A method and device for reconstructing abnormal data of a wind speed sensor

Through the variational modal decomposition and adaptive noise complete ensemble empirical modal decomposition algorithm combined with the Crazy Butterfly algorithm to train the LSTM neural network, the problem of unsatisfactory reconstruction of the abnormal data of the wind speed sensor is solved, and the efficiency improvement of the wind farm is achieved.

CN118839117BActive Publication Date: 2024-11-22HUNAN UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411311197.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-11-22
Estimated Expiration
2044-09-20

AI Technical Summary

Technical Problem

At this stage, the reconstruction of wind speed sensor abnormal data through artificial intelligence models is not ideal, which affects the stable operation of the wind turbine.

Method used

Variable mode decomposition and improved adaptive noise fully ensemble empirical mode decomposition algorithm decompose historical wind speed sequences, and combine the crazy butterfly algorithm based on adaptive perturbation to train the LSTM neural network to realize the reconstruction of wind speed sequences.

Benefits of technology

The efficiency of the wind farm is improved, and the accurate reconstruction of the abnormal data of wind speed sensors in actual projects is achieved, which solves the problem of unsatisfactory reconstruction of artificial intelligence models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118839117B_ABST
    Figure CN118839117B_ABST
Patent Text Reader

Abstract

The present application provides a method and device for reconstructing abnormal data of a wind speed sensor. The method decomposes a historical wind speed sequence once through variational mode decomposition, and then decomposes the decomposed residual components twice through an improved adaptive noise complete set empirical mode decomposition algorithm. Subsequently, the decomposed historical wind speed subsequences obtained once and the secondary decomposition historical wind speed subsequences obtained are respectively input into an LSTM neural network, and trained in combination with a crazy butterfly algorithm based on adaptive perturbation to obtain a first target LSTM neural network and a second target LSTM neural network. Finally, the wind speed sequence to be measured is reconstructed through the first target LSTM neural network and the second target LSTM neural network to obtain a corresponding wind speed sequence reconstruction result, thereby achieving an accurate reconstruction effect of abnormal data of wind speed sensors in actual projects, thereby improving the efficiency of wind farms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to a method and a device for reconstructing abnormal data of a wind speed sensor. Background Art

[0002] The randomness and instability of wind speed will affect the stability of wind turbine operation, thus seriously restricting the development of wind power. In this case, the reconstruction of abnormal data of wind speed sensors is particularly urgent.

[0003] The existing methods for reconstructing abnormal data of wind speed sensors can basically be divided into physical models, statistical models and artificial intelligence models. The physical model uses physical parameters such as temperature, humidity, density and terrain information to solve equations. These equations require a lot of calculations and are complicated to solve. The statistical model reconstructs the abnormal data of wind speed sensors by analyzing a large amount of historical data to establish a statistical model. Artificial intelligence models are widely used in many engineering fields due to their associative and learning capabilities, but the reconstruction effect of abnormal data of wind speed sensors through artificial intelligence models is not ideal at this stage. Summary of the invention

[0004] The present invention provides a method and a device for reconstructing abnormal data of a wind speed sensor, which are used to reconstruct abnormal data of a wind speed sensor in an actual project, and solve the technical problem that the reconstruction effect of abnormal data of a wind speed sensor is not ideal when an artificial intelligence model is used at the current stage.

[0005] In a first aspect, the present invention provides a method for reconstructing abnormal data of a wind speed sensor, comprising: step S1, performing variational mode decomposition on the acquired historical wind speed sequence to obtain multiple first-decomposition historical wind speed subsequences; step S2, decomposing the residual components of the first-decomposition historical wind speed subsequence through an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a second-decomposition historical wind speed subsequence; step S3, inputting the first-decomposition historical wind speed subsequence into an initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence; step S4, using a crazy butterfly algorithm based on adaptive perturbation, combining the historical wind speed sequence corresponding to the first-decomposition historical wind speed subsequence and the first predicted wind speed sequence, training the first LSTM neural network to obtain a first predicted wind speed sequence. A target LSTM neural network; step S5, inputting the secondary decomposition historical wind speed subsequence into the initially constructed second LSTM neural network to obtain a corresponding second predicted wind speed sequence; step S6, using the crazy butterfly algorithm based on adaptive perturbation, combined with the historical wind speed sequence corresponding to the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence, to train the second LSTM neural network to obtain a second target LSTM neural network; step S7, inputting the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network respectively, to obtain a corresponding first reconstructed wind speed sequence and a second reconstructed wind speed sequence; step S8, superimposing the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

[0006] Optionally, the step S4 includes: step S41, determining a first training error based on the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence; step S42, using the crazy butterfly algorithm based on adaptive perturbation to optimize the network parameters of the first LSTM neural network based on the first training error until the first training termination condition is met to obtain the first optimal network parameters; step S43, using the first optimal network parameters to generate the first target LSTM neural network.

[0007] Optionally, the step S6 includes: step S61, determining a second training error based on the historical wind speed sequence corresponding to the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence; step S62, using the crazy butterfly algorithm based on adaptive perturbation to optimize the network parameters of the second LSTM neural network based on the second training error until the second training termination condition is met to obtain the second optimal network parameters; step S63, using the second optimal network parameters to generate the second target LSTM neural network.

[0008] Optionally, the step S42 includes: step S421, determining a first search method for the butterfly individual based on a random number assigned to the butterfly individual and a size relationship between the preset first switching probability; step S422, updating the butterfly individual position through the first search method in combination with a first fitness value corresponding to the butterfly individual; the first fitness value is determined based on the first training error; step S423, determining whether the first training termination condition is met; if not, returning to step S41; if so, outputting the first optimal network parameters.

[0009] Optionally, the step S62 includes: step S621, determining a second search method for the butterfly individual based on a random number assigned to the butterfly individual and a size relationship between the preset second switching probability; step S622, updating the butterfly individual position through the second search method in combination with a second fitness value corresponding to the butterfly individual; the second fitness value is determined based on the second training error; step S623, determining whether the second training termination condition is met; if not, returning to step S61; if so, outputting the second optimal network parameters.

