CN114268129A - Wind power frequency modulation gradual inertia control method based on deep neural network - Google Patents

Abstract

The invention discloses a wind power frequency modulation gradual inertia control method based on a deep neural network, which comprises the following steps: obtaining wind power frequency modulation optimal gradual inertia control parameters in time domain simulation by using a parasitic-predation algorithm and generating a data set; b, extracting the characteristics of the data set generated in the step A based on a noise reduction automatic encoder; and C, learning the features extracted in the step B based on a deep neural network and generating an optimal gradual inertia control scheme of wind power frequency modulation. When the method is faced with different load disturbance events, a corresponding frequency control scheme can be rapidly, efficiently and economically provided, all strength of gradual inertia control is exerted, and the requirement of excellent gradual inertia control during wind power frequency modulation is met.

Description

Wind power frequency modulation gradual inertia control method based on deep neural network

Technical Field

The invention relates to the technical field of wind power frequency modulation, in particular to a wind power frequency modulation gradual inertia control method based on a deep neural network.

Background

In order to cope with various problems of climate, environment, energy and the like, the proportion of renewable energy sources in power systems is increasing. Wind power has been steadily and continuously developed as an important renewable energy source. The wind power provides clean energy, meanwhile, the inertia of the power system is reduced, the capability of the system for coping with active power unbalance is weakened, and a serious challenge is brought to the frequency control quality and the frequency stability of the system. The wind turbine is generally controlled by maximum power point tracking, is connected to a power grid through a power electronic converter, is decoupled from system frequency, cannot respond to power deviation by releasing or absorbing energy, does not have inertia response characteristics, and cannot actively provide inertia support for the power grid under active power disturbance. In order to improve the quality and efficiency of the system frequency control, it is necessary to have the wind turbine participate in this control and in this way compensate for the reduction in the inertia of the whole system. Recently, new grid codes also require that wind farms must contribute to power system frequency control. To meet the requirements, when an active power imbalance occurs in the grid, gradual inertial control can be used to participate in grid frequency control, which can quickly provide transient power support.

In the process of frequency modulation by using gradual inertia control, the fan provides increased power for a period of time to the power grid by releasing the kinetic energy of the rotor, and the rotating speed of the rotor is reduced. If the fan rotor speed drops to its defined lower limit, the fan will be disconnected from the power system. Therefore, the wind turbine can only provide a short power support to the grid in addition by releasing limited kinetic energy of the rotor. Gradual inertia control must also ensure that the fan rotor terminates power overshoot before reaching the minimum rotational speed. However, the sudden termination of a power overshoot may lead to a new drop in the grid frequency, also referred to as a Secondary Frequency Droop (SFD). Under the gradual inertia control, the frequency modulation effect is in complex relation with the secondary frequency drop, and different methods can be adopted to reduce the secondary frequency drop on the premise of ensuring the good frequency modulation effect.

At present, smooth termination of the gradual inertia control is proposed, the smooth termination being achieved by reducing the wind turbine output active power, which is affected by the rotor speed and wind penetration. The validity of the smooth termination is verified in different cases. The main drawback of a smooth termination is the reduction of kinetic energy released to the power system, which will limit the contribution of the wind turbine to the frequency recovery of the system. It has also been proposed to use a high voltage dc link to compensate for power reduction of the wind turbine at the terminals. Due to the fast response characteristic of the high-voltage direct-current link, the power attenuation compensation technology has good performance in the aspect of relieving secondary frequency drop. However, this method requires a large investment.

The existing improvement method under the gradual inertia control is difficult to provide a corresponding frequency control scheme quickly, efficiently and economically in the face of different load disturbance events, exerts all strength of the gradual inertia control, and is difficult to meet the requirement of excellent gradual inertia control in wind power frequency modulation.

Disclosure of Invention

The invention aims to provide a wind power frequency modulation gradual inertia control method based on a deep neural network. When the method is faced with different load disturbance events, a corresponding frequency control scheme can be rapidly, efficiently and economically provided, all strength of gradual inertia control is exerted, and the requirement of excellent gradual inertia control during wind power frequency modulation is met.

The technical scheme of the invention is as follows: a wind power frequency modulation gradual inertia control method based on a deep neural network comprises the following steps:

A. obtaining wind power frequency modulation optimal gradual inertia control parameters in time domain simulation by using a parasitic-predation algorithm and generating a data set;

B. b, extracting the characteristics of the data set generated in the step A based on a noise reduction automatic encoder;

C. and C, learning the features extracted in the step B based on a deep neural network and generating an optimal gradual inertia control scheme of wind power frequency modulation.

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the wind power frequency modulation optimal gradual inertia control parameter in the step a is used when a gradual inertia control strategy is used in wind power frequency modulation;

the gradual inertia control is a control strategy for wind power to participate in power grid frequency adjustment, and comprises two main stages: a short-time over-sending stage and a rotating speed recovery stage.

