient; 、 is a random vector in [0,1]. Therefore, the position of the golden eagle at iteration t+1 can be calculated by adding the step vector at iteration t to the position at iteration t:

[0041] (9)

[0042] Where: is the t+1th position of the Golden Eagle; is the t-th position of the Golden Eagle; The step size of the golden eagle's movement.

[0043] If the adaptability of the eagle's new position is better than the position in its memory, the eagle's memory will be updated with the new position. Otherwise, the memory remains intact, but the eagle will be in the new position. In the new iteration, each eagle randomly selects a eagle from the population, rotates around its best visited position, calculates the attack vector, the cruise vector, and finally calculates the step vector and the new position for the next iteration. This loop will continue until any termination condition is met. There are two coefficients in Equation (7), namely the attack coefficient and the cruise coefficient, which are used to control the influence of the attack vector and the cruise vector on the step vector, as shown in Equation (10):

[0044] (10)

[0045] Where: t represents the current iteration; T represents the maximum iteration; and is the attack tendency the initial value and the termination value; and is the cruise tendency the initial value and the termination value:

[0046] Furthermore, the basic steps of obtaining the optimal wind power frequency modulation step-by-step inertia control parameters and generating a data set using the golden eagle optimization algorithm in the time domain simulation in step 1 are as follows:

[0047] Step 1) Initialize the number of golden eagles;

[0048] Step 2) Calculate the fitness function and initialize the population memory;

[0049] Step 3) Initialize 、 ;

[0050] Step 4) Calculate formula (10) and update 、 ;

[0051] Step 5) Calculate formula (2), and randomly select prey from the attack vectors calculated from the memory of the population;

[0052] Step 6) Calculate the cruise vector (formulas (3)-(6)), calculate the step size vector (formulas (7)-(9)), update the position (formula (9)), and evaluate the fitness function of the new position;

[0053] Step 7) Update the optimal solution and the optimal position;

[0054] Step 8) Determine whether the maximum number of iterations is satisfied. If so, output the optimal golden eagle position and the global optimal solution. Otherwise, return to step 4 to recalculate iteratively;

[0055] Furthermore, the feature extraction of the data set in step 2 is performed by a stacked denoising autoencoder.

[0056] An autoencoder (AE) is as Figure 4 shown. An autoencoder consists of an encoder and a decoder. The encoder and the decoder have hidden layers: the input is converted into a latent representation in the hidden layer through the encoder, and then the internal representation is converted into an output through the decoder. The output is equivalent to the reconstruction of the input and should be as close to the input as possible.

[0057] Select the input sample set , which consists of N groups of samples , where N is the number of sample groups and n is the number of samples in each group. Let the set of hidden layer feature vectors be , which consists of N groups of feature vectors , where m is the number of vectors in each group of feature vectors. Then the encoding relationship between and is

[0058] (11)

[0059] In the formula: is the weight matrix between the input layer and the hidden layer; is the bias matrix between the input layer and the hidden layer; is the neuron activation function of the encoder, usually the sigmoid function, which has good feature recognition ability.

[0060] (12)

[0061] In the formula: is the input vector.

[0062] The decoder is the inverse operation of the encoder, using the feature vectors of the hidden layer as the input vector. Let be the set of output vectors, with a total of N groups and a dimension of n. Then the expression of the decoder is

[0063] (13)

[0064] In the formula: is the weight matrix between the hidden layer and the output layer; is the bias matrix between the hidden layer and the output layer; is the neuron activation function of the decoder.

[0065] The autoencoder achieves the purpose of feature learning by minimizing the reconstruction error between the output vector and the input vector, and continuously adjusts the network weights and biases using the gradient descent algorithm to reduce the reconstruction error. However, the learning of the autoencoder may simply retain the original input data information and cannot ensure obtaining an effective feature information expression.

[0066] The denoising autoencoder (DAE) is an improvement of the autoencoder, and its architecture is as shown in Figure 5As shown in Figure 2, this approach builds on the autoencoder by adding noise to a portion of the input and training it to recover the original noise-free input. Using a denoising autoencoder allows the hidden layer to learn more effective data representations and uncover deeper hidden information.