[0010] Optionally, the first search method includes local search and global search; the step S422 includes: step S4221, determining the first nectar source position according to the size relationship of the first fitness value, and then determining the corresponding first fragrance value; step S4222, when the first search method is the local search, updating the individual position of the butterfly according to a preset local search expression; step S4223, when the first search method is the global search, based on the first nectar source position, combined with a preset global search expression, update the individual position of the butterfly.

[0011] Optionally, the second search method includes local search and global search; the step S622 includes: step S6221, determining the second nectar source position according to the size relationship of the second fitness value, and then determining the corresponding second fragrance value; step S6222, when the second search method is the local search, updating the individual position of the butterfly according to the local search expression; step S6223, when the second search method is the global search, based on the second nectar source position, combined with the global search expression, update the individual position of the butterfly.

[0012] In a second aspect, the present invention provides a device for reconstructing abnormal data of a wind speed sensor, comprising: a primary decomposition module, used to perform variational mode decomposition on the acquired historical wind speed sequence to obtain multiple primary decomposition historical wind speed subsequences; a secondary decomposition module, used to decompose the residual components of the primary decomposition historical wind speed subsequence through an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a secondary decomposition historical wind speed subsequence; a first input module, used to input the primary decomposition historical wind speed subsequence into an initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence; a first training module, used to train the first LSTM neural network using a crazy butterfly algorithm based on adaptive perturbation, in combination with the historical wind speed sequence corresponding to the primary decomposition historical wind speed subsequence and the first predicted wind speed sequence, to obtain to the first target LSTM neural network; a second input module, used to input the secondary decomposition historical wind speed subsequence into the initially constructed second LSTM neural network, to obtain the corresponding second predicted wind speed sequence; a second training module, used to use the crazy butterfly algorithm based on adaptive perturbation, combined with the historical wind speed sequence corresponding to the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence, to train the second LSTM neural network, to obtain the second target LSTM neural network; a prediction module, used to input the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network, respectively, to obtain the corresponding first reconstructed wind speed sequence and the second reconstructed wind speed sequence; a reconstruction module, used to superimpose the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

[0013] In a third aspect, the present invention provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in the method provided in the first aspect are performed.

[0014] In a fourth aspect, the present invention provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, runs the steps of the method provided in the first aspect.

[0015] It can be seen from the above technical solutions that the present invention has the following advantages:

[0016] The present invention provides a method and device for reconstructing abnormal data of a wind speed sensor. The method decomposes a historical wind speed sequence once through variational mode decomposition, and then decomposes the residual components after decomposition twice through an improved adaptive noise complete set empirical mode decomposition algorithm. Subsequently, the decomposed historical wind speed subsequence and the secondary decomposition historical wind speed subsequence obtained by decomposition are respectively input into an LSTM neural network, and trained in combination with a crazy butterfly algorithm based on adaptive perturbation to obtain a first target LSTM neural network and a second target LSTM neural network. Finally, the wind speed of the wind speed sequence to be measured is reconstructed through the first target LSTM neural network and the second target LSTM neural network to obtain a corresponding wind speed sequence reconstruction result, thereby achieving an accurate reconstruction effect of abnormal data of wind speed sensors in actual projects, thereby improving the efficiency of wind farms, and solving the technical problem that the reconstruction effect of abnormal data of wind speed sensors is not ideal when an artificial intelligence model is used to reconstruct abnormal data of wind speed sensors at this stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0018] Figure 1 A flowchart of a first embodiment of a method for reconstructing abnormal data of a wind speed sensor according to the present invention;

[0019] Figure 2 This is a flowchart of a second embodiment of a method for reconstructing abnormal data of a wind speed sensor according to the present invention;

[0020] Figure 3 A flowchart of the steps of constructing a sensor abnormal data reconstruction model based on VMD, ICEEMDAN and CIBOA-LSTM in Example 1 of a method for reconstructing abnormal data of a wind speed sensor of the present invention;

[0021] Figure 4 Schematic diagram of the reconstruction result of one-step prediction of VMD-ICEEMDAN-GA-LSTM in Site1;

[0022] Figure 5 Schematic diagram of the reconstruction result of one-step prediction of VMD-ICEEMDAN-PSO-LSTM in Site1;

[0023] Figure 6 Schematic diagram of the reconstruction result of one-step prediction of VMD-ICEEMDAN-CIBOA-LSTM in Site1;

[0024] Figure 7 Schematic diagram of the reconstruction results of VMD-ICEEMDAN-GA-LSTM two-step prediction in Site2;

[0025] Figure 8 Schematic diagram of the reconstruction results of VMD-ICEEMDAN-PSO-LSTM two-step prediction in Site2;

[0026] Fig. 9 Schematic diagram of the reconstruction results of VMD-ICEEMDAN-CIBOA-LSTM two-step prediction in Site2;

[0027] Fig.10 Schematic diagram of the reconstruction results of BP two-step prediction in Site1;

[0028] Fig.11 This is a schematic diagram of the reconstruction results of LSTM two-step prediction in Site1;

[0029] Fig.12 This is a schematic diagram of the reconstruction results of the two-step prediction of BOA-LSTM in Site1;

[0030] Fig.13 This is a schematic diagram of the reconstruction results of the two-step prediction of CIBOA-LSTM in Site1;

[0031] Fig.14 The present invention is a structural block diagram of an embodiment of a device for reconstructing abnormal data of a wind speed sensor according to the present invention. DETAILED DESCRIPTION

[0032] The embodiments of the present invention provide a method and a device for reconstructing abnormal data of a wind speed sensor, which are used to reconstruct abnormal data of a wind speed sensor in an actual project.

[0033] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Example 1: Please refer to Figure 1 , Figure 1 This is a flow chart of a first embodiment of a method for reconstructing abnormal data of a wind speed sensor according to the present invention. The method includes:

[0035] Step S1, performing variational mode decomposition on the acquired historical wind speed sequence to obtain a plurality of primary decomposition historical wind speed subsequences;

[0036] It should be noted that variational mode decomposition (VMD) is a signal decomposition estimation method. In the process of obtaining decomposed components, this method determines the frequency center and bandwidth of each component by iteratively searching for the optimal solution of the variational model, so as to adaptively realize the frequency domain decomposition of the signal and the effective separation of each component.

