Wind power missing data complementation method based on improved WGAIN

Abstract

The invention belongs to the field of wind power missing data complementation. The invention provides a data completion method based on an improved generative adversarial interpolation network in order to solve the problem of missing wind power data. According to the method, a convolutional neural network, a long-short term memory network and a self-attention mechanism are fused in a generator, so that long-term dependency of time sequence data and potential correlation among multiple features are fully mined. And the discriminator optimizes the difference measurement between the generated data distribution and the real data distribution by introducing a Wasserstein distance, so that the complementation precision and the overall performance of the model for missing data are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of wind power data processing, and specifically relates to a method for completing missing data of wind power based on improving WGAIN based on convolutional neural network, long short-term memory network and self-attention mechanism. Background Art

[0002] As energy crisis and environmental problems become increasingly serious, the global energy structure is accelerating the transition from fossil fuels to clean energy such as wind power. Although wind power generation has significant potential, its intermittent and volatile nature poses severe challenges to the planning, operation and control of wind farms. In addition, the large-scale grid connection of wind power further increases the difficulty of stable operation of the power system. Therefore, accurate prediction of wind power is of great significance for the reasonable allocation of system backup capacity, ensuring the safe operation of the power grid and improving economic benefits. However, affected by the uncertainty of the natural environment where wind farms are located, harsh natural conditions may cause abnormal operation of system equipment, communication errors and sensor failures, thereby causing data loss in the supervisory control and data acquisition (SCADA) system. Data loss not only significantly reduces the performance of the SCADA system, but also makes it difficult to obtain complete wind farm data, resulting in a reduction in the number of valid data in the data set, a decrease in data accuracy, affecting the effect of model training, and even causing overfitting risks. The above problems pose a huge challenge to wind power prediction. Therefore, the effective completion of missing data has become a key issue that needs to be solved urgently.

[0003] At present, data completion methods are mainly divided into two categories: classical statistical methods and data-driven methods. Among statistical methods, the autoregressive integrated moving average model (ARIMA) is widely used to establish linear relationships to complete missing values ​​such as power, voltage or current in a specific time data set. However, a single fixed mathematical model is difficult to effectively describe high-dimensional wind turbine data, and the effectiveness of statistical completion methods decreases significantly as the data missing rate increases. In recent years, with the rapid development of deep learning technology, data-driven completion methods based on deep learning have received widespread attention, among which generative adversarial networks (GANs) have shown excellent performance in the field of data generation. However, the traditional GAN ​​framework does not fully consider the time dependence of time series data and ignores the potential correlation between different features of multivariate time series. In addition, due to the significant nonlinear and non-stationary characteristics of wind power series, these methods are often inefficient in dealing with missing wind power data and cannot accurately capture the distribution characteristics of the data, resulting in potential errors in the reconstruction results. Therefore, future research should focus on developing more targeted structured neural network models to further improve the effect of missing wind power data completion. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention proposes a method for improving the missing data completion of wind power of WGAIN. The method realizes high-precision completion of missing data by introducing convolutional neural network (CNN), long short-term memory network (LSTM) and self-attention mechanism (Self-Attention), combined with an improved generative adversarial network (GAN) architecture.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A method for completing missing data of wind power based on improved WGAIN, the method comprising the following steps:

[0007] Step 1: Define the problem of missing wind power data completion.

[0008] Step 2: Design an improved WGAIN model, including optimization of the generator network and discriminator network structures and loss function design.

[0009] Step 3: Optimize the model based on the game training strategy between the generator and the discriminator.

[0010] Step 4: Evaluate the error of the wind power missing data completion results.

[0011] This technology mainly takes into account that wind power data has strong volatility and randomness, and wind power is easily affected by weather. Therefore, the difficulty lies in how to capture the time dependence of time series data and the potential correlation between different characteristics of multivariate time series to achieve high-precision completion of missing data.

[0012] The technical solution is further optimized, and the step 1 specifically includes the following contents:

[0013] In actual projects, the multivariate time series samples of SCADA data collected in wind turbines are X = (x ij )∈R T×D , where T is the number of time steps and D is the number of features. Given the possibility of missing values, a missing matrix M = (m ij )∈R T×D To indicate the missing data. In this matrix, m ij =1 means x ij has been observed, and m ij =0 indicates missing value.

