Sample cleaning method for wind turbine icing prediction model training

Abstract

The present invention relates to the field of data processing technology, and in particular to a sample cleaning method for training a fan icing prediction model, the method comprising: obtaining a sample to be cleaned during the operation of a fan; wherein the sample to be cleaned includes multiple factors; using a fan icing prediction model to determine the temporal coupling between the multiple factors; and cleaning the sample to be cleaned based on the temporal coupling between the multiple factors. The method can effectively eliminate the influence of noise on the sample and achieve efficient cleaning of sample data.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a sample cleaning method for training a wind turbine icing prediction model in the technical field of data processing. Background Art

[0002] As the temperature drops, the power decreases. Due to freezing rain or snowfall, the water mist entering the wind turbine condenses into ice and covers the blades, forming wind turbine icing, which has a great impact on the load change of the wind power generation system and the operation status of the wind turbine. Therefore, in order to ensure the normal and safe operation of the wind turbine and reduce maintenance costs, the prediction of wind turbine blade icing becomes very important.

[0003] The historical operation data of wind turbines is the basis for prediction and early warning. However, in the actual operation of wind turbines, a large amount of abnormal data will be generated during operation due to factors such as geographical location, meteorological factors, errors in data collection and transmission, and equipment failures. These abnormal data cannot accurately reflect the actual operating status of wind turbines, resulting in the inability to establish an accurate connection with wind turbine parameters.

[0004] Existing wind turbine data cleaning methods mainly include statistical methods, density-based or clustering methods. When the amount of abnormal data is large, the statistical method is more likely to fail to identify; the selection of model parameters based on density and clustering methods is usually difficult; the local outlier detection algorithm can only identify some points with very low density values. When the density of outliers is similar to that of normal values, they will be divided into the same cluster and cannot be effectively identified, thus failing to accurately and effectively complete data cleaning. Summary of the invention

[0005] In order to solve the technical problem that the existing wind turbine data cleaning method cannot accurately and effectively complete data cleaning, the purpose of the present invention is to provide a sample cleaning method for wind turbine icing prediction model training. The technical solution adopted is as follows:

[0006] In a first aspect, an embodiment of the present invention provides a sample cleaning method for training a wind turbine icing prediction model, wherein the wind turbine icing prediction model used in the method is composed of a time series prediction model and a prediction model for predicting wind turbine icing phenomena, and the method comprises:

[0007] Acquire a sample to be cleaned during the operation of the fan; wherein the sample to be cleaned includes at least two factors;

[0008] Using the time series prediction model, determining the time series coupling between different factors;

[0009] Based on the temporal coupling between different factors, the sample to be cleaned is cleaned.

[0010] In combination with the first aspect above, in some possible implementations, obtaining at least two factors during the operation of the wind turbine includes:

[0011] Acquiring geographic location information data, meteorological element data, and icing shutdown data of the wind turbine during operation;

[0012] At least two factors are determined based on the geographic location information data, the meteorological element data, and the icing shutdown data.

[0013] In combination with the first aspect, in some possible implementations, the determining of at least two factors based on the geographic location information data, the meteorological element data, and the icing shutdown data includes:

[0014] The geographic location information data, the meteorological element data and the icing shutdown data are cleaned to obtain at least two factors.

[0015] In the above method, by collecting geographical location information data, meteorological element data and icing shutdown data of the wind turbine during operation, and performing data cleaning operations such as normalization on these data, a variety of factors that are convenient for subsequent calculations are obtained.

[0016] In combination with the first aspect above, in some possible implementations, the method further includes:

[0017] Eliminate any row element in the training set of the sample to be cleaned to obtain a training sequence;

[0018] Based on the training sequence, the prediction model to be trained is trained to obtain a time series prediction model for any row element.

[0019] In the above method, by eliminating the training sequence of any row element in the training set and training the prediction model to be trained, a time series prediction model for the row element can be obtained to improve the prediction accuracy of the prediction model for predicting wind turbine icing phenomenon in the wind turbine icing prediction model.