[0037] Step S2, decomposing the residual component of the primary decomposition historical wind speed subsequence by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a secondary decomposition historical wind speed subsequence;

[0038] It should be noted that the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is an improved algorithm based on empirical mode decomposition (EMD). It solves the frequency aliasing and pseudo-modal problems existing in the EMD decomposition method - by adding white noise, the continuity of frequencies between adjacent scales is achieved, thereby reducing the impact of mixing.

[0039] Step S3, inputting the once-decomposed historical wind speed subsequence into the initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence;

[0040] It should be noted that the Long Short-Term Memory (LSTM) neural network is an improved version of the Recurrent Neural Network (RNN). In order to solve the problem that RNN is not suitable for long sequence information, in addition to the existing input and output gates of RNN, LSTM also sets a forget gate to control memory transfer and update. The forget gate can prevent the movement of meaningless information in the past while retaining useful information. This measure optimizes the long-term memory performance of RNN and optimizes the performance of the algorithm.

[0041] Specifically, the computational expressions involved in the LSTM neural network include:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] in, is the time step The output of the next forget gate, is the weight, is the sigmoid function, is the current time step Input, The previous time step The intermediate state is the bias, is the time step The input of the lower input-output gate, is the time step The input of the lower input node, is the hyperbolic tangent function, is the time step The output of the lower input-output gate, is the time step The next output is the intermediate state of the face change, is the time step The unit status, is the time step The unit status.

[0049] The embodiment of the present invention inputs the once-decomposed historical wind speed subsequence into the initially constructed first LSTM neural network to predict the corresponding first predicted wind speed sequence.

[0050] Step S4, using a crazy butterfly algorithm based on adaptive disturbance, combining the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence, training the first LSTM neural network to obtain a first target LSTM neural network;

[0051] It should be noted that the Crazy Invasive Butterfly Optimization Algorithm (CIBOA) based on adaptive perturbation is an improved butterfly optimization algorithm. The butterfly optimization algorithm is an innovative intelligent optimization algorithm inspired by the foraging behavior of butterflies. Butterflies determine the potential direction of food sources or mating partners by sensing and analyzing the smell in the air. Each butterfly releases a certain intensity of fragrance, which can spread far away and can be sensed by other butterflies. The intensity of this fragrance is related to the fitness of the butterfly, so when the butterfly moves, its fitness will also change. In the algorithm, when a butterfly senses the fragrance released by other butterflies, it will move towards the target, which corresponds to the global search stage in the algorithm. When the butterfly cannot sense the surrounding smell, it will move randomly, and this stage is called local search.

[0052] The self-cognitive flight part in the butterfly optimization algorithm reflects that the rear butterfly traces back the position of the previous butterfly. This position update mechanism is similar to the cognitive part in the particle swarm optimization algorithm. In the particle swarm optimization algorithm, the inertia weight is introduced to adjust the inertia degree of the particle at the current position, thereby affecting the search behavior of the particle. However, if the inertia weight is set too large or too small, it may cause the algorithm to fall into a local optimum, thereby reducing the efficiency of the algorithm. In order to better balance the local and global search capabilities of the algorithm, the embodiment of the present invention introduces a linearly decreasing inertia weight to update the position of the butterfly.

[0053] At the same time, in the global search phase of the butterfly optimization algorithm, all butterfly individuals fly towards the same global optimal position. The disadvantage of this operation is that it is easy to cause the algorithm to fall into a local optimum, especially when dealing with complex multi-peak functions, which may cause problems of premature convergence and insufficient precision. In order to solve this problem, the embodiment of the present invention introduces a multi-stage perturbation strategy to update the global optimal nectar position. Specifically, the global optimal nectar position is perturbed according to a normal random distribution with adjustable variance to generate a new global optimal nectar position.

[0054] In addition, the concentration of the scent released by butterflies during flight will not remain constant, but will fluctuate with the change of distance, thus affecting the behavior of the population. In order to describe this phenomenon, the embodiment of the present invention introduces the concept of "crazy factor", the core idea of ​​which is to model the behavior of butterflies through crazy variables. In the specific implementation, the crazy factor is introduced into the position information update formula of the butterfly optimization algorithm to ensure that the butterfly has a predetermined crazy probability, thereby maintaining the diversity of the population.

[0055] In an optional embodiment, the step S4 includes:

[0056] Step S41, determining a first training error according to the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence;

[0057] Step S42, using the crazy butterfly algorithm based on adaptive perturbation to optimize the network parameters of the first LSTM neural network based on the first training error until the first training termination condition is met, thereby obtaining the first optimal network parameters;

[0058] Step S43: using the first optimal network parameters to generate the first target LSTM neural network.

[0059] In an embodiment of the present invention, a first training error is determined based on the gap between an actual historical wind speed sequence of a once-decomposed historical wind speed subsequence and a first predicted wind speed sequence, and then a crazy butterfly algorithm based on adaptive perturbation is used in combination with the first training error to optimize the network parameters of the first LSTM neural network until a preset termination condition is reached, thereby finding the first optimal network parameters, and applying the first optimal network parameters to generate a first target LSTM neural network.

[0060] Step S5, inputting the secondary decomposed historical wind speed subsequence into the initially constructed second LSTM neural network to obtain a corresponding second predicted wind speed sequence;

[0061] The embodiment of the present invention inputs the secondary decomposed historical wind speed subsequence into the initially constructed second LSTM neural network to predict the corresponding second predicted wind speed sequence.

[0062] Step S6, using the crazy butterfly algorithm based on adaptive disturbance, combined with the historical wind speed sequence corresponding to the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence, to train the second LSTM neural network to obtain a second target LSTM neural network;

[0063] In an optional embodiment, step S6 includes:

[0064] Step S61, determining a second training error according to the historical wind speed sequence corresponding to the secondary decomposed historical wind speed subsequence and the second predicted wind speed sequence;

[0065] Step S62, using the crazy butterfly algorithm based on adaptive perturbation to optimize the network parameters of the second LSTM neural network based on the second training error until the second training termination condition is met, thereby obtaining second optimal network parameters;

[0066] Step S63: Use the second optimal network parameters to generate the second target LSTM neural network.

[0067] In an embodiment of the present invention, a second training error is determined based on the gap between the actual historical wind speed sequence of the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence, and then a crazy butterfly algorithm based on adaptive perturbation is used in combination with the second training error to optimize the network parameters of the second LSTM neural network until a preset termination condition is reached, thereby finding the second optimal network parameters, and applying the second optimal network parameters to generate a second target LSTM neural network.