[0014] The step 2 is as follows:

[0015] Step 2.1: Generator Network Design:

[0016] The goal of the generator is to generate missing data based on observed values. In the generator module, the input includes three matrices: the data matrix Use 0 to represent missing samples; the noise matrix Z randomly generates noise in the missing places; the mask matrix M uses 0 and 1 to distinguish missing data from real data. A random matrix with consistent dimensions, whose elements are sampled from the interval [0,0.1]. The generator consists of CNN layer, LSTM layer, self-attention mechanism layer and regression layer. The CNN layer is used to extract time series features, the LSTM layer captures the time dependency of sequence data, the self-attention mechanism layer is used to reveal the potential correlation between multivariate features, and the fully connected layers output the results. The generator network structure diagram is shown in the figure below. Figure 1 shown.

[0017] Step 2.2: Generator loss function design:

[0018] The goal of the generator G is to generate realistic data to deceive the discriminator D. The improved WGAIN generator loss function is composed of the reconstruction loss Loss R and generate loss Loss G The generation loss ensures that the generator continuously improves the quality of generated samples, making the distribution of generated data closer to the distribution of real data. The reconstruction loss is the mean square error loss between the generated samples and the real samples, which is used to constrain the performance of the generator G on known data.

[0019] Step 2.3: Discriminator network design:

[0020] The discriminator uses Wasserrestein distance to distinguish between generated values ​​and true values. Its network structure includes two convolutional layers and one fully connected layer. The activation function of the two convolutional layers is LeakyReLU, and the activation function of the fully connected layer is Sigmoid. The discriminator network structure is shown in the figure below. Figure 2 shown.

[0021] Step 2.4: Design a prompt mechanism:

[0022] In order to improve the ability of the discriminator D to distinguish between original data and generated data, a hint matrix H is designed to provide the discriminator with some data authenticity information to further improve the completion accuracy.

[0023] Step 2.5: Discriminator loss function design:

[0024] The task of the discriminator D is to judge the completed data Which values ​​in are original observations and which are generated values. The discriminator determines the authenticity of the data based on the completed data and the prompt matrix. The loss function uses the Wasserstein distance to optimize the adversarial training of the generator and the discriminator.

[0025] A wind power missing data completion method based on improved WGAIN is shown in the following figure: Figure 3 shown.

[0026] The step 3 is as follows:

[0027] The alternating optimization strategy of generative adversarial network is adopted. The discriminator parameters are fixed to optimize the generator, and then the generator parameters are fixed to optimize the discriminator, so as to achieve effective update of model parameters.

[0028] The step 4 is as follows:

[0029] The root mean square error (RMSE) is used as an indicator to evaluate the error of the completed wind power data.

[0030] Compared with the prior art, the beneficial effects of the present invention are mainly manifested in:

[0031] Considering the volatility and randomness of wind power data, this paper designs a generator network structure with convolutional neural network (CNN), long short-term memory network (LSTM) and self-attention mechanism to fully explore the long-term dependency of time series data and the potential correlation between multiple features. The discriminator optimizes the difference measure between the generated data distribution and the real data distribution by introducing Wasserstein distance, which significantly improves the model's missing data completion accuracy and overall performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the generator network structure diagram;

[0033] Figure 2 It is the structure diagram of the discriminator network;

[0034] Figure 3 This is the architecture diagram of the wind power missing data completion method based on the improved WGAIN. DETAILED DESCRIPTION

[0035] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments.

[0036] Taking into account the significant time series characteristics of wind power data, the present invention combines the advantages of long short-term memory network (LSTM), convolutional neural network (CNN) and self-attention mechanism to design an improved generator network structure. LSTM can effectively capture the time-dependent characteristics in time series, CNN can extract multi-scale and multi-directional feature information, and the self-attention mechanism can reveal the potential correlation between multivariate time series features. In addition, the discriminator loss function introduces Wasserstein distance to optimize the similarity between the generated data distribution and the real data distribution, thereby proposing a wind power missing data completion model based on improved WGAIN.