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the specific process of the wind power participating in the power grid frequency modulation by using the gradual inertia control method is as follows:

1) the fan normally works at the point A on the MPPT curve, and the output electromagnetic power is P_w0The rotor speed is omega₀Output power P_w＝P_MPPT＝P_w0；

At the time of T0, the power imbalance of the power system suddenly occurs to cause the frequency to drop, the fan monitors that the frequency drops to exceed the dead zone limit value, then the mode is switched to the gradual inertia control mode, and the output power delta P is immediately increased_upOutput power P_w＝P_w0+ΔP_upThe output power is increased from the point A to the point B, and a short-time over-transmitting stage is entered;

2) in the short-time over-sending stage, the output power of the fan is within a period of time delta T_upInternally held constant, i.e. output power always being P_w0+ΔP_upFrom point B to point C;

at the stage, the electromagnetic power output by the fan is larger than the mechanical power, the rotor is decelerated immediately, and the equation of motion of the rotor of the fan is

H_wIs the fan inertia constant;

duration of short-time superhair phase Δ T_upThe value of (A) is to ensure that the rotor speed is at T_offDoes not reach the minimum rotating speed omega at the moment_min；

3)T_offAt the moment, the short-time over-sending stage is ended, and the output power of the fan is suddenly reduced by delta P_offAt this time

Rotor speed of omega_offThe output power is suddenly reduced from the point C to the point D and enters a recovery stage;

4) the fan enters a rotating speed recovery stage, the fan operates in a maximum power point tracking mode at the moment, and the output power of the fan is

The generated system power shortage is delta P_off＝-P_w0+ΔP_off-P_wAt this stage, the mechanical power of the fan is greater than the electromagnetic power, the rotor starts to accelerate and passes through delta T_rAt recovery time of T_endThe rotor speed at the moment is controlled by omega_offRestore to omega₀The electromagnetic power output by the fan is recovered to P along with the increase of the rotating speed of the rotor_w0The output power slowly returns from point D to point a along the MPPT trajectory.

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the parasitic-predation algorithm simulates predation and parasitic behaviors among cats, cuckoos and crows to form a parasitic and predation system of the crows-cuckoos-cats, and the parasitic-predation algorithm specifically comprises the following contents:

1) randomly performing population initialization within a boundary range:

2) at the beginning, the number of crows will decrease with time; simulating crow to fly through two states, randomly generating candidate crow, and generating a new position for crow i

r1 is a random number, LF is a Levy flight function;

3) in the nesting stage, the optimal bird nest is used for updating, and a bird nest solution is found through the Levy flight process; the population initialization state beyond the dimension range is carried out according to the formula (4);

4) entering a parasitism (crow-cuckoo) phase;

4.1) in the initial stage, the predation efficiency is low, and cats try to kill cuckoos; the predation efficiency is high, so that cuckoos are killed; the efficiency of cuckoos is assumed to be small/medium, the efficiency of cats decreases; at this stage, partial crow eggs are replaced by cuckoo eggs which are similar to the crow eggs and are difficult to be found; in addition, the nests that are to be populated are selected according to fitness, the better the nests are, the more easily they are populated. Create a new solution/nest to replace part of the nest and with probability P_aPartially inferior nests were found; a new nest of cuckoo is available,

selecting the position of the bird' S nest using a wheel, S_GObeying a uniform gaussian distribution, k being a binary matrix k-rand [0, 1%]＞P_a(7)，P_aThe incremental factor is T/2T or G/2G, wherein T or G is the iteration number at the moment, and T or G is the maximum iteration number;

at the beginning of the parasitic phase, matrix k is filled with 1; k is gradually increased after the last time to keep the diversity of the population;

5) entering a predation stage;

5.1) at the initial stage, the predation efficiency is high, the number of cats and crows is rapidly increased, sufficient living resources cannot be provided for cuckoos, and the cuckoos are extinct; the stage is a crow-cat stage based on a cat tracking mode; the search mode need not be executed because the cat knows the search space is empty; at this stage, the cuckoo found a cat-repelling compound; the cat tracks the nest where the cuckoo is absent by using low-odor secretion, selects an unparasitized nest and tracks the nest randomly; the cat moves to each dimension according to the speed of the cat; cats have high predation efficiency, resulting in a rapid growth, while crows and cuckoos have a slow growth;

this phase comprises three steps:

step a, updating the speed of each dimension as follows:

v_k，d＝v_k，d+r*c*(x_best，d-x_k，d)(8)，v_k，dis cat_kVelocity in d dimension, x_best，dThe location of the cat with the best fitness value, x_k，dIs cat_kC is a constant, r is [0, 1 ]]A random number within a range;

b, checking whether the updating speed exceeds the maximum speed range; setting the new speed equal to the limit if it is greater than the maximum speed;