[0068] Step S7, inputting the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network respectively to obtain a corresponding first reconstructed wind speed sequence and a second reconstructed wind speed sequence;

[0069] In an embodiment of the present invention, the wind speed sequence to be measured is respectively input into the first target LSTM neural network and the second target LSTM neural network to obtain the corresponding first reconstructed wind speed sequence and the second reconstructed wind speed sequence, in preparation for the final wind speed reconstruction result.

[0070] Step S8: superimpose the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

[0071] The embodiment of the present invention obtains the final wind speed sequence reconstruction result by superimposing the first reconstructed wind speed sequence and the second reconstructed wind speed sequence and comprehensively considering the prediction results based on the first target LSTM neural network and the second target LSTM neural network.

[0072] In summary, a method for reconstructing abnormal data of a wind speed sensor provided by an embodiment of the present invention decomposes a historical wind speed sequence once through variational mode decomposition, and then decomposes the decomposed residual components twice through an improved adaptive noise complete set empirical mode decomposition algorithm, and then inputs the decomposed historical wind speed subsequence and the secondary decomposition historical wind speed subsequence obtained by decomposition into the LSTM neural network respectively, and trains them in combination with the crazy butterfly algorithm based on adaptive perturbation to obtain a first target LSTM neural network and a second target LSTM neural network, and finally reconstructs the wind speed of the measured wind speed sequence through the first target LSTM neural network and the second target LSTM neural network to obtain the corresponding wind speed sequence reconstruction result, thereby achieving accurate reconstruction effect of abnormal data of wind speed sensors in actual projects, thereby improving the efficiency of wind farms, and solving the technical problem of unsatisfactory reconstruction effect when abnormal data of wind speed sensors are reconstructed through artificial intelligence models at this stage.

[0073] Example 2: Please refer to Figure 2 , Figure 2 The flowchart of Embodiment 2 of a method for reconstructing abnormal data of a wind speed sensor according to the present invention includes:

[0074] S201, performing variational mode decomposition on the acquired historical wind speed sequence to obtain a plurality of primary decomposition historical wind speed subsequences;

[0075] In the embodiment of the present invention, variational modal decomposition decomposes the input historical wind speed sequence into multiple modal functions, that is, multiple primary decomposition historical wind speed subsequences.

[0076] In the specific implementation, variational mode decomposition requires that the finite bandwidth of each mode is around its center frequency so that the sum of all modal estimation bandwidths is minimized. The variational optimization model of variational mode decomposition is:

[0077] ;

[0078] in, is the historical wind speed series, is the number of modal functions, for The unit impulse function at time, is an imaginary unit, is the convolution operator, For the function The partial guide, is the center frequency set, To decompose the historical wind speed subsequence once, Used to describe the constraints in the optimization problem of this model.

[0079] At the same time, in order to transform the constrained optimization problem into an unconstrained optimization problem, the quadratic penalty function term and Lagrange multiplier are introduced. The expression of the augmented Lagrangian function is:

[0080] ;

[0081] in, is the augmented Lagrangian function, λ is the Lagrangian multiplier, is the penalty parameter.

[0082] In the specific implementation, the alternating inverse multiplier method is used to solve the optimization problem, that is, by alternating optimization and , to solve the unconstrained optimization problem. The specific update formula is:

[0083] ;

[0084] ;

[0085] in, is the historical wind speed sequence after Fourier transformation, is the mode function after Fourier transformation, is the Lagrange multiplier after Fourier transformation, For the The center frequency at the iteration is No. modal components, for The Fourier transform of .

[0086] S202, decomposing the residual component of the primary decomposition historical wind speed subsequence by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a secondary decomposition historical wind speed subsequence;

[0087] In the implementation of the present invention, the basic steps of the ICEEMDAN algorithm are as follows:

[0088] (1) Based on the original signal, a new sequence is constructed by decomposing the historical wind speed subsequence once. The formula is expressed as:

[0089] ;

[0090] in, is the original signal, is generated by the EMD algorithm The modal components, is the adjustment factor of the original noise intensity, To add to the original signal Group white noise, To add The sequence constructed after the group white noise.

[0091] (2) Determine the first set of residuals, expressed as:

[0092] ;

[0093] in, is the first set of residuals, is the local mean of the signal generated by the EMD algorithm, To find the average value for the whole.

[0094] (3) Calculate the first modal component, and its formula is expressed as:

[0095] ;

[0096] in, is the first modal component.

[0097] (4) On this basis, white noise is continued to be added, and the second set of residuals and modal vectors are calculated using local mean decomposition, which is expressed as follows:

[0098] ;

[0099] ;

[0100] in, is the second set of residuals, is the adjustment factor of the first group of noise intensity, is the second modal component.

[0101] (5) Similarly, determine The residuals and modal components of the group are combined to obtain all modal components and the worst residuals, which can be expressed as:

[0102] ;

[0103] ;

[0104] in, For the Group residuals, For the Group residuals, for The adjustment factor for the group noise intensity, For the Group modal components.

[0105] In addition, the formula for determining the noise intensity is:

[0106] ;

[0107] in, is the standard deviation, is the inverse of the signal-to-noise ratio between the first added noise and the analyzed signal, It is a sequence constructed by adding white noise to the original signal.

[0108] S203, inputting the once-decomposed historical wind speed subsequence into an initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence;

[0109] S204, determining a first training error according to the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence;

[0110] S205, determining a first search method for the butterfly individual based on a magnitude relationship between a random number assigned to the butterfly individual and a preset first switch probability;

[0111] The butterfly in the butterfly optimization algorithm determines the potential direction of food sources or mating partners by sensing and analyzing the smell in the air. The fragrance is determined by the physical intensity of the stimulus, and its calculation formula is:

[0112] ;

[0113] in, is the fragrance value, is the sensory factor, usually with a value of 0.01. is the stimulus intensity, which is related to the optimal fitness. is the power exponent, usually taken as 0.1.

[0114] In the initial stage of the algorithm, the position of each butterfly is randomly generated, and then after the fragrance is generated at the respective positions of the butterflies according to the above formula, the algorithm enters the global search and local search stages. In the global search process, each butterfly moves towards the current global optimal position, and the formula is expressed as:

[0115] ;

[0116] in, For the A butterfly in the The position vector in the iteration, For the A butterfly in the The position vector in the iteration, is a random number in [0, 1], is the current global optimal position, For the The scent value of a butterfly.