[0020] In combination with the first aspect above, in some possible implementations, using the time series prediction model to determine the time series coupling between different factors includes:

[0021] Input each factor in the training set into the time series prediction model to obtain a predicted value;

[0022] Determining the contribution of each of the factors to the predicted value;

[0023] Based on the contribution of each factor to the predicted value, determining the similarity between the contribution of different factors;

[0024] The similarity and the contribution are combined to obtain the temporal coupling between different factors.

[0025] In the above manner, by analyzing the similarities between the contribution degrees corresponding to different kinds of factor data, the similarities are merged, so that the coupling relationship between various factor data can be accurately analyzed, that is, the temporal coupling between the various factor data can be obtained.

[0026] In combination with the first aspect, in some possible implementations, determining the similarity between the contribution degrees corresponding to different factors based on the contribution degree of each factor to the predicted value includes:

[0027] Based on the contribution of each factor to the predicted value, determining the similarity between the contribution corresponding to any two factors;

[0028] The similarities between the contribution degrees corresponding to the arbitrary two factors are merged to obtain the similarities between the contribution degrees corresponding to the different types of factors.

[0029] In combination with the first aspect above, in some possible implementations, the cleaning of the sample to be cleaned based on the timing coupling between the different factors includes:

[0030] Determining edge samples among the samples to be cleaned;

[0031] Determining adjacent samples corresponding to the edge samples based on the temporal coupling between different factors in the edge samples;

[0032] Based on the neighboring samples, interpolating and synthesizing the edge samples to obtain synthesized samples;

[0033] The sample to be cleaned is cleaned based on the synthetic sample.

[0034] In the above method, by analyzing the edge samples in the samples to be cleaned and the neighboring samples of the edge samples, the edge samples are interpolated and synthesized to obtain the corrected generated samples, and the abnormal samples are cleaned, thereby achieving data cleaning of the samples to be cleaned.

[0035] In combination with the first aspect, in some possible implementations, determining edge samples in the samples to be cleaned includes:

[0036] Clustering the samples to be cleaned to obtain clustering results;

[0037] Based on the clustering result, edge samples in the samples to be cleaned are determined.

[0038] In the above manner, by analyzing the clustering results of the samples to be cleaned, the edge samples on the boundary of each cluster can be marked quickly and accurately.

[0039] In combination with the first aspect, in some possible implementations, determining the neighboring samples corresponding to the edge samples based on the temporal coupling between different factors in the edge samples includes:

[0040] Based on the temporal coupling between different factors in the edge sample, determining the decoding neighbor distance between the edge sample and the candidate sample; wherein the candidate sample is a generated sample corresponding to a preset cluster in the clustering result;

[0041] Based on the decoded neighbor distance, a neighboring sample corresponding to the edge sample is determined.

[0042] In the above manner, by analyzing the decoded neighbor distance between the edge sample and the generated sample, the generated sample with the shortest decoded neighbor distance is used as the neighboring sample corresponding to the edge sample.

[0043] In combination with the first aspect, in some possible implementations, the interpolating and synthesizing the edge samples based on the neighboring samples to obtain the synthesized samples includes:

[0044] Using the neighboring samples to perform interpolation on different factors in the edge samples to obtain an interpolation result;

[0045] The interpolation results are synthesized to obtain the synthesized samples.

[0046] In the above method, edge samples are interpolated and synthesized through neighboring samples to obtain synthetic samples, thereby providing more samples of wind turbines in different states for the icing prediction model and reducing the impact of sample imbalance.

[0047] In a second aspect, a sample cleaning device for training a wind turbine icing prediction model is provided, wherein the wind turbine icing prediction model used by the device is composed of a time series prediction model and a prediction model for predicting wind turbine icing phenomena, and the device comprises:

[0048] An acquisition module is used to acquire a sample to be cleaned during the operation of the fan; wherein the sample to be cleaned includes at least two factors;

[0049] A determination module, used to determine the timing coupling between different factors using the timing prediction model;

[0050] The cleaning module is used to clean the sample to be cleaned based on the timing coupling between the different factors.

[0051] According to a third aspect, a computer program product is provided, comprising: a computer program code, which, when executed on a computer, enables the computer to execute the method according to the first aspect or any possible implementation of the first aspect.