step c, updating cat_kPosition x of_k，d＝v_k，d+x_k，d(9)。

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the basic steps of obtaining the optimal gradual inertia control parameter of the wind power frequency modulation in the time domain simulation by using the parasitic-predation algorithm and generating the data set in the step a are as follows:

a1, initializing two control parameters delta P_up、ΔT_upPopulation and algorithm parameters of;

a2, evaluation parameter Δ P_up、ΔT_upFitness function value ofDetermining an optimal parameter solution and an optimal parameter bird nest;

a3, calculating parameter delta P_up、ΔT_upThe corresponding number of cat groups, crow groups and bird nests;

a4, entering a nesting stage according to a formula

And

updating the control parameter Δ P_up、ΔT_upThe bird nest position;

a5, entering a parasitic stage according to the formula

S_G＝(X_r2-X_r3) Rand and k ═ rand [0, 1 ]]＞P_aUpdating the control parameter Δ P_up、ΔT_upThe bird nest position;

a6, entering the predation stage according to the formula v_k，d＝v_k，d+r*c*(x_best，d-x_k，d) And x_k，d＝v_k，d+x_k，dUpdating the control parameter Δ P_up、ΔT_upThe bird nest position;

a7, reevaluating Δ P_up、ΔT_upAnd updating the global optimal solution;

a8, judging whether an iteration condition is met, if so, outputting an optimal solution, and otherwise, returning to the step A2 to repeat iteration updating calculation again;

a9, optimal Delta P to be output_up、ΔT_upAnd the wind speed, the wind power ratio and the load disturbance quantity under the corresponding scene are combined in a pairing mode to generate a data set.

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the feature extraction of the data set based on the noise reduction automatic encoder in the step B is performed by the noise reduction automatic encoder;

the automatic encoder consists of an encoder and a decoder, wherein the encoder and the decoder are provided with hidden layers, the encoder converts input into potential representation in the hidden layers, then the decoder converts internal representation into output, and the output is equivalent to reconstruction of the input and is as close to the input as possible;

selecting an input sample set X consisting of N sets of samples X₁、x₂、…、x_nThe method comprises the following steps of (1) forming a sample group, wherein N is the number of the 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 the set of hidden layer feature vectors is composed of N sets of feature vectors H₁、h₂、…、h_mWherein m is the number of vectors in each group of eigenvectors, and the coding relationship between X and H is H ═ s_f(WX + B) (10), wherein W is a weight matrix of the input layer and the hidden layer; b is an input layer and hidden layer threshold matrix; s_fFor the neuron activation function of the encoder, a sigmoid function is usually used, which has good feature discrimination,

z is an input vector;

the decoder is the inverse operation of the encoder, the characteristic vector of the hidden layer is used as an input vector, Y is set as an output vector set, N groups are shared, the dimension is N, and the expression of the decoder is

Y＝s_g(W ' X + B ') (9), wherein W ' is a weight matrix of the hidden layer and the output layer; b' is a threshold matrix of a hidden layer and an output layer; s_gActivating a function for a neuron of a decoder;

the noise reduction automatic encoder is an improvement of an automatic encoder, and on the basis of the automatic encoder, noise is added to a part of input values, and the noise reduction automatic encoder is trained to restore the original noise-free input.

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the basic steps of extracting the features of the data set based on the noise reduction automatic encoder in the step B are as follows:

b1, carrying out normalization processing on the data set obtained in the time domain simulation by using a parasitic-predation algorithm;

b2, dividing a training set and a verification set, taking 80% of original data as the training set, and taking the remaining 20% as test data to generate training and test data;

b3, greedy unsupervised pre-training layer by layer, taking the data added with noise as input data, and randomly selecting part of neurons to temporarily stop working by adopting a Dropout technology;

b4, fine-tuning the weight and the threshold value of the network by adopting a top-down small-batch Adagarad optimization method for the initialized network parameters until the iteration number reaches a set value.

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the learning of the data characteristics based on the deep neural network in the step C to generate the wind power frequency modulation optimal gradual inertia control scheme is performed on the basis of the deep neural network;

the deep neural network is a neural network with a plurality of hidden layers, also called a deep feedforward network; according to the position division of different layers, the neural network layer inside the DNN can be divided into: the input layer, the hidden layer and the output layer, wherein the first layer is the input layer, the last layer is the output layer, and the middle layers are all the hidden layers; the layers are all connected, namely any neuron on the ith layer is connected with any neuron on the (i + 1) th layer; the complex DNN is from the local model, namely a linear relation as the perceptron

Plus an activation function sigma (z).