[0067] The stacked denoising autoencoder (SDAE) consists of multiple DAEs, such as Figure 6 shown. W n is the weight matrix between the n-1th hidden layer and the nth hidden layer, B n is the bias matrix between the n-1th hidden layer and the nth hidden layer. Training is accomplished through two steps: unsupervised greedy layer-by-layer pre-training and supervised refinement training, achieving feature learning. The input to the first DAE layer is the raw data, and the output data of the hidden layer serves as the input to the upper DAE layer. Noise is added to the input data before training each layer. Through training, the stacked denoising autoencoder can efficiently extract high-order features from the data, better fit complex functions, accelerate the acquisition of network parameters, and enhance the deep feature learning capabilities of neural networks.

[0068] Furthermore, the basic steps of extracting features from the dataset based on the stacked denoising autoencoder in step 2 are as follows:

[0069] Step 1) normalize the data set obtained using the Golden Eagle optimization algorithm in the time domain simulation;

[0070] Step 2) Divide the training set and validation set, using 80% of the original data as the training set and the remaining 20% as the test data to generate training and test data;

[0071] Step 3) Build a stacked denoising autoencoder model and select the number of hidden layers and neurons;

[0072] Step 4) Greedy unsupervised pre-training layer by layer uses the noised data as input and uses the dropout technique to randomly select some neurons to temporarily stop working.

[0073] Step 5) Supervised refinement training is performed to fine-tune the weights and biases of the network until the number of iterations reaches the set value.

[0074] Furthermore, the generation of the optimal wind power frequency regulation step-by-step inertia control scheme based on deep neural network learning data features in step 3 is performed on the basis of the deep neural network.

[0075] A deep neural network (DNN) is a neural network with multiple hidden layers and is also known as a deep feedforward neural network (DFN). Divided by the position of different layers, the neural network layers inside the DNN can be classified into: an input layer, hidden layers, and an output layer. Generally, the first layer is the input layer, the last layer is the output layer, and the middle layers are all hidden layers. The layers are fully connected, that is, any neuron in the i-th layer must be connected to any neuron in the (i + 1)-th layer. The basic structure is as Figure 7 shown. From a local model perspective, a complex DNN, like a perceptron, is a linear relationship , plus an activation function .

[0076] Furthermore, the basic steps for generating an optimal wind power frequency modulation step-by-step inertia control scheme based on learning data features by a deep neural network in step 3 are as follows:

[0077] Step 1) Use the features extracted by the stacked denoising autoencoder as the input of the deep neural network;

[0078] Step 2) Divide the training set and the validation set, use 80% of the original data as the training set, and the remaining 20% as the test data to generate training and test data;

[0079] Step 3) Build a deep neural network model and select the number of hidden layers and the number of neurons in each hidden layer;

[0080] Step 4) Use Adam to optimize the input weight matrix, hidden layer feature vectors, and output weight matrix of the network, and selectively add regularization to the hidden layer;

[0081] Step 5) Use MSE as the evaluation index to judge whether it no longer decreases. If it is satisfied, the training is completed; otherwise, return to step 3;

[0082] Step 6) Save the trained model. When a load disturbance event is encountered next time, directly use the model to generate an optimal wind power frequency modulation step-by-step inertia control scheme.

[0083] Compared with the prior art by adopting the above technical solutions, the present invention has the following advantages:

[0084] 1. When performing wind power frequency modulation step-by-step inertia control, compared with the wind power frequency modulation strategy with a single consideration factor, the method used in this patent comprehensively considers the influence of different wind speeds, wind power ratios, and load disturbance amounts on the system frequency characteristics. The reference factors are more comprehensive, and it can efficiently obtain the optimal step-by-step inertia control parameters in multiple scenarios, meeting the usage requirements of wind power frequency modulation in different scenarios of the power system and having good generalization ability.

[0085] 2. The method used in this patent obtains the optimal step-by-step inertia parameters through feature extraction of the stacked denoising autoencoder and deep neural network learning, which has extremely high accuracy and meets the requirements of rapid frequency control during the actual operation of wind power frequency modulation. Compared with the time-domain simulation method, the method used in this patent, based on the advantage of fast computing speed of deep learning, greatly saves the frequency control time for online decision-making and improves the efficiency of online frequency control.