[0052] In a fourth aspect, a computer-readable storage medium is provided, which stores a computer program code. When the computer program code runs on a computer, the computer executes the method in the above-mentioned first aspect or any possible implementation manner of the first aspect.

[0053] The present invention has the following beneficial effects: after obtaining the samples to be cleaned during the operation of the fan, the timing prediction model in the fan icing prediction model is used to evaluate the timing coupling between multiple factors; and based on the timing coupling between the multiple factors, the samples to be cleaned are cleaned; in this way, by evaluating the timing coupling between the factors in the samples to be cleaned sampled at different times, an analysis result with a more fine-grained data correlation can be obtained, so as to more efficiently clean the data to be cleaned. In this way, after accurately analyzing the timing coupling between various factors, the samples to be cleaned are cleaned, so that the influence of noise and the like on the samples can be effectively eliminated to achieve efficient cleaning of the sample data. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are 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 work.

[0055] Figure 1 It is a schematic diagram of the implementation flow of a sample cleaning method for training a wind turbine icing prediction model provided by an embodiment of the present invention;

[0056] Figure 2 It is another implementation flow diagram of the sample cleaning method for wind turbine icing prediction model training provided by an embodiment of the present invention;

[0057] Figure 3 Schematic diagram of the implementation framework of the variational autoencoder model provided by an embodiment of the present invention;

[0058] Figure 4 It is another implementation flow diagram of a sample cleaning method for training a wind turbine icing prediction model provided by an embodiment of the present invention;

[0059] Figure 5It is a schematic diagram of an application scenario of a sample cleaning method for training a wind turbine icing prediction model provided by an embodiment of the present invention;

[0060] Figure 6 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the sample cleaning method for wind turbine icing prediction model training proposed by the present invention, its specific implementation method, structure, features and effects are described in detail as follows in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0062] Among them, in the description of the embodiments of the present invention, unless otherwise specified, " / " means or, for example, A / B can mean A or B: "and / or" in the text is only a way to describe the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present invention, "multiple" refers to two or more than two.

[0063] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as suggesting or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features.

[0064] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0065] The following is a detailed description of the sample cleaning method for wind turbine icing prediction model training provided by the present invention in conjunction with the accompanying drawings. Figure 1 , which shows a flow chart of a sample cleaning method for training a wind turbine icing prediction model provided by an embodiment of the present invention, the method comprising:

[0066] 101, obtaining a sample to be cleaned during the operation of the fan.

[0067] The sample to be cleaned includes at least two factors, namely multiple factors, including: geographic location information data, meteorological element data and ice-covered shutdown data.

[0068] In some possible implementations, first, the geographical location information data, meteorological element data and ice-covered shutdown data of the wind turbine during operation are obtained; for example, the geographical location information data, meteorological element data and ice-covered shutdown data of the wind turbine during operation are collected at a certain time interval. Then, based on the geographical location information data, the meteorological element data and the ice-covered shutdown data, the multiple factors are determined. In a specific example, the multiple factors are obtained by cleaning the geographical location information data, the meteorological element data and the ice-covered shutdown data. In some possible implementations, the wind farm data acquisition and monitoring and control system (SCADA) is used to continuously record the factors in the operation of the wind turbine. During the data collection period, the sampling intervals of various factors are the same. The factors include the geographical location information data of the wind farm, including longitude, latitude and altitude; the meteorological element data of the wind farm, including the temperature, humidity, wind speed, wind direction and air pressure on the wind turbine; the ice-covered shutdown data of the wind farm, including the shutdown frequency and shutdown duration of the wind turbine; and the output power of the wind turbine. The downtime duration is the time interval between the shutdown moment and the normal operation moment. In this embodiment, the meteorological element data of the wind turbine farm is recorded continuously, that is, when the wind turbine is shut down due to ice coating, the recording of the meteorological element data will not stop. For the sampling data when it is not in the shutdown state, the shutdown frequency and shutdown duration at the time of sampling are set to 0. In order to eliminate the influence of the dimension and reduce the amount of calculation during data cleaning, each type of meteorological element data and ice-covered shutdown data is normalized separately. The normalization process includes but is not limited to maximum-minimum value normalization and standard score (z-score) normalization. In this way, by collecting the geographic location information data, meteorological element data and ice-covered shutdown data of the wind turbine during operation, and performing data cleaning operations such as normalization on these data, a variety of factors that are convenient for subsequent calculations are obtained.