[0117] The position update formula of the local search process is:

[0118] ;

[0119] in, is a random number in [0, 1], and is randomly selected from the solution space and Butterfly.

[0120] S206, updating the position of the individual butterfly by using the first search method and combining the first fitness value corresponding to the individual butterfly; the first fitness value is determined according to the first training error;

[0121] In an optional embodiment, the first search method includes local search and global search; through the first search method, the individual butterfly position is updated in combination with the first fitness value corresponding to the individual butterfly, including: determining the first nectar source position according to the size relationship of the first fitness value, and then determining the corresponding first fragrance value; when the first search method is the local search, the individual butterfly position is updated according to a preset local search expression; when the first search method is the global search, the individual butterfly position is updated based on the first nectar source position in combination with a preset global search expression.

[0122] S207, determining whether the first training termination condition is met; if not, returning to step S204; if so, outputting the first optimal network parameters.

[0123] S208, using the first optimal network parameters to generate the first target LSTM neural network;

[0124] S209, inputting the secondary decomposed historical wind speed subsequence into the initially constructed second LSTM neural network to obtain a corresponding second predicted wind speed sequence;

[0125] S210, determining a second training error according to the historical wind speed sequence corresponding to the secondary decomposed historical wind speed subsequence and the second predicted wind speed sequence;

[0126] S211, determining a second search mode for the butterfly individual based on a magnitude relationship between a random number assigned to the butterfly individual and a preset second switching probability;

[0127] In the foraging process of butterflies, both global and local searches occur. Therefore, the embodiment of the present invention switches between ordinary global search and intensive local search by setting the first switch probability and the second switch probability to 0.8. Each iteration uses a random number with a size interval of [0,1] assigned to the butterfly individual to compare with the switch probability to determine whether the search method is global search or local search.

[0128] S212, updating the individual butterfly position by combining the second fitness value corresponding to the individual butterfly through the second search method; the second fitness value is determined according to the second training error;

[0129] In an optional embodiment, the individual butterfly position is updated through the second search method in combination with the second fitness value corresponding to the individual butterfly, including: determining the second nectar source position according to the size relationship of the second fitness value, and then determining the corresponding second fragrance value; when the second search method is the local search, updating the individual butterfly position according to the local search expression; when the second search method is the global search, updating the individual butterfly position based on the second nectar source position in combination with the global search expression.

[0130] In order to better balance the local and global search capabilities of the algorithm, the embodiment of the present invention introduces a linearly decreasing inertia weight to update the position of the butterfly, which is expressed as follows:

[0131] ;

[0132] ;

[0133] in, is a random number in [0, 1], For the The adaptive inertia weight at the iteration, is the inertia weight at the end of the iteration, is the initial inertia weight, is the maximum number of iterations.

[0134] In an embodiment of the present invention, =0.9, = 0.4, the algorithm has the best performance. Therefore, as the iteration proceeds, the inertia weight decreases linearly from 0.9 to 0.4. The larger inertia weight in the initial stage of the iteration allows the algorithm to maintain a better search capability, while the smaller inertia weight in the later stage of the iteration helps the algorithm to develop better capabilities.

[0135] At the same time, the embodiment of the present invention introduces a multi-stage perturbation strategy, which perturbs the global optimal nectar position according to a normal random distribution with adjustable variance, thereby generating a new global optimal nectar position. The position update formula is expressed as:

[0136] ;

[0137] in, The global optimal nectar position after introducing the multi-stage perturbation strategy. is the uncertainty relative to the global optimal nectar position, which is a non-increasing function of the number of iterations, is the perturbation update rule.

[0138] The update formula is:

[0139] ;

[0140] in, , and is the radius parameter of the normal perturbation, and > > , and is the control parameter of radius variation, and < , is the maximum number of iterations.

[0141] In addition, the embodiment of the present invention introduces a madness factor into the position information update formula of the butterfly optimization algorithm to ensure that the butterfly has a predetermined madness probability, thereby maintaining the diversity of the population, wherein the position update formula after the madness factor is introduced is:

[0142] ;

[0143] in, is a random number in [0, 1], It is usually a small constant, indicating that the butterfly is less likely to move the nectar location during flight. and The definitions are:

[0144] ;

[0145] ;

[0146] in, For the set crazy probability, if The smaller the value, the random number There is a high probability that it will be greater than , the craziness factor will also be 0, which is what one actually expects.

[0147] S213, determining whether the second training termination condition is met; if not, returning to step S210; if so, outputting the second optimal network parameters;

[0148] S214, using the second optimal network parameters to generate the second target LSTM neural network;

[0149] The crazy butterfly algorithm based on adaptive perturbation used in the embodiment of the present invention introduces an adaptive inertia weight in its own cognitive flight part to balance the local and global search capabilities of the algorithm; a perturbation strategy is introduced at the global optimal position to prevent the algorithm from falling into the local optimum; a crazy operator is introduced at the nectar position to increase population diversity and obtain a better optimal solution.

[0150] In a specific implementation, the steps of determining the optimal network parameters based on the adaptive perturbation crazy butterfly algorithm include:

[0151] (1) Set the algorithm parameters and initialize the individual butterfly positions. Randomly generate an N×d-dimensional moment based on the upper and lower limits of the search space.

[0152] (2) Calculate the initial fitness value. Calculate the fitness values ​​of N butterflies based on the test function and apply them to the first LSTM neural network and the second LSTM neural network in the wind speed prediction problem.

[0153] (3) Arrange the fitness values ​​in ascending order. Select the butterfly position with the best fitness value as the initial nectar source position and calculate its corresponding fragrance size.

[0154] (4) According to the position update formula after the introduction of the crazy factor, combined with the preset first switch probability and second switch probability, the search method is determined and the position of each butterfly is updated accordingly.

[0155] (5) Apply the updated butterfly positions to the network parameters of the first LSTM neural network and the second LSTM neural network, calculate the fitness value of the network parameters corresponding to each butterfly, and update the optimal network parameter position.