[0037] The method specifically comprises the following steps:

[0038] Step 1: Define the problem of missing wind power data completion.

[0039] In actual engineering, the SCADA data collected from wind turbines forms a multivariate time series sample X = (x ij )∈R T×D , where T is the number of time steps and D is the number of features. In case of missing data, a missing matrix M=(m ij )∈R T×D To indicate the missing data. In this matrix, m ij =1 indicates the corresponding data x ij has been observed, and m ij =0 indicates the corresponding data x ij Missing.

[0040]

[0041] Observational data It can be represented by input X and M as shown below:

[0042]

[0043] Where ⊙ is the Hadamard product operation of the matrix.

[0044] Suppose the output of the generator is While generating missing data, the generator also generates and replaces the original non-missing data. It is not the final completed data. Suppose the final output completed data is The original data is retained in the non-missing part, and the missing part is filled by the data generated by the model. The calculation method is as follows:

[0045]

[0046] Where E is a matrix with the same dimension as M and all elements are 1.

[0047] Step 2: Design an improved WGAIN model, including optimization of the generator network and discriminator network structures and loss function design.

[0048] The step 2 specifically includes the following steps:

[0049] Step 2.1: Generator Network Design

[0050] The goal of the generator is to generate missing data based on observed values. In the generator module, the input includes three matrices: the data matrix Use 0 to represent missing samples; the noise matrix Z randomly generates noise in the missing places; the mask matrix M uses 0 and 1 to distinguish missing data from real data. A random matrix of uniform dimension whose elements are sampled from the interval [0,0.1]. The generator consists of CNN layer, LSTM layer, self-attention mechanism layer and fully connected layer. The CNN layer is used to extract features, the LSTM layer captures the temporal dependency of the time series, and the self-attention mechanism layer is used to reveal the potential correlation between multivariate features.

[0051] In multivariate time series data, there are different variables. By analyzing the correlation between different variables, the final completion accuracy can be improved. Therefore, in order to exploit the potential correlation between different variables, a self-attention mechanism is designed to capture the potential correlation between different variables. The self-attention mechanism automatically assigns different weights to each element in the sequence to better capture the relationship between different elements. Although irrelevant features have been removed, the input data still contains multiple different features or indicators. In order to model the interaction between these features, a self-attention mechanism is designed to capture the potential relationship between data and features. The calculation of its self-attention can be described by the following formula:

[0052] Q=XW q

[0053] K=XW k

[0054] V=XW v

[0055]

[0056] Where Q and K are the query vector and key vector for calculating similarity, V is the value vector, and d k is the dimension of the key. X is the dimension of R n×m data, and is the training parameter.

[0057] Step 2.2: Generator loss function design:

[0058] The goal of the generator G is to generate realistic data to deceive the discriminator D. The improved WGAIN generator loss function is composed of the reconstruction loss Loss R and generate loss Loss G The generation loss ensures that the generator continuously improves the quality of generated samples, making the distribution of generated data closer to the distribution of real data. The reconstruction loss is the mean square error loss between the generated samples and the real samples, which is used to constrain the performance of the generator G on known data. Let C = (c ij )∈R T×D is the output of the discriminator. The generator loss function L G The calculation formula is as follows:

[0059]

[0060] Loss G =-E[c ij |m ij =0]

[0061] L G =Loss G +λLoss R

[0062] Where E is the expectation and λ is a hyperparameter that controls the ratio between generation loss and reconstruction loss.

[0063] Step 2.3: Discriminator network design

[0064] The discriminator network structure includes two convolutional layers and one fully connected layer. The activation function of the two convolutional layers is LeakyReLU, and the activation function of the fully connected layer is Sigmoid, so that the output value range is between 0 and 1. The structure of the discriminator is as follows Figure 2 shown.

[0065] Step 2.4: Design a prompt mechanism:

[0066] In order to improve the ability of the discriminator D to distinguish between original data and generated data, a hint matrix H is designed. The calculation formula of the hint matrix H is as follows:

[0067] H=W⊙M+0.5(1-W)

[0068] Where W is a random binary moment with the same dimension as the mask matrix M, and its elements are sampled from {0,1}.