[0086] 3. Compared with other step-by-step inertia control methods, the method used in this patent can obtain the optimal step-by-step inertia control parameters with extremely high precision through the above technical solutions, resulting in a better frequency modulation effect. It can efficiently participate in the frequency control of the power system, promptly suppress frequency dips, is beneficial to compensating for system inertia, and improves the quality and efficiency of system frequency control, which is of great significance to the safety and stability of power system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 is the abstract drawing;

[0088] Figure 2 is the relationship diagram of the rotor speed and output power of the step-by-step inertia control strategy fan under the conditions of constant wind speed and fixed wind power ratio;

[0089] Figure 3 is the variation diagram of the output power with time during the step-by-step inertia control process;

[0090] Figure 4 is the structure diagram of the autoencoder;

[0091] Figure 5 is the structure diagram of the denoising autoencoder;

[0092] Figure 6 is the structure diagram of the stacked denoising autoencoder;

[0093] Figure 7 is the structure diagram of the deep neural network;

[0094] Figure 8 is the IEEE 9-bus test system;

[0095] Figure 9 is the schematic diagram of the prediction result of the step-by-step inertia control method for wind power frequency modulation based on deep learning;

[0096] Figure 10 is the schematic diagram of different step-by-step inertia control results. DETAILED DESCRIPTION OF THE INVENTION

[0097] The technical solutions of the present invention will be further described in detail below with reference to the drawings:

[0098] The present invention can be implemented in many different forms and should not be construed as limited to the embodiments described herein. On the contrary, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present invention to those skilled in the art.

[0099] The present invention provides a stepwise inertial control method for wind power frequency regulation based on deep learning, which is characterized by including the following three steps: Step 1, obtaining the optimal stepwise inertial control parameters for wind power frequency regulation using the Golden Eagle optimization algorithm in time-domain simulation and generating a data set; Step 2, extracting features from the data set based on a stacked denoising autoencoder; Step 3, generating an optimal stepwise inertial control scheme for wind power frequency regulation by learning data features based on a deep neural network.

[0100] The optimal stepwise inertial control parameters in Step 1 are used when the stepwise inertial control strategy is used in wind power frequency regulation.

[0101] Stepwise inertial control (SIC) is a control strategy for wind power to participate in power system frequency regulation, including two main stages: a short-term over-generation stage and a rotational speed recovery stage. Figure 2 、 Figure 3 respectively represent the relationship between the wind turbine rotor speed and output power and the change of output power during the stepwise inertial control process when the stepwise inertial control strategy is used under the conditions of a fixed wind power ratio and a constant wind speed. The specific process of participating in frequency regulation is as follows:

[0102] 1) The wind turbine normally operates at point A on the MPPT curve, and the output electromagnetic power is , the rotor speed is , and the output power is . T At time 0, a sudden power imbalance occurs in the system, resulting in a frequency drop. When the wind turbine detects that the frequency drop exceeds the dead zone limit, it immediately switches to the stepwise inertial control mode and immediately increases the output power by , and the output power is . The output power is increased from point A to point B, entering the short-term over-generation stage.

[0103] 2) In the short-term over-generation stage, the output power of the wind turbine remains unchanged for a period of time , that is, the output power is always , lasting from point B to point C. In this stage, the electromagnetic power output by the wind turbine is greater than the mechanical power, and the rotor then decelerates. At this time, the rotor motion equation of the wind turbine is:

[0104] (1)

[0105] In the formula: refers to the inertia constant of the wind turbine.

[0106] It should be noted that the short-term over-generation stage must be terminated before the rotor decelerates excessively to avoid the occurrence of rotor stall events in the wind turbine. Therefore, the duration of the short-term over-generation stage should be such that the rotor speed does not reach the minimum speed at the moment , which is generally 0.7 p.u.

[0107] 3) At the moment, the short-term over-generation stage ends, and the output power of the wind turbine suddenly drops . At this time , the rotor speed is . The output power suddenly drops from point C to point D, and the wind turbine enters the recovery stage.

[0108] 4) When the wind turbine enters the speed recovery stage, it operates in the maximum power point tracking mode at this time, and the output power of the wind turbine is . The resulting system power deficit is: . In this stage, the mechanical power of the wind turbine is greater than the electromagnetic power, and the rotor starts to accelerate. After of the recovery time, at the rotor speed changes from to . The electromagnetic power output by the wind turbine will also recover to as the rotor speed increases, and the output power slowly returns from point D to point A along the MPPT trajectory.