In the wind power frequency modulation gradual inertia control method based on the deep neural network, the basic steps of learning the data characteristics based on the deep neural network in the step C to generate the optimal wind power frequency modulation gradual inertia control scheme are as follows:

c1, taking the features extracted by the noise reduction automatic encoder as the input of the deep neural network;

c2, dividing a training set and a verification set, taking 80% of original data as the training set, and taking the remaining 20% as test data to generate training and test data;

c3, constructing a deep neural network model, and selecting the number of hidden layer layers and the number of neurons of each hidden layer;

c4, taking MSE as an evaluation index, judging whether the MSE is not reduced any more, finishing training if the MSE is not reduced any more, and returning to C3 if the MSE is not reduced any more;

and C5, storing the trained model, and directly using the model to generate the optimal gradual inertia control scheme of the wind power frequency modulation when encountering a load disturbance event next time.

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

1. when wind power frequency modulation gradual inertia control is carried out, compared with a wind power frequency modulation strategy with a single consideration factor, the method comprehensively considers the influence of different wind speeds, wind power occupation ratios and load disturbance quantities on the frequency characteristics of the system, has more reference factors, can comprehensively and effectively obtain the optimal gradual inertia control parameters, meets the use requirements of the wind power frequency modulation in different scenes in the power system, and has good generalization capability;

2. compared with a time domain simulation method, the method has the advantages that the calculation speed is high based on deep learning, and the frequency control time for online decision making is greatly saved; meanwhile, the optimal gradual inertia parameter obtained through the feature extraction of the noise reduction automatic encoder and the deep neural network learning has extremely high accuracy, and the requirement of quickly realizing frequency control in the actual operation of wind power frequency modulation is met.

3. Compared with other gradual inertia control methods, the method has the advantages that the optimal gradual inertia control parameters are obtained, a better frequency modulation effect can be obtained, the method can effectively participate in power grid frequency control, frequency drop is restrained in time, compensation of system inertia is facilitated, quality and efficiency of system frequency control are improved, and the method has great significance on safety and stability of power grid operation.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a diagram showing the relationship between the rotor speed and the output power of a fan according to a gradual inertia control strategy under the conditions of constant wind speed and fixed wind power ratio;

FIG. 3 is a graph of output power over time during the progressive inertial control of the present invention;

FIG. 4 is a block diagram of an automatic encoder of the present invention;

FIG. 5 is a block diagram of a noise reduction autoencoder of the present invention;

FIG. 6 is a diagram of a deep neural network architecture of the present invention;

FIG. 7 is an IEEE 9 node test system of the present invention;

FIG. 8 is a schematic diagram illustrating the prediction results of the optimal parameters for the stepwise inertial control of the present invention;

FIG. 9 is a diagram illustrating the results of different stepwise inertial control.

Detailed Description

The invention is further illustrated by the following figures and examples, which are not to be construed as limiting the invention.

Examples are given. A wind power frequency modulation gradual inertia control method based on a deep neural network is disclosed, as shown in figure 1, and comprises the following steps:

A. obtaining wind power frequency modulation optimal gradual inertia control parameters in time domain simulation by using a parasitic-predation algorithm and generating a data set;

B. b, extracting the characteristics of the data set generated in the step A based on a noise reduction automatic encoder;

C. and C, learning the features extracted in the step B based on a deep neural network and generating an optimal gradual inertia control scheme of wind power frequency modulation.

The optimal gradual inertia control parameter of the wind power frequency modulation is used when a gradual inertia control strategy is used in the wind power frequency modulation;

the gradual inertia control is a control strategy for wind power to participate in power grid frequency adjustment, and comprises two main stages: a short-time over-sending stage and a rotating speed recovery stage.

The attached figures 2 and 3 respectively show the relation between the rotating speed of a fan rotor and output power of a step-by-step inertia control strategy under the conditions of constant wind speed and fixed wind power ratio and the change of the output power along with time in the step-by-step inertia control process, and the specific process that wind power participates in power grid frequency modulation by using a step-by-step inertia control method is as follows:

1) the fan normally works at the point A on the MPPT curve, and the output electromagnetic power is P_w0The rotor speed is omega₀Output power P_w＝P_MPPT＝P_w0；

At the time of T0, the power imbalance of the power system suddenly occurs to cause the frequency to drop, the fan monitors that the frequency drops to exceed the dead zone limit value, then the mode is switched to the gradual inertia control mode, and the output power delta P is immediately increased_upOutput power P_w＝P_w0+ΔP_upThe output power is increased from the point A to the point B, and a short-time over-transmitting stage is entered;

2) in the short-time over-sending stage, the output power of the fan is within a period of time delta T_upInternally held constant, i.e. output power always being P_w0+ΔP_upFrom point B to point C;

at the stage, the electromagnetic power output by the fan is larger than the mechanical power, the rotor is decelerated immediately, and the equation of motion of the rotor of the fan is

H_wIs the fan inertia constant;

the short-time overjet phase must be terminated before the rotor is excessively decelerated, and therefore the duration Δ T of the short-time overjet phase_upThe value of (A) is to ensure that the rotor speed is at T_offDoes not reach the minimum rotating speed omega at the moment_min，ω_minTypically 0.7 p.u.;