[0156] (6) Repeat the iterative process from step (1) to step (5). If the set accuracy requirement or the specified maximum number of iterations is reached, the algorithm is terminated and the global optimal solution, i.e., the first optimal network parameters and the second optimal network parameters, is output.

[0157] S215, inputting the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network respectively to obtain a corresponding first reconstructed wind speed sequence and a second reconstructed wind speed sequence;

[0158] S216: Superimpose the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

[0159] Sensor failures mainly include fixed deviation failures, failure failures, and drift deviation failures. Since the wind speed sensor is the most important sensor for a wind turbine, a wind speed sensor failure will cause other sensors to become abnormal one after another, which is not conducive to the safe and stable operation of the wind turbine. Based on this, a method for reconstructing abnormal data of a wind speed sensor disclosed in an embodiment of the present invention is based on the above steps. When a wind speed sensor fails, the wind speed sequence reconstruction result is used to enable the wind speed sensor to continue working for a period of time until the staff replaces it, thereby avoiding the abnormal data generated by the wind speed sensor failure affecting other sensors. Thereby, the accurate reconstruction effect of the abnormal data of the wind speed sensor in the actual project is achieved, thereby improving the efficiency of the wind farm and solving the technical problem that the reconstruction effect of abnormal data of the wind speed sensor is not ideal when the artificial intelligence model is used at this stage.

[0160] To facilitate those skilled in the art to understand the construction process and beneficial effects of the present invention, a method for reconstructing abnormal data of a wind speed sensor of the present invention is exemplified below.

[0161] Example 1: This example builds sensor abnormal data reconstruction models based on VMD, ICEEMDAN and CIBOA-LSTM. The construction process is as follows: Figure 3 As shown, the specific steps include:

[0162] (1) The center frequency ratio method is used to determine the number K of VMD decomposition subsequences, and VMD is used to decompose the wind speed series into K subsequences.

[0163] VMD was used to decompose the original wind speed series. After repeated experiments, the penalty factor α was set to 2000 and the tolerance was set to 10 -6 , only the modal decomposition layer number K is preferred. The K value is determined by observing the center frequency, and the K value is preset from small to large. When the center frequency of the last layer of IMF components remains relatively stable, it can be considered that K has obtained the optimal value at this time. The VMD decomposition of the original wind speed sequence at different K values ​​produces the center frequency of the IMF components of each data set, as shown in the table of the center frequencies of each IMF component at different K values ​​in Table 1 below. It can be seen that for Site1 and Site2, the center frequency of the last layer of IMF components remains relatively stable after K>6, so K=7 is the optimal value.

[0164] Table 1 Central frequency of each IMF component under different K values

[0165]

[0166] (2) As previous methods usually discard the residual components after VMD decomposition and ignore the rich information carried in the residual components, the sequence decomposition is incomplete and the model accuracy is limited. Therefore, making full use of the rich information of the residual components after the first decomposition and carrying out secondary decomposition can effectively improve the prediction accuracy of the model. After obtaining the residual components after VMD processing of the wind speed sequence, the residual components are subjected to ICEEMDAN secondary decomposition. Among them, the parameter signal-to-noise ratio of ICEEMDAN is set to 0.2, and the maximum number of iterations is set to 1000.

[0167] (3) A CIBOA-LSTM model is established for all subsequences obtained after the first decomposition by VMD and the second decomposition by ICEEMDAN. CIBOA is used to optimize the initial learning rate, number of hidden layers and regularization parameters of the LSTM model for training and prediction.

[0168] (4) The prediction results of each subsequence are superimposed to obtain the final result of wind speed reconstruction.

[0169] Example 2: This example uses wind speed data from two wind farms in Chenzhou, Hunan, China as experimental data, with a time interval of 10 minutes. In each data set, we select 2,000 points as experimental data, of which the first 1,700 points are used as training sets and the remaining 300 points are used as test sets. The reconstruction system uses both single-step prediction and multi-step prediction.

[0170] In order to verify the prediction ability of the model obtained by the crazy butterfly algorithm based on adaptive perturbation involved in the reconstruction method of abnormal data of a wind speed sensor of the present invention, different intelligent optimization combination models are proposed, namely: VMD-ICEEMDAN-CIBOA-LSTM (i.e., a composite signal processing and prediction model combining variational mode decomposition, improved adaptive noise complete set empirical mode decomposition, crazy butterfly algorithm based on adaptive perturbation and long short-term memory neural network, i.e., the model involved in the above-mentioned embodiment and example one), VMD-ICEEMDAN-BOA-LSTM (i.e., a composite signal processing and prediction model combining variational mode decomposition, improved adaptive noise complete set empirical mode decomposition, crazy butterfly algorithm based on adaptive perturbation and long short-term memory neural network, i.e., the model involved in the above-mentioned embodiment and example one), VMD-ICEEMDAN-BOA-LSTM (i.e., a composite signal processing and prediction model combining variational mode decomposition, improved adaptive noise complete set empirical mode decomposition, butterfly optimization algorithm and long short-term memory neural network), The composite signal processing and prediction model of the wind speed sensor is composed of VMD-ICEEMDAN-PSO-LSTM (combining variational mode decomposition, improved adaptive noise complete set empirical mode decomposition, particle swarm optimization algorithm and long short-term memory neural network), VMD-ICEEMDAN-GA-LSTM (combining variational mode decomposition, improved adaptive noise complete set empirical mode decomposition, genetic algorithm and long short-term memory neural network), and the selected fault types are fixed deviation fault and failure fault. By merging the fault data into the normal 300 test data, the fault state of the wind speed sensor is simulated. Among them, the definition of fixed deviation fault is that the difference between the test value and the actual value is a roughly fixed amount. In this example, a deviation of 0.7m / s is added between 105 and 135 in the test set of Site1 to simulate a fixed deviation fault; a failure fault means that the sensor suddenly fails from the normal working state. In the test set of Site2, between 74 and 106, the test value drops to the specified value, and the wind speed gradually drops to 2m / s. The evaluation results are shown in the table of comparison results of different intelligent optimization combination models in Table 2 below: It can be seen that the model obtained by the crazy butterfly algorithm based on adaptive perturbation and its reconstruction method involved in this example are better than other models at any prediction step number. Specifically, for fixed deviation faults, when the prediction step size is 1, VMD-ICEEMDAN-CIBOA-LSTM produced the most satisfactory results at Site 1: RMSE=0.1973, MAE=0.1764, MAPE=2.2670. In terms of two-step prediction, the MAE values ​​of VMD-ICEEMDAN-GA-LSTM, VMD-ICEEMDAN-PSO-LSTM and VMD-ICEEMDAN- BOA -LSTM are 0.4276, 0.3752 and 0.3104, respectively. In addition, the MAE value of the proposed model is the smallest, MAE=0.2128. When the prediction step size is 3, the evaluation index value obtained by the proposed reconstruction model is the lowest as before, indicating that the proposed model is more accurate. Please participate Figure 4~Figure 6 , Figure 4 This is a schematic diagram of the reconstruction results of VMD-ICEEMDAN-GA-LSTM one-step prediction in Site1. Figure 5 This is a schematic diagram of the reconstruction results of VMD-ICEEMDAN-PSO-LSTM one-step prediction in Site1. Figure 6 This is a schematic diagram of the reconstruction result of one-step prediction of VMD-ICEEMDAN-CIBOA-LSTM in Site1. In Example 2 of the present invention, Figure 4~Figure 6 The results of reconstructing data for different horizontal coordinates (different test sample numbers) corresponding to one-step prediction in Site1 using different optimization methods are shown. Regardless of the number of prediction steps, the proposed model can obtain the best reconstruction result.