[0069] 102. Using a wind turbine icing prediction model, determine the temporal coupling between the multiple factors.

[0070] Here, by obtaining a wind turbine icing prediction model, and using the wind turbine icing prediction model, the temporal coupling between the multiple factors is evaluated. The temporal coupling between the multiple factors is used to represent the coupling relationship between different factors.

[0071] It should be understood that the wind turbine icing prediction model in this embodiment includes two parts: one part is a time series prediction model 1 for analyzing the coupling between collected data, and the time series prediction model 1 can be a single-factor time series prediction model or a multi-factor time series prediction model. In some possible implementations, the time series prediction model 1 can be an exponential moving average model (EMA). The wind turbine icing prediction model used to determine the time series coupling between different factors actually refers to the above-mentioned time series prediction model 1 in the wind turbine icing prediction model. The other part is a prediction model 2 for predicting the wind turbine icing phenomenon. The prediction model 2 is trained using the samples obtained after cleaning the samples to be cleaned, and finally realizes the prediction of the wind turbine icing phenomenon. The prediction model 2 for predicting the wind turbine icing phenomenon is an existing prediction model. In some possible implementations, the existing prediction model can be an autoregressive moving average (AMRIA) model, a random forest, or a prediction model based on a neural network.

[0072] In some possible implementations, a training sequence is obtained by removing any row of elements from the training set of the samples to be cleaned; and based on the training sequence, the prediction model to be trained is trained to obtain a fan icing prediction model for any row of elements. For example, the first row of elements is first removed from the training set of the samples to be cleaned to obtain a fan icing prediction model for the first row of elements (i.e., the timing prediction model 1 in the above-mentioned fan icing prediction model). Here, the first row of elements of all samples in the training set is used as the target value, and the sequence composed of the remaining elements after removing the target value from each sample is used as the actual input sequence of the sample to train the prediction model for the first row of elements; in this way, by removing the training sequence of any row of elements from the training set to train the prediction model to be trained, it is possible to obtain a fan icing prediction model for the row of elements, so as to improve the prediction accuracy of the fan icing prediction model.

[0073] In some possible implementations, when using a prediction model to predict the wind turbine icing phenomenon, a variational autoencoder (VAE) is used to generate high-quality sample data. However, as a non-sequential model, VAE does not have the ability to capture the temporal dynamics of data, resulting in low quality of generated data and difficulty in obtaining high-confidence prediction results. The dynamic change characteristics of high-quality generated data should be directly related to the wind turbine icing phenomenon, rather than being affected by other factors to produce dynamic changes. However, for the collected historical data, there is also a coupling relationship between different factors. For example, there is a correlation between altitude, air pressure and condensation point, and there is also a correlation between temperature and altitude, which leads to the formation of wind turbine icing phenomenon earlier in high altitude areas than in low altitude areas. Therefore, consider analyzing the coupling relationship between factors and clarifying the direct correlation between various factors and wind turbine icing phenomenon. All factors sampled each time in the non-icing shutdown state of the wind turbine are taken as a column element in the matrix, and the data matrix is ​​composed of all the sampled data. In the embodiment of the present invention, the data matrix is ​​divided into a training set and a test set by random sampling, and the prediction model is trained by the training set. The ratio of the training set to the test set is in the range of 6:4-8:2. The specific ratio can be determined by the implementer according to the amount of data. The coupling relationship between the elements is evaluated by the prediction model. The above step 102 can be performed by Figure 2 The steps shown achieve:

[0074] 201, input each factor in the training set into a wind turbine icing prediction model to obtain a prediction value.

[0075] 202, determine the contribution of each factor to the predicted value.

[0076] 203 , based on the contribution of each factor to the predicted value, determine the similarity between the contribution degrees corresponding to different factors.