3)T_offat the moment, the short-time over-sending stage is ended, and the output power of the fan is suddenly reduced by delta P_offAt this time

Rotor speed of omega_offThe output power is suddenly reduced from the point C to the point D and enters a recovery stage;

4) the fan enters a rotating speed recovery stage, the fan operates in a maximum power point tracking mode at the moment, and the output power of the fan is

Produced byThe power shortage of the system is delta P_off＝-P_w0+ΔP_off-P_wAt this stage, the mechanical power of the fan is greater than the electromagnetic power, the rotor starts to accelerate and passes through delta T_rAt recovery time of T_endThe rotor speed at the moment is controlled by omega_offRestore to omega₀The electromagnetic power output by the fan is recovered to P along with the increase of the rotating speed of the rotor_w0The output power slowly returns from point D to point a along the MPPT trajectory.

The parasitic-predation algorithm simulates predation and parasitic behaviors among the cat, the cuckoo and the crow to form a parasitic and predation system of the crow, the cuckoo and the cat, and has the advantages of high precision, high convergence speed and the like; the specific contents of the parasitic-predation algorithm are as follows:

1) randomly performing population initialization within a boundary range:

2) at the beginning, the number of crows will decrease with time; simulating crow to fly through two states, randomly generating candidate crow, and generating a new position for crow i

r1 is a random number, LF is a Levy flight function;

3) in the nesting stage, the optimal bird nest is used for updating, and a bird nest solution is found through the Levy flight process; levy flights are random walks observed by most species; the step length of the Levy flight is controlled by the probability distribution of the heavy tail, which is called Levy distribution; levy flight is superior to uniform random distribution in the aspect of exploring a search space, so that the Levy flight replaces uniform random motion to simulate the avoidance behavior of local optimum trapping and premature convergence, and the overall exploration capacity is improved; the population initialization state beyond the dimension range is carried out according to the formula (4);

the random change of the population is displayed through reinitialization, and the diversity of the search space is increased; this phase is designed as a pure exploration phase, Levy flights are used in the first state, so that crow is scattered throughout the search space;

4) entering a parasitic stage;

4.1) in the initial stage, the predation efficiency is low, and cats try to kill cuckoos; the predation efficiency is high, so that cuckoos are killed; the efficiency of cuckoos is assumed to be small/medium, the efficiency of cats decreases; at this stage, partial crow eggs are replaced by cuckoo eggs which are similar to the crow eggs and are difficult to be found; in addition, the nests that are to be populated are selected according to fitness, the better the nests are, the more easily they are populated. Create a new solution/nest to replace part of the nest and with probability P_aPartially inferior nests were found; a new nest of cuckoo is available,

S_G＝(X_r2-X_r3)*rand(6)，

the position of the bird nest is selected by using a wheel disc, SG follows uniform Gaussian distribution, and k is a binary matrix k ═ rand [0, 1 ]]＞P_a(7)，P_aThe incremental factor is T/2T or G/2G, wherein T or G is the iteration number at the moment, and T or G is the maximum iteration number;

at the beginning of the parasitic phase, matrix k is filled with 1; k is gradually increased after the last time to keep the diversity of the population;

5) entering a predation stage:

5.1) at the initial stage, the predation efficiency is high, the number of cats and crows is rapidly increased, sufficient living resources cannot be provided for cuckoos, and the cuckoos are extinct; the stage is a crow-cat stage based on a cat tracking mode; the search mode need not be executed because the cat knows the search space is empty; at this stage, the cuckoo found a cat-repelling compound; the cat tracks the nest where the cuckoo is absent by using low-odor secretion, selects an unparasitized nest and tracks the nest randomly; the cat moves to each dimension according to the speed of the cat; cats have high predation efficiency, resulting in a rapid growth, while crows and cuckoos have a slow growth;

this phase comprises three steps:

step a, updating the speed of each dimension as follows:

v_k，d＝v_k，d+r*c*(x_best，d-x_k，d)(8)，v_k，dis cat_kVelocity in d dimension, x_best，dThe location of the cat with the best fitness value, x_k，dIs cat_kC is a constant, r is [0, 1 ]]A random number within a range;

b, checking whether the updating speed exceeds the maximum speed range; if the new speed is greater than the maximum speed, it is set equal to the limit (the speed limit is modified to decrease linearly from 1 to 0.25);

step c, updating cat_kPosition x of_k，d＝v_k，d+x_k，d(9)。

The basic steps of obtaining the optimal gradual inertia control parameter of the wind power frequency modulation in the time domain simulation by using the parasitic-predation algorithm and generating the data set in the step A are as follows:

a1, initializing two control parameters delta P_up、ΔT_upPopulation and algorithm parameters of;

a2, evaluation parameter Δ P_up、ΔT_upDetermining an optimal parameter solution and an optimal parameter bird nest;

a3, calculating parameter delta P_up、ΔT_upThe corresponding number of cat groups, crow groups and bird nests;