[0109] The optimal wind power frequency modulation step-by-step inertia control parameters obtained in step 1 are obtained using the Golden Eagle Optimization Algorithm in time-domain simulation.

[0110] The Golden Eagle Optimization Algorithm (GEO) is inspired by the wisdom of the golden eagle and has advantages such as strong optimization ability and fast convergence speed. When hunting, the golden eagle adjusts its speed at different hovering trajectory stages. In the initial stage of hunting, it shows more tendency to cruise and search for prey. In the later stage, it shows more tendency to attack. The golden eagle adjusts these two parts to capture the best prey within the shortest time and feasible area. This kind of behavior is written as a mathematical model and developed by a method that emphasizes exploration and global optimization.

[0111] Hovering action: The Golden Eagle Optimization Algorithm is based on the hovering motion of the golden eagle. Each golden eagle remembers the best position it has reached so far. The golden eagle wants to attack the prey and cruise to find better food. In each iteration, each golden eagle randomly selects the prey of another golden eagle and hovers at the best position it has visited so far. The golden eagle can also choose to circle its own memory.

[0112] Prey Selection: In each iteration, each golden eagle must select a prey to perform cruising and attacking operations. In the Golden Eagle Optimization Algorithm, the prey is modeled as the best solution found by the golden eagle swarm so far. Each golden eagle can remember the best solution found so far. In each iteration, each search agent selects the target prey from the memory of the entire group. Then, the attack and cruising vectors for each golden eagle's selected prey are calculated. If the new position calculated through the attack and cruising vectors is better than the previous position in the memory, the memory is updated. The prey selection strategy plays an important role in the Golden Eagle Optimization Algorithm. The selection can be carried out in a basic way, where each golden eagle only selects prey from its own memory. To enable the golden eagles to better explore the land, a random one-to-one mapping scheme is proposed, where each golden eagle randomly selects the prey in the current iteration from the memories of other group members. It should be noted that the selected prey is not necessarily the nearest or farthest prey. In this scheme, each prey in the memory is assigned or mapped to a unique golden eagle. Then each golden eagle attacks and cruises on the selected prey.

[0113] Attack Behavior: The attack behavior can be simulated by a vector starting from the current position of the golden eagle and ending at the position of the prey in the golden eagle's memory. The attack vector of the golden eagle can be calculated by Equation (2):

[0114] (2)

[0115] Where: is the attack vector of the i-th golden eagle; is the golden eagle f the best location (prey) reached so far; is the current position of the i-th golden eagle.

[0116] Cruising (Exploration) Behavior: The cruising vector is calculated based on the attack vector. The cruising vector is the tangent vector of the circle and perpendicular to the attack vector. Cruising can also be considered as the linear velocity of the golden eagle relative to the prey. The n-dimensional cruising vector lies in the hyperplane tangent to the circle. Therefore, in order to calculate the cruising vector, the equation of the tangent hyperplane must be calculated first. The equation of the n-dimensional hyperplane can be determined by any point on the hyperplane and a vector perpendicular to the hyperplane, and this vector is called the normal vector of the hyperplane. The scalar form of the hyperplane equation in n n-dimensional space is shown in Equation (3):

[0117] (3)

[0118] Where: is the normal vector; is the variable vector; is any point on the hyperplane; . Substitute If the (attack vector) is regarded as the normal vector of the hyperplane, then the (Golden Eagle cruise vector in the iteration) can represent the hyperplane to which it belongs according to Equation (4).

[0119] (4)

[0120] In the formula: is the attack vector; is the decision / design variable vector; is the position of the selected prey.

[0121] To find a random vector on the cruise hyperplane, we must first find a random destination point C on this hyperplane, rather than the point we already have (the current position of the Golden Eagle). Note that the starting point of its cruise vector is the current position of the Golden Eagle. Since the hyperplane is one dimension smaller than their ambient space, a random 1×n point cannot be simply generated. A simple random point in the n-dimensional space cannot be guaranteed to lie on the cruise hyperplane. A new point on the n-dimensional cruise hyperplane has n - 1 degrees of freedom, meaning that n - 1 dimensions can be freely chosen, but the hyperplane equation determines the last 1 dimension, as shown in Equation (3). The last dimension must be chosen to satisfy the hyperplane equation. Therefore, there are n - 1 free variables and 1 fixed variable. Use the following steps to find a random n n-dimensional destination point C on the cruise hyperplane of the Golden Eagle, and the specific steps are as follows:

[0122] Step 1: Randomly select one variable from the n variables as the fixed variable;

[0123] Step 2: Assign random values to all variables except the k-th variable, because the k-th variable is fixed.