[0171] For the failure of Site 2, in the one-step prediction, the proposed reconstruction model has the best results in three indicators, namely MAE=0.1787, RMSE=0.2104 and MAPE=2.1095. Figure 7~Figure 9 , Figure 7 This is a schematic diagram of the reconstruction results of VMD-ICEEMDAN-GA-LSTM two-step prediction in Site2. Figure 8 This is a schematic diagram of the reconstruction results of VMD-ICEEMDAN-PSO-LSTM two-step prediction in Site2. Fig. 9 Schematic diagram of reconstruction results of two-step prediction of VMD-ICEEMDAN-CIBOA-LSTM in Site2. In Example 2 of the present invention, Figure 7~Figure 9 The figure shows the results of reconstructing data for different horizontal coordinates (different test sample numbers) corresponding to different optimization methods for two-step prediction in Site2. Specifically, the reconstruction accuracy of the relevant models from high to low is VMD-ICEEMDAN-BOA-LSTM, VMD-ICEEMDAN-PSO-LSTM and VMD-ICEEMDAN-GA-LSTM.

[0172] Table 2 Comparison results of different intelligent optimization combination models

[0173]

[0174] Example 3: This example uses reconstruction models of various combination technologies, including: Backpropagation Network (BP), LSTM and BOA-LSTM. Through comparative experiments, it is verified whether the basic reconstruction model CIBOA-LSTM proposed in the above embodiments and examples performs better than the other three reconstruction models. The final index value of the experiment is shown in the final index value table of the experiment in Table 3 below. By comparing the final index values ​​obtained in the experiment, the following conclusions can be drawn:

[0175] (1) In the one-step prediction of the fixed deviation fault of Site1, the CIBOA-LSTM method achieved the most satisfactory results: MAE=0.4958, RMSE=0.6352, MAPE=6.3176. Please participate Figure 10~Figure 13 , Fig.10 This is a schematic diagram of the reconstruction results of BP's two-step prediction in Site1. Fig.11 This is a schematic diagram of the reconstruction results of LSTM two-step prediction in Site1. Fig.12 This is a schematic diagram of the reconstruction results of BOA-LSTM two-step prediction in Site1. Fig.13 This is a schematic diagram of the reconstruction results of CIBOA-LSTM two-step prediction in Site1. Figure 10-13 The reconstruction results of the two-step prediction corresponding to different horizontal axes (different test sample numbers) in Site1 are shown. In the two-step prediction, compared with LSTM, the MAE of the BOA-LSTM model is reduced by 24.05%, the RMSE of the BOA-LSTM model is reduced by 18.03%, and the MAPE of the BOA-LSTM model is reduced by 26.23%. In the three-step prediction, the MAE of the BP model is 1.7435, RMSE is 2.1043, and MAPE is 23.7016.

[0176] (2) For failures, at Site 2, the CIBOA-LSTM model performs better than the other three methods in terms of the three indicators in the two-step prediction, namely, MAE=0.6366, RMSE=0.8375, and MAPE=8.1664. In the three-step prediction, except for the CIBOA-LSTM model, the BOA-LSTM model (MAPE=12.8894, RMSE=1.1123, and MAE=0.9007) is more accurate than the other models.

[0177] Table 3 Final index value table of the test

[0178]

[0179] It should be added that the mean absolute error calculation expression is:

[0180] ;

[0181] The root mean square error calculation expression is:

[0182] ;

[0183] The mean absolute percentage error calculation expression is:

[0184] ;

[0185] Among them, MAE is the mean absolute error, RMSE is the root mean square error, and MAPE is the mean absolute percentage error. is the total number of predictions, is the predicted value, is the tag value.

[0186] Example 3: Please refer to Fig.14 , Fig.14 The structural block diagram of an embodiment of a device for reconstructing abnormal data of a wind speed sensor according to the present invention comprises:

[0187] A primary decomposition module 301 is used to perform variational mode decomposition on the acquired historical wind speed sequence to obtain multiple primary decomposition historical wind speed subsequences;

[0188] A secondary decomposition module 302 is used to decompose the residual component of the primary decomposition historical wind speed subsequence by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a secondary decomposition historical wind speed subsequence;

[0189] A first input module 303 is used to input the once-decomposed historical wind speed subsequence into the initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence;

[0190] A first training module 304 is used to train the first LSTM neural network by using a crazy butterfly algorithm based on adaptive perturbation, combining the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence, to obtain a first target LSTM neural network;

[0191] The second input module 305 is used to input the secondary decomposed historical wind speed subsequence into the initially constructed second LSTM neural network to obtain a corresponding second predicted wind speed sequence;

[0192] A second training module 306 is used to train the second LSTM neural network by using the crazy butterfly algorithm based on adaptive perturbation, in combination with the historical wind speed sequence corresponding to the secondary decomposed historical wind speed subsequence and the second predicted wind speed sequence, to obtain a second target LSTM neural network;

[0193] A prediction module 307, used to input the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network respectively, to obtain a corresponding first reconstructed wind speed sequence and a second reconstructed wind speed sequence;

[0194] The reconstruction module 308 is used to superimpose the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

[0195] Embodiment 4: An embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a method for reconstructing abnormal data of a wind speed sensor in any embodiment.