a4, entering a nesting stage according to a formula

And

updating the control parameter Δ P_up、ΔT_upThe bird nest position;

a5, enter and sendGrowth stage according to the formula

S_G＝(X_r2-X_r3) Rand and k ═ rand [0, 1 ]]＞P_aUpdating the control parameter Δ P_up、ΔT_upThe bird nest position;

a6, entering the predation stage according to the steps a to c, namely the formula v_k，d＝v_k，d+r*c*(x_best，d-x_k，d) And x_k，d＝v_k，d+x_k，dUpdating the control parameter Δ P_up、ΔT_upThe bird nest position;

a7, reevaluating Δ P_up、ΔT_upAnd updating the global optimal solution;

a8, judging whether an iteration condition is met, if so, outputting an optimal solution, and otherwise, returning to the step A2 to repeat iteration updating calculation again;

a9, optimal Delta P to be output_up、ΔT_upAnd the wind speed, the wind power ratio and the load disturbance quantity under the corresponding scene are combined in a pairing mode to generate a data set.

The characteristic extraction of the data set based on the noise reduction automatic encoder in the step B is carried out by the noise reduction automatic encoder;

the automatic encoder is composed of an encoder and a decoder as shown in fig. 4, wherein the encoder and the decoder are provided with hidden layers, the input is converted into potential representation in the hidden layers through the encoder, then the internal representation is converted into output through the decoder, and the output is equivalent to reconstruction of the input and is as close to the input as possible;

selecting an input sample set X consisting of N sets of samples X₁、x₂、…、x_nThe method comprises the following steps of (1) forming a sample group, wherein N is the number of the 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 the set of hidden layer feature vectors is composed of N sets of feature vectors H₁、h₂、…、h_mWherein m is the number of vectors in each group of eigenvectors, and the coding relationship between X and H is H ═ s_f(WX + B) (10), wherein W is a weight matrix of the input layer and the hidden layer; b is an inputLayer and hidden layer threshold matrices; s_fFor the neuron activation function of the encoder, a sigmoid function is usually used, which has good feature discrimination,

z is an input vector;

the decoder is the inverse operation of the encoder, the characteristic vector of the hidden layer is used as an input vector, Y is set as an output vector set, N groups are shared, the dimension is N, and the expression of the decoder is

Y＝s_g(W ' X + B ') (9), wherein W ' is a weight matrix of the hidden layer and the output layer; b' is a threshold matrix of a hidden layer and an output layer; s_gActivating a function for a neuron of a decoder;

the automatic encoder achieves the purpose of feature learning by minimizing the reconstruction error between the output vector and the input vector, and the gradient descent algorithm is used for continuously adjusting the network weight and the threshold value to reduce the reconstruction error. However, the learning of the auto-encoder may simply retain the original input data information and does not ensure that an effective representation of the feature information is obtained;

the noise reduction auto-encoder architecture is an improvement of the auto-encoder, as shown in fig. 5, and on the basis of the auto-encoder, by adding noise to a part of the input values, it is trained to recover the original noise-free input.

The basic steps of extracting the features of the data set based on the noise reduction automatic encoder in the step B are as follows:

b1, carrying out normalization processing on the data set obtained in the time domain simulation by using a parasitic-predation algorithm;

b2, dividing a training set and a verification set, taking 80% of original data as the training set, and taking the remaining 20% as test data to generate training and test data;

b3, greedy unsupervised pre-training layer by layer, taking the data added with noise as input data, and randomly selecting part of neurons to temporarily stop working by adopting a Dropout technology;

b4, fine-tuning the weight and the threshold value of the network by adopting a top-down small-batch Adagarad optimization method for the initialized network parameters until the iteration number reaches a set value.

C, learning data characteristics based on the deep neural network to generate the optimal gradual inertia control scheme of wind power frequency modulation on the basis of the deep neural network;

the deep neural network is a neural network with a plurality of hidden layers, also called a deep feedforward network; according to the position division of different layers, the neural network layer inside the DNN can be divided into: the input layer, the hidden layer and the output layer, wherein the first layer is the input layer, the last layer is the output layer, and the middle layers are all the hidden layers; the layers are all connected, namely any neuron at the ith layer is necessarily connected with any neuron at the (i + 1) th layer, and the basic structure is shown in figure 6; the complex DNN is from the local model, namely a linear relation as the perceptron

Plus an activation function sigma (z).

The basic steps of learning the data characteristics based on the deep neural network in the step C to generate the optimal wind power frequency modulation gradual inertia control scheme are as follows:

c1, taking the features extracted by the noise reduction automatic encoder as the input of the deep neural network;

c2, dividing a training set and a verification set, taking 80% of original data as the training set, and taking the remaining 20% as test data to generate training and test data;

c3, constructing a deep neural network model, and selecting the number of hidden layer layers and the number of neurons of each hidden layer;

c4, taking MSE as an evaluation index, judging whether the MSE is not reduced any more, finishing training if the MSE is not reduced any more, and returning to C3 if the MSE is not reduced any more;

and C5, storing the trained model, and directly using the model to generate the optimal gradual inertia control scheme of the wind power frequency modulation when encountering a load disturbance event next time.