[0124] Step 3: Calculate the value of the fixed variable using Equation (5).

[0125] (5)

[0126] In the formula: is the destination point C of the k-th element; is the attack vector of the j-th element; k is the number of the fixed variable. To find a random destination point on the cruise hyperplane, the general representation of the destination point on the cruise hyperplane is shown in Equation (6):

[0127] (6)

[0128] Transfer to a new position: The displacement of the golden eagle consists of an attack and a vector. We define the step vector of the golden eagle iteration as equations (7) and (8).

[0129] (7)

[0130] (8)

[0131] Where: is the attack coefficient; is the cruise coefficient; , are random vectors within [0,1]. Therefore, by adding the step vector in iteration t to the position in iteration t, the position of the golden eagle in iteration t+1 can be calculated:

[0132] (9)

[0133] Where: is the position of the golden eagle at the (t + 1)-th time; is the position of the golden eagle at the t-th time; is the step size of the movement of the golden eagle.

[0134] If the fitness of the new position of the golden eagle is better than the position in its memory, the memory of this eagle will be updated with the new position. Otherwise, the memory remains intact, but the eagle will be at the new position. In the new iteration, each golden eagle randomly selects a golden eagle from the population, rotates around its best visited position, calculates the attack vector, the cruise vector, and finally calculates the step vector and the new position for the next iteration. This loop will continue to execute until any termination condition is met. There are two coefficients in equation (7), namely the attack coefficient and the cruise coefficient, which are used to control the influence of the attack vector and the cruise vector on the step vector, as shown in equation (10):

[0135] (10)

[0136] Where: t represents the current iteration; T represents the maximum iteration; and are the initial value and the termination value of the attack tendency ; and are the initial value and the termination value of the cruise tendency :

[0137] Furthermore, the basic steps of obtaining the optimal wind power frequency modulation step inertia control parameters and generating a data set using the golden eagle optimization algorithm in the time domain simulation in step 1 are as follows:

[0138] Step 1) Initialize the number of golden eagles;

[0139] Step 2) Calculate the fitness function and initialize the population memory;

[0140] Step 3) Initialize 、 ;

[0141] Step 4) Calculate Equation (10) and update 、 ;

[0142] Step 5) Calculate Equation (2), and randomly select a prey from the attack vectors calculated from the population memory;

[0143] Step 6) Calculate the cruise vector (Equations (3)-(6)), calculate the step vector (Equations (7)-(9)), update the position (Equation (9)), and evaluate the fitness function of the new position;

[0144] Step 7) Update the optimal solution and the optimal position;

[0145] Step 8) Determine whether the maximum number of iterations is satisfied. If so, output the optimal golden eagle position and the global optimal solution. Otherwise, return to Step 4 for iterative calculation;

[0146] Furthermore, the feature extraction of the dataset based on the stacked denoising autoencoder in Step 2 is performed by the stacked denoising autoencoder.

[0147] The autoencoder (AE) is as Figure 4 shown. The autoencoder consists of an encoder and a decoder. The encoder and the decoder have hidden layers: the input is converted into a latent representation in the hidden layer through the encoder, and then the internal representation is converted into an output through the decoder. The output is equivalent to the reconstruction of the input and should be as close to the input as possible.

[0148] Select the input sample set , which is composed of N groups of samples . Among them, N is the number of sample groups, and n is the number of samples in each group of samples. Let the set of hidden layer feature vectors be , which is composed of N groups of feature vectors . Among them, m is the number of vectors in each group of feature vectors. Then the encoding relationship between and is

[0149] [[ID=�8]] (11)

[0150] In the formula: is the weight matrix between the input layer and the hidden layer; is the bias matrix between the input layer and the hidden layer; is the neuron activation function of the encoder, usually the sigmoid function is used, which has good feature recognition.