[0196] An embodiment of the present invention further provides a computer storage medium on which a computer program is stored. When the computer program is executed by the processor, the steps of a method for reconstructing abnormal data of a wind speed sensor in any embodiment are implemented.

[0197] Embodiment 5: Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0198] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing abnormal data of a wind speed sensor, characterized in that: include: Step S1, performing variational mode decomposition on the acquired historical wind speed sequence to obtain a plurality of primary decomposition historical wind speed subsequences; Step S2, decomposing the residual component of the primary decomposition historical wind speed subsequence by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a secondary decomposition historical wind speed subsequence; Step S3, inputting the once-decomposed historical wind speed subsequence into the initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence; Step S4, using a crazy butterfly algorithm based on adaptive disturbance, combining the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence, training the first LSTM neural network to obtain a first target LSTM neural network; Step S5, inputting the secondary decomposed historical wind speed subsequence into the initially constructed second LSTM neural network to obtain a corresponding second predicted wind speed sequence; Step S6, using the crazy butterfly algorithm based on adaptive disturbance, combined with the historical wind speed sequence corresponding to the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence, to train the second LSTM neural network to obtain a second target LSTM neural network; Step S7, inputting the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network respectively to obtain a corresponding first reconstructed wind speed sequence and a second reconstructed wind speed sequence; Step S8: superimpose the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

2. The method for reconstructing abnormal data of a wind speed sensor according to claim 1, characterized in that: The step S4 comprises: Step S41, determining a first training error according to the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence; Step S42, using the crazy butterfly algorithm based on adaptive perturbation to optimize the network parameters of the first LSTM neural network based on the first training error until the first training termination condition is met, thereby obtaining the first optimal network parameters; Step S43: using the first optimal network parameters to generate the first target LSTM neural network.

3. The method for reconstructing abnormal data of a wind speed sensor according to claim 2, characterized in that: The step S6 comprises: Step S61, determining a second training error according to the historical wind speed sequence corresponding to the secondary decomposed historical wind speed subsequence and the second predicted wind speed sequence; Step S62, using the crazy butterfly algorithm based on adaptive perturbation to optimize the network parameters of the second LSTM neural network based on the second training error until the second training termination condition is met, thereby obtaining second optimal network parameters; Step S63: Use the second optimal network parameters to generate the second target LSTM neural network.

4. The method for reconstructing abnormal data of a wind speed sensor according to claim 2, characterized in that: The step S42 comprises: Step S421, determining a first search mode for the butterfly individual based on a random number assigned to the butterfly individual and a magnitude relationship between the random number assigned to the butterfly individual and a preset first switch probability; Step S422, updating the individual butterfly position by using the first search method and combining the first fitness value corresponding to the individual butterfly; the first fitness value is determined according to the first training error; Step S423, determine whether the first training termination condition is met; if not, return to step S41; if so, output the first optimal network parameters.

5. The method for reconstructing abnormal data of a wind speed sensor according to claim 3, characterized in that: The step S62 comprises: Step S621, determining a second search method for the butterfly individual based on a random number assigned to the butterfly individual and a magnitude relationship between the random number assigned to the butterfly individual and a preset second switch probability; Step S622, updating the individual butterfly position by combining the second fitness value corresponding to the individual butterfly through the second search mode; the second fitness value is determined according to the second training error; Step S623, determine whether the second training termination condition is met; if not, return to step S61; if so, output the second optimal network parameters.

6. The method for reconstructing abnormal data of a wind speed sensor according to claim 4, characterized in that: The first search method includes local search and global search; the step S422 includes: Step S4221, determining the first nectar source location according to the magnitude relationship of the first fitness values, and then determining the corresponding first fragrance value; Step S4222: when the first search mode is the local search, updating the individual butterfly position according to a preset local search expression; Step S4223, when the first search method is the global search, based on the first nectar source position, the individual butterfly position is updated in combination with a preset global search expression.

7. The method for reconstructing abnormal data of a wind speed sensor according to claim 5, characterized in that: The second search method includes local search and global search; the step S622 includes: Step S6221, determining the second nectar source location according to the magnitude relationship of the second fitness values, and then determining the corresponding second fragrance value; Step S6222, when the second search mode is the local search, updating the individual butterfly position according to the local search expression; Step S6223, when the second search mode is the global search, based on the second nectar source position, the individual butterfly position is updated in combination with the global search expression.

8. A device for reconstructing abnormal data of a wind speed sensor, characterized in that: include: A primary decomposition module is used to perform variational mode decomposition on the acquired historical wind speed sequence to obtain multiple primary decomposition historical wind speed subsequences; A secondary decomposition module is used to decompose the residual component of the primary decomposition historical wind speed subsequence by using an improved adaptive noise complete set empirical mode decomposition algorithm to obtain a secondary decomposition historical wind speed subsequence; A first input module is used to input the once-decomposed historical wind speed subsequence into an initially constructed first LSTM neural network to obtain a corresponding first predicted wind speed sequence; A first training module is used to train the first LSTM neural network by using a crazy butterfly algorithm based on adaptive perturbation, combining the historical wind speed sequence corresponding to the once-decomposed historical wind speed subsequence and the first predicted wind speed sequence, to obtain a first target LSTM neural network; A second input module is used to input the secondary decomposed historical wind speed subsequence into the initially constructed second LSTM neural network to obtain a corresponding second predicted wind speed sequence; A second training module is used to train the second LSTM neural network by using the crazy butterfly algorithm based on adaptive perturbation, in combination with the historical wind speed sequence corresponding to the secondary decomposition historical wind speed subsequence and the second predicted wind speed sequence, to obtain a second target LSTM neural network; A prediction module, used for inputting the wind speed sequence to be measured into the first target LSTM neural network and the second target LSTM neural network respectively, to obtain a corresponding first reconstructed wind speed sequence and a second reconstructed wind speed sequence; The reconstruction module is used to superimpose the first reconstructed wind speed sequence and the second reconstructed wind speed sequence to obtain a wind speed sequence reconstruction result.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is executed.