Example testing:

taking the IEEE 9 node system as an example test system, the wind turbine model is connected to a line L3 in the system, as shown in FIG. 7; the studies involved were all obtained in a simulated operation of the wind turbine at sub-rated power;

in order to meet the requirements of using the method in different scenes, the influences of wind speed, wind power ratio and load disturbance quantity on the frequency characteristics of the system are considered, and the SIC is subjected to parameter setting from the three aspects.

In this example: setting the wind speed to be 4m/s to 10m/s, and increasing by 1m/s in each case; setting the wind power percentage from 5% to 60%, and increasing by 5% in each case; the load disturbance amount was set from 1.005 to 1.25, in increments of 0.05 in each case; in total 4200 cases.

The schematic diagram of the optimal SIC parameter combination prediction result is shown in figure 8, and the effect comparison with other frequency modulation control methods is added in a typical scene; the scene in the system is 5m/s of wind speed, 25% of wind power proportion and 0.035p.u.

In this scenario, the frequency control effect is shown in fig. 9 in comparison with other SIC control and frequency control without wind power participation.

Claims (8)

1. A wind power frequency modulation gradual inertia control method based on a deep neural network comprises the following steps:

A. obtaining wind power frequency modulation optimal gradual inertia control parameters in time domain simulation by using a parasitic-predation algorithm and generating a data set;

B. b, extracting the characteristics of the data set generated in the step A based on a noise reduction automatic encoder;

C. and C, learning the features extracted in the step B based on a deep neural network and generating an optimal gradual inertia control scheme of wind power frequency modulation.

2. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 1, wherein: the optimal gradual inertia control parameter of the wind power frequency modulation is used when a gradual inertia control strategy is used in the wind power frequency modulation;

the step-by-Step Inertial Control (SIC) is a control strategy for wind power to participate in grid frequency adjustment, and comprises two main stages: a short-time over-sending stage and a rotating speed recovery stage.

3. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 2, wherein the specific process of the wind power participating in the frequency modulation of the power grid is as follows:

1) the fan normally works at the point A on the MPPT curve, and the output electromagnetic power is P_w0The rotor speed is omega₀Output power P_w=P_MPPT＝P_w0；

At the time of T0, the power imbalance of the power system suddenly occurs to cause the frequency to drop, the fan monitors that the frequency drops to exceed the dead zone limit value, then the mode is switched to the gradual inertia control mode, and the output power delta P is immediately increased_upOutput power P_w＝P_w0+ΔP_upThe output power is increased from the point A to the point B, and a short-time over-transmitting stage is entered;

2) in the short-time over-sending stage, the output power of the fan is within a period of time delta T_upInternally held constant, i.e. output power always being P_w0+ΔP_upFrom point B to point C;

at the stage, the electromagnetic power output by the fan is larger than the mechanical power, the rotor is decelerated immediately, and the rotor motion equation of the fan is 2H_w

H_wIs the fan inertia constant;

duration of short-time superhair phase Δ T_upThe value of (A) is to ensure that the rotor speed is at T_offDoes not reach the minimum rotating speed omega at the moment_min；

3)T_offAt the moment, the short-time over-sending stage is ended, and the output power of the fan is suddenly reduced by delta P_offAt this time

Rotor speed of omega_offThe output power is suddenly reduced from the point C to the point D and enters a recovery stage;

4) the fan enters a rotating speed recovery stage, the fan operates in a maximum power point tracking mode at the moment, and the output power of the fan is

The generated system power shortage is delta P_off＝-P_w0+ΔP_off-P_wAt this stage, the mechanical power of the fan is greater than the electromagnetic power, the rotor starts to accelerate and passes through delta T_rAt recovery time of T_endThe rotor speed at the moment is controlled by omega_offRestore to omega₀The electromagnetic power output by the fan is recovered to P along with the increase of the rotating speed of the rotor_w0The output power slowly returns from point D to point a along the MPPT trajectory.

4. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 1, wherein the basic steps of obtaining the optimal gradual inertia control parameter of wind power frequency modulation in time domain simulation by using a parasitic-predation algorithm and generating a data set in the step A are as follows:

a1, initializing two control parameters delta P_up、ΔT_upPopulation and algorithm parameters of;

a2, evaluation parameter Δ P_up、ΔT_upDetermining an optimal parameter solution and an optimal parameter bird nest;

a3, calculating parameter delta P_up、ΔT_upThe corresponding number of cat groups, crow groups and bird nests;

a4, entering a nesting stage according to a formula

And

updating the control parameter Δ P_up、ΔT_upThe bird nest position;

a5, entering a parasitic stage according to the formula

S_G＝(X_r2-X_r3) Rand and k ═ rand [0, 1 ]]＞P_aUpdating the control parameter Δ P_up、ΔT_upThe bird nest position;

a6, entering the predation stage according to the formula v_k，d＝v_k，d+r*c*(x_best，d-x_k，d) And x_k，d＝v_k，d+x_k，dUpdating the control parameter Δ P_up、ΔT_upThe bird nest position;

a7, reevaluating Δ P_up、ΔT_upAnd updating the global optimal solution;

a8, judging whether an iteration condition is met, if so, outputting an optimal solution, and otherwise, returning to the step A2 to repeat iteration updating calculation again;

a9, optimal Delta P to be output_up、ΔT_upAnd the wind speed, the wind power ratio and the load disturbance quantity under the corresponding scene are combined in a pairing mode to generate a data set.

5. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 1, wherein: the characteristic extraction of the data set based on the noise reduction automatic encoder in the step B is carried out by the noise reduction automatic encoder;

the Automatic Encoder (AE) consists of an encoder and a decoder, wherein the encoder and the decoder are provided with hidden layers, the encoder converts input into potential tokens in the hidden layers, and then the decoder converts internal tokens into output, and the output is equivalent to reconstruction of the input and needs to be as close to the input as possible;

selecting an input sample set X consisting of N sets of samples X₁、x₂、…、x_nThe method comprises the following steps of (1) forming a sample group, wherein N is the number of the sample groups, and N is the number of samples in each group of samples; setting the hidden layer feature vector set as H from N groups of feature vectorsAmount h₁、h₂、…、h_mWherein m is the number of vectors in each group of eigenvectors, and the coding relationship between X and H is H ═ s_f(WX + B) (10), wherein W is a weight matrix of the input layer and the hidden layer; b is an input layer and hidden layer threshold matrix; s_fFor the neuron activation function of the encoder, a sigmoid function is usually used, which has good feature discrimination,

z is an input vector;

the decoder is the inverse operation of the encoder, and takes the characteristic vector of the hidden layer as the input vector, and sets Y as the output vector set, and has N groups in total, and the dimension is N, then the expression of the decoder is Y ═ s_g(W ' X + B ') (9), wherein W ' is a weight matrix of the hidden layer and the output layer; b' is a threshold matrix of a hidden layer and an output layer; s_gActivating a function for a neuron of a decoder;

the noise reduction auto-encoder (DAE) is an improvement on an auto-encoder that is trained to recover the original noise-free input by adding noise to a portion of the input values.

6. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 1, wherein the basic steps of extracting the features of the data set based on the noise reduction automatic encoder in the step B are as follows:

b1, carrying out normalization processing on the data set obtained in the time domain simulation by using a parasitic-predation algorithm;

b2, dividing a training set and a verification set, taking 80% of original data as the training set, and taking the remaining 20% as test data to generate training and test data;

b3, greedy unsupervised pre-training layer by layer, taking the data added with noise as input data, and randomly selecting part of neurons to temporarily stop working by adopting a Dropout technology;

b4, fine-tuning the weight and the threshold value of the network by adopting a top-down small-batch Adagarad optimization method for the initialized network parameters until the iteration number reaches a set value.

7. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 1, wherein: c, learning data characteristics based on the deep neural network to generate the optimal gradual inertia control scheme of wind power frequency modulation on the basis of the deep neural network;

the Deep Neural Network (DNN) is a neural network having a plurality of hidden layers, also referred to as a deep feed-forward network (DFN); according to the position division of different layers, the neural network layer inside the DNN can be divided into: the input layer, the hidden layer and the output layer, wherein the first layer is the input layer, the last layer is the output layer, and the middle layers are all the hidden layers; the layers are all connected, namely any neuron on the ith layer is connected with any neuron on the (i + 1) th layer; the complex DNN is from the local model, namely a linear relation as the perceptron

Plus an activation function sigma (z).

8. The wind power frequency modulation gradual inertia control method based on the deep neural network as claimed in claim 1, wherein the basic steps of learning the data characteristics based on the deep neural network in the step C to generate the optimal wind power frequency modulation gradual inertia control scheme are as follows:

c1, taking the features extracted by the noise reduction automatic encoder as the input of the deep neural network;

c2, dividing a training set and a verification set, taking 80% of original data as the training set, and taking the remaining 20% as test data to generate training and test data;

c3, constructing a deep neural network model, and selecting the number of hidden layer layers and the number of neurons of each hidden layer;

c4, taking MSE as an evaluation index, judging whether the MSE is not reduced any more, finishing training if the MSE is not reduced any more, and returning to C3 if the MSE is not reduced any more;

and C5, storing the trained model, and directly using the model to generate the optimal gradual inertia control scheme of the wind power frequency modulation when encountering a load disturbance event next time.

Images