[0151] (12)

[0152] Where: is the input vector.

[0153] The decoder is the inverse operation of the encoder, taking the feature vector of the hidden layer as the input vector. Let is the output vector set, there are N groups, the dimension is n, then the expression of the decoder is

[0154] (13)

[0155] Where: is the weight matrix of the hidden layer and the output layer; is the bias matrix of the hidden layer and the output layer; is the neuron activation function of the decoder.

[0156] Autoencoders achieve feature learning by minimizing the reconstruction error between the output vector and the input vector. They use a gradient descent algorithm to continuously adjust the network weights and biases to reduce the reconstruction error. However, autoencoder learning may simply retain the original input data information and cannot ensure an effective representation of feature information.

[0157] The denoising autoencoder (DAE) is an improvement on the autoencoder, and its architecture is as follows Figure 5 As shown in Figure 2, this approach builds on the autoencoder by adding noise to a portion of the input and training it to recover the original noise-free input. Using a denoising autoencoder allows the hidden layer to learn more effective data representations and uncover deeper hidden information.

[0158] The stacked denoising autoencoder (SDAE) consists of multiple DAEs, such as Figure 6 shown. W n is the weight matrix between the n-1th hidden layer and the nth hidden layer, B$n$ is the bias matrix between the $(n - 1)$-th hidden layer and the $n$-th hidden layer. The training is completed through two steps: unsupervised greedy layer-by-layer pre-training and supervised refinement training to achieve feature learning. The input of the first-layer DAE is the original data, and the output data of the hidden layer is used as the input data of the upper-layer DAE. Noise is added to the input data before the training of each layer. The stacked denoising autoencoder can efficiently extract the high-order features of the data through training, better fit complex functions, improve the speed of obtaining network parameters, and enhance the deep feature learning ability of the neural network.

[0159] The basic steps of feature extraction from the dataset based on the stacked denoising autoencoder in Step 2 are as follows:

[0160] Step 1) Normalize the dataset obtained using the Golden Eagle optimization algorithm in the time-domain simulation.

[0161] Step 2) Divide the training set and the validation set. Use 80% of the original data as the training set and the remaining 20% as the test data to generate the training and test data.

[0162] Step 3) Construct a stacked denoising autoencoder model and select the number of hidden layers and the number of neurons.

[0163] Step 4) Layer-by-layer greedy unsupervised pre-training. Use the data with added noise as the input data and adopt the dropout technique to randomly select some neurons to be temporarily inactive.

[0164] Step 5) Supervised refinement training. Fine-tune the weights and biases of the network until the number of iterations reaches the set value.

[0165] The generation of the optimal wind power frequency regulation step-by-step inertia control scheme based on learning the data features by the deep neural network in Step 3 is carried out on the basis of the deep neural network.

[0166] A deep neural network (DNN) is a neural network with multiple hidden layers and is also known as a deep feedforward network (DFN). Divided by the position of different layers, the neural network layers inside the DNN can be classified into: the input layer, the hidden layer, and the output layer. Generally, the first layer is the input layer, the last layer is the output layer, and the middle layers are all hidden layers. The layers are fully connected, that is, any neuron in the $i$-th layer must be connected to any neuron in the $(i + 1)$-th layer. The basic structure is as Figure 7 shown. From the perspective of a local model, a complex DNN is the same as a perceptron, that is, a linear relationship , plus an activation function .

[0167] The basic steps for generating the optimal wind power frequency modulation step-by-step inertia control scheme based on learning data features by a deep neural network in Step 3 are as follows:

[0168] Step 1) Use the features extracted by the stacked denoising autoencoder as the input of the deep neural network;

[0169] Step 2) Divide the training set and the validation set. Use 80% of the original data as the training set and the remaining 20% as the test data to generate the training and test data;

[0170] Step 3) Build a deep neural network model and select the number of hidden layers and the number of neurons in each hidden layer;

[0171] Step 4) Use Adam to optimize the input weight matrix, the hidden layer feature vector, and the output weight matrix of the network, and selectively add regularization to the hidden layer;

[0172] Step 5) Use MSE as the evaluation index to judge whether it no longer decreases. If it is satisfied, the training is completed; otherwise, return to Step 3;

[0173] Step 6) Save the trained model. When encountering a load disturbance event next time, directly use the model to generate the optimal wind power frequency modulation step-by-step inertia control scheme.