[0069] When W is a matrix of all 1s, H = M; when W is a matrix of all 0s, all elements of H are 0.5. Therefore, the sampling space of the hint matrix H is {0, 0.5, 1}. When the elements of H are 0 or 1, the hint matrix clearly identifies the input data as a generated sample or a real sample; when the elements are 0.5, the hint matrix does not provide additional information and the discriminator needs to make its own judgment. By adjusting W, the amount of reference information that the discriminator relies on can be flexibly controlled.

[0070] Step 2.5: Design the discriminator loss function:

[0071] The task of the discriminator D is to judge the completed data Which values ​​in are original observations and which values ​​are generated values. Its input is the completed data and the prompt matrix H, the output is a matrix of the same size as the original data, and each value in the matrix represents the probability that the corresponding position of the input data is the original observation value. When the output value is P = 1, the discriminator judges that the position is the original observation value; when the output value is p = 0, it is judged to be generated data. In this case, the discriminator is equivalent to the prediction mask matrix M, in which the value corresponding to the original observation data is 1 and the value corresponding to the missing value is 0. In order to further improve the stability and performance of the generative adversarial network, the Wasserstein distance is introduced to replace the traditional cross entropy loss. The Wasserstein distance can better measure the difference between the distribution of generated data and the distribution of real data, while alleviating the problem of gradient disappearance. Specifically, the goal of the discriminator is to make the distribution of generated data closer to that of real data by minimizing the Wasserstein distance. It should be emphasized that the mask matrix M is predefined, so the training of the discriminator can be regarded as supervised learning, thereby effectively improving its performance. The loss function of the discriminator is:

[0072] D loss =E[c ij |m ij =0]-E[c ij |m ij =1]

[0073] Step 3: Optimize the model based on the game training strategy between the generator and the discriminator.

[0074] The step 3 specifically comprises the following steps:

[0075] The generative adversarial network adopts an alternating optimization strategy, that is, after optimizing the generator G once, the discriminator D is optimized k times. When optimizing the generator, the discriminator parameters are kept fixed so that the generator generates samples as realistic as possible to deceive the discriminator. When optimizing the discriminator, the generator parameters are fixed so that the discriminator can distinguish between real samples and generated samples as accurately as possible.

[0076] Step 4: Evaluate the error of the wind power missing data completion results.

[0077] The step 4 specifically comprises the following steps:

[0078] The wind power data containing missing values ​​is input into the improved WGAIN model in step 2, and after the game training in step 3, the generative adversarial network model gradually converges. Finally, the generator generates the completed data, and the experimental results are evaluated by the root mean square error (RMSE). The smaller the RMSE value, the closer the generated data is to the true value. The following is the mathematical definition of the root mean square error.

[0079]

[0080] Where: Y i and are the true value and generated value of the i-th sample respectively; N is the total number of samples.

[0081] Compared with the prior art, the beneficial effects of the present invention are mainly manifested in:

[0082] In response to the problem of missing wind power data, the present invention proposes a data completion method based on an improved generative adversarial interpolation network (WGAIN) of CNN-LSTM-Self-Attention. In response to the strong volatility and randomness of wind power, the invention integrates convolutional neural network (CNN), long short-term memory network (LSTM) and self-attention mechanism (Self-Attention) in the generator to fully explore the long-term dependency of time series data and the potential correlation between multiple features. The discriminator optimizes the difference measure between the generated data distribution and the real data distribution by introducing the Wasserstein distance, which significantly improves the model's completion accuracy and overall performance for missing data.

[0083] It should be noted that, in this article, terms such as "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "include..." or "include..." do not exclude the existence of additional elements in the process, method, article or terminal device including the elements. In addition, in this article, "greater than", "less than", "exceed" and the like are understood to exclude the number itself; "above", "below", "within" and the like are understood to include the number itself.

[0084] Although the above embodiments have been described, once those skilled in the art know the basic creative concepts, they can make additional changes and modifications to these embodiments. Therefore, the above description is only an embodiment of the present invention and does not limit the patent protection scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the specification and drawings of the present invention, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for completing missing data of wind power based on improved WGAIN, characterized in that: The following steps are involved: Step 1: Define the problem of missing wind power data completion; Step 2: Design an improved WGAIN model, including optimization of the generator network and discriminator network structures and loss function design; Step 3: Optimize the model based on the game training strategy between the generator and the discriminator; Step 4: Evaluate the error of the wind power missing data completion results.

2. The method for completing missing data of wind power based on improved WGAIN according to claim 1, characterized in that: The step 1 specifically includes: Define a wind power data X = (x ij )∈R T×D Matrix M of the same size = (m ij )∈R T×D To indicate the missing data; in this matrix, m ij =1 means x ij has been observed, and m ij =0 indicates missing value; observed data It can be represented by input X and M as shown below: Where ⊙ is the Hadamard product operation of the matrix; Define the completed data for the final output as Then there is Where E is a matrix with the same dimension as M and all elements are 1. is the output of the generator.

3. The method for completing missing data of wind power based on improved WGAIN according to claim 1, characterized in that: The step 2 specifically includes the following contents: Step 2.1: Generator Network Design The generator receives a matrix of data The noise matrix Z and the mask matrix M are composed of CNN layer, LSTM layer, self-attention mechanism layer and fully connected layer. The CNN layer is used to extract features, the LSTM layer captures the time dependency of the time series, and the self-attention mechanism layer reveals the potential correlation between multiple variables by assigning different weights to the sequence. The specific calculation formula is as follows: Q=XW q K=XW k V=XW v Where Q and K are the query vector and key vector for calculating similarity, V is the value vector, and d k is the dimension of the key, X is the dimension of R n×m data, and is the training parameter; Step 2.2: Generator loss function design: The improved WGAIN generator loss function is composed of the reconstruction loss Loss R and generate loss Loss G Composition, including Loss R Ensure that the generated data is close to the true distribution, Loss G Limit the performance of the generator on known observations, let C = (c ij )∈R T×D C is the output of the discriminator, and the generator loss function L G The calculation formula is as follows: Loss G =-E[c ij |m ij =0] L G =Loss G +λLoss R Where E is the expectation and λ is a hyperparameter that controls the ratio between generation loss and reconstruction loss. Where λ is a hyperparameter that controls the ratio between generation loss and reconstruction loss; Step 2.3: Discriminator network design The discriminator network structure consists of two convolutional layers and one fully connected layer. The activation function of the two convolutional layers is LeakyReLU, while the activation function of the fully connected layer is Sigmoid, which is used to generate the final probability value to ensure that the output is in the range of [0,1]. Step 2.4: Design a prompt mechanism: In order to improve the ability of the discriminator D to distinguish between original data and generated data, a hint matrix H is designed. The calculation formula of the hint matrix H is as follows: H=W⊙M+0.5(1-W) Where W is a random binary matrix with the same dimension as the mask matrix M, and its elements are sampled from {0,1}; Step 2.5: Discriminator loss function design: The task of the discriminator D is to judge the completed data Which values ​​in are original observations and which values ​​are generated values, and the input is the completed data And the prompt matrix H, the output is a matrix of the same size as the original data. Each value in the matrix represents the probability that the corresponding position of the input data is the original observation value. The discriminator uses the Wasserstein distance to distinguish between generated values ​​and true values. By improving the stability of learning and solving the mode collapse problem, the model is easier to train than other GANs. Its loss function is: D loss =E[c ij |m ij =0]-E[c ij |m ij =1]。 4. The method for completing missing data of wind power based on improved WGAIN according to claim 1, characterized in that: The step 3 adopts the generator and discriminator alternating optimization strategy: Optimize the generator while fixing the discriminator parameters to generate data that is as realistic as possible; The discriminator is optimized while the generator parameters are fixed to accurately distinguish between real data and generated data.

5. The method for completing missing data of wind power based on improved WGAIN according to claim 1, characterized in that: The step 4 uses the root mean square error to evaluate the error of the wind power missing data completion result.

6. The method for completing missing data of wind power based on improved WGAIN according to claim 5, characterized in that: The smaller the value of the root mean square error is, the closer the generated value is to the true value. The following is the mathematical definition of the evaluation index: Where: Y i and are the true value and generated value of the i-th sample respectively; N is the total number of samples.

Images