[0174] Take the IEEE 9-bus system as the example test system and connect the wind turbine model to line L3 in the system, as Figure 8 shown. All the research involved is obtained from the simulation operation of the wind turbine at a power lower than the rated power. To meet the requirements of using this method in different scenarios, consider the influence of wind speed, wind power proportion, and load disturbance amount on the system frequency characteristics. Set the parameters of SIC from these three aspects respectively. In this example, set the wind speed from 4 m / s to 11 m / s, increasing by 1 m / s for each case. Set the wind power proportion from 5% to 60%, increasing by 5% for each case. Set the load disturbance amount from 1.005 to 1.25, increasing by 0.005 for each case. There are a total of 4800 cases. The schematic diagram of the prediction results of the optimal SIC parameter combination is as Figure 9 shown. And add the effect comparison with other frequency modulation control methods in the typical scenario. The scenario in the system is wind speed 5 m / s, wind power proportion 25%, and load disturbance amount 0.035 p.u. In this scenario, compare with other SIC controls and the frequency control without wind power participation. The frequency control effect is shown in Figure 10 shown.

[0175] The above embodiments are only for illustrating the technical idea of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention in accordance with the technical idea proposed by the present invention shall be included within the protection scope of the present invention.

Claims

1. A stepwise inertia control method for wind power frequency regulation based on deep learning, characterized in that It includes the following three steps: Step 1: Use the Golden Eagle Optimization Algorithm in time-domain simulation to obtain the optimal step-by-step inertial control parameters for wind power frequency regulation and generate a dataset; Step 2: Extract features from the dataset based on the stacked denoising autoencoder; Step 3: Generate the optimal step-by-step inertial control scheme for wind power frequency regulation by learning data features based on a deep neural network; The basic steps of Step 1 are as follows: Step 1.1: Initialize the number of golden eagles; Step 1.2: Calculate the fitness function and initialize the population memory; Step 1.3: Initialize the attack coefficient p of the Golden Eagle a and the cruise coefficient p c ; Step 1.4, calculate formula (10) and update p a , p c ; Step 1.5: Calculate Equation (2), and randomly select a prey from the attack vectors calculated from the memory of the population; Step 1.6: Calculate the cruise vector (Equations (3)-(⑥)), calculate the step-size vector (Equations (⑦)-(⑨)), update the position (Equation (⑨)), and evaluate the fitness function of the new position; Step 1.7: Update the optimal solution and the optimal position; Step 1.8: Determine whether the maximum number of iterations is satisfied. If so, output the optimal golden eagle position and the global optimal solution. Otherwise, return to Step 1.4 for re-iterative calculation; Formulas (2)-(10) in the above steps are defined as: In the formula: is the attack vector of the i-th golden eagle; is the best location (prey) that the golden eagle f has reached so far; is the current position of the i-th golden eagle; where: H = [h1, h2, …, h n is the normal vector; X = [x1, x2, …, x n is the variable vector; P = [p1, p2, …, p n is an arbitrary point on the hyperplane; Where: A = [a1, a2, …, a n is the attack vector; X = [x1, x2, …, x n is the decision / design variable vector; is the position of the selected prey; Where: c k is the k-th element of the target point C; a j is the j-th element of the attack vector ; k is the number of the fixed variable; where: p a is the attack coefficient; p c is the cruise coefficient; is a random vector within [0, 1]; Therefore, by adding the step vector in iteration t to the position in iteration t, the position of the golden eagle in iteration t + 1 can be calculated: where: x t+1 is the position of the golden eagle at the (t + 1)-th time; x t is the position of the golden eagle at the t-th time; is the step size of the movement of the golden eagle; where: t represents the current iteration; T represents the maximum iteration; and are the initial value and the termination value of the attack tendency p a ; and are the initial value and the termination value of the cruising tendency p c ; The basic steps in Step 2 are as follows: Step 2.1: Normalize the dataset obtained by using the Golden Eagle Optimization Algorithm in time-domain simulation; Step 2.2: Divide the training set and the validation set. Use 80% of the original data as the training set and the remaining 20% as the test data to generate training and test data; Step 2.3: Construct a stacked denoising autoencoder model and select the number of hidden layers and neurons; Step 2.4: Perform layer-by-layer greedy unsupervised pre-training. Use the data with added noise as the input data and adopt the dropout technique to randomly select some neurons to be temporarily inactive; Step 2.5: Perform supervised refinement training to fine-tune the weights and biases of the network until the number of iterations reaches the set value.

2. A step-by-step inertial control method for wind power frequency regulation based on deep learning according to Claim 1, wherein the optimal step-by-step inertial control parameters in Step 1 are used when the step-by-step inertial control strategy is used in wind power frequency regulation.

3. A step-by-step inertial control method for wind power frequency regulation based on deep learning according to Claim 1, wherein the feature extraction of the dataset in Step 2 is performed by a stacked denoising autoencoder; An autoencoder (AE) consists of an encoder and a decoder. The encoder and the decoder have hidden layers: the input is converted into a latent representation in the hidden layer through the encoder, and then the internal representation is converted into an output through the decoder. The output is equivalent to the reconstruction of the input and should be as close as possible to the input; Select the input sample set X, which consists of N groups of samples x1, x2, …, x n and, N is the number of sample groups, and n is the number of samples in each group of samples; let the set of hidden layer feature vectors be H, and H is composed of N groups of feature vectors h1, h2, …, h m which are composed of m vectors in each group of feature vectors. Then, the coding relationship between X and H is H = s f (WX + B)(11) Where: W is the weight matrix between the input layer and the hidden layer; B is the bias matrix between the input layer and the hidden layer; s f is the neuron activation function of the encoder, usually the sigmoid function, which has good feature recognition ability; In the formula: z is the input vector; The decoder is the inverse operation of the encoder. Taking the feature vector of the hidden layer as the input vector, let Y be the set of output vectors, with a total of N groups and a dimension of n. Then the expression of the decoder is Y = s g (W'X + B') (13) Where: W' is the weight matrix between the hidden layer and the output layer; B' is the bias matrix between the hidden layer and the output layer; s g is the neuron activation function of the decoder; The autoencoder achieves the purpose of feature learning by minimizing the reconstruction error between the output vector and the input vector, and uses the gradient descent algorithm to continuously adjust the network weights and bias to reduce the reconstruction error. However, the learning of the autoencoder may simply retain the original input data information and cannot ensure an effective representation of feature information. The denoising autoencoder (DAE) is an improvement on the autoencoder. It adds noise to a portion of the input value and trains it to restore the original noise-free input. After using the denoising autoencoder, the hidden layer can learn more effective data representation and mine deeper hidden information. The stacked denoising autoencoder (SDAE) consists of multiple DAEs, where Wn is the weight matrix between the n-1th hidden layer and the nth hidden layer, and Bn is the bias matrix between the n-1th hidden layer and the nth hidden layer. Training is completed through two steps: unsupervised greedy layer-by-layer pre-training and supervised refinement training to achieve feature learning. The input of the first layer DAE is the original data, and the output data of the hidden layer is used as the input data of the upper layer DAE. Noise must be added to the input data before training each layer. Through training, the stacked denoising autoencoder can efficiently extract high-order features of the data, better fit complex functions, speed up the acquisition of network parameters, and improve the deep feature learning ability of the neural network.

4. According to the method for wind power frequency regulation step-by-step inertia control based on deep learning in claim 1, the basic steps for generating the optimal wind power frequency regulation step-by-step inertia control scheme based on deep neural network learning data features in step 3 are as follows: Step 1) using the features extracted by the stacked denoising autoencoder as input to the deep neural network; Step 2) Divide the training set and the validation set, using 80% of the original data as the training set and the remaining 20% as the test data to generate training and test data; Step 3) Build a deep neural network model and select the number of hidden layers and the number of neurons in each hidden layer; Step 4) Use Adam to optimize the network's input weight matrix, hidden layer feature vectors, and output weight matrix, and selectively add regularization to the hidden layer; Step 5) Use MSE as the evaluation indicator to determine whether it no longer decreases. If it does, complete the training; otherwise, return to step 3. Step 6) Save the trained model. Next time a load disturbance event occurs, directly use the model to generate the optimal wind power frequency regulation step-by-step inertia control scheme.

Images