[0077] Here, the similarity between the contribution of any two factors is calculated by the contribution of each factor to the predicted value; wherein the similarity includes but is not limited to cosine similarity, Peel correlation coefficient, and mutual correlation coefficient. The similarity between the contribution of any two factors is merged to obtain the similarity between the contribution of different factors. Here, by calculating the similarity between the contribution of any two factors, multiple similarity values ​​can be obtained, and the multiple similarity values ​​are summed to obtain the similarity between the contribution of different factors.

[0078] 204 , the similarity and the contribution are integrated to obtain the temporal coupling between the multiple factors.

[0079] In the above steps 201 to 204, the prediction model for each factor is trained by the training set, and the elements in each sample are involved in the calculation of the model explanation (Shapley Additive ExPlanations, SHAP) value to obtain the temporal coupling between the two factors. By evaluating the influence of the factors in the sampled data at different times, a more fine-grained data correlation analysis result can be obtained to facilitate subsequent correlation filtering.

[0080] In the above steps 201 to 204, the actual input sequence of each sample (i.e., each factor) in the test set is input into the time series prediction model 1 in the wind turbine icing prediction model, and the time series prediction model 1 outputs the predicted value of the actual input sequence, and the SHAP attribution analysis method is used to determine the contribution of each element in the actual input sequence of each factor to the predicted value, i.e., the SHAP value of each element. After that, the remaining row elements of all samples in the training set are respectively used as target values, and the above process is repeated in sequence to determine the SHAP value of each element in the actual input sequence of each factor in the test set for each predicted value; the reason for doing so is that training the prediction model for each factor can obtain a more fine-grained data correlation analysis result by evaluating the degree of influence between factors in the sampled data at different times, so as to facilitate subsequent correlation filtering. A clustering algorithm is used to divide all samples in the training set into multiple clusters, and the coupling between the two factors is determined based on the SHAP of the elements in the samples in different clusters. The clustering algorithm includes but is not limited to k-shape clustering (Example 1, clustering based on sequences), k-means clustering (Example 2, clustering based on discrete points in space), and chameleon clustering (Example 3, clustering based on graph theory). If the SHAP value of the prediction value output by the prediction model of the U-th factor for the A-th factor is large in different clusters, it means that no matter the wind turbine is in a high output power, low output power or stable operation state, the change of the U-th factor has a greater impact on the value of the A-th factor, which means that the change of the factor value will not only be affected by the icing phenomenon, but also by other factors, and the greater the impact on determining the correlation between the A-th factor and the wind turbine icing shutdown phenomenon. The maximum, minimum, mean, and third quartile sequence of all SHAP values ​​of the U-th factor in each cluster with respect to the predicted value of the A-th factor is taken as the characteristic sequence of the contribution of the U-th factor to the A-th factor in each cluster. This can avoid the problem of ignoring all features by comparing only some points, and solve the large amount of calculations caused by directly comparing all SHAP values ​​in two clusters.

[0081] Based on the above analysis, the temporal coupling is constructed here to characterize the coupling strength between two factors. The temporal coupling between the Uth factor and the Ath factor is calculated as shown in formula (1):

[0082] (1);

[0083] Among them, U and A represent any two factors respectively. is the temporal coupling between the Uth factor and the Ath factor, is the sum of the similarity measurement results of the contribution feature sequences of the U-th factor to the A-th factor among all clusters (i.e., the similarity between the contributions of different factors), k is the number of clusters of samples in the training set, i is the i-th cluster, n is the type of factor, the size of n is equal to the number of elements in any sample, B is the B-th factor, It is the absolute value of the difference between the same order elements of the contribution feature sequences of the Uth factor in the i-th cluster to the Ath factor and the Bth factor. In this way, by analyzing the similarity between the contributions corresponding to different factor data and fusing the similarity, the coupling relationship between various factor data can be accurately analyzed, that is, the temporal coupling between multiple factors can be obtained.

[0084] 103 . Clean the sample to be cleaned based on the timing coupling between the multiple factors.

[0085] Here, the influence degree between the multiple factors is reflected by the temporal coupling between the multiple factors; and based on the influence degree between the multiple factors, the correlation between the multiple factors is further analyzed. Among them, the influence degree between the multiple factors is used to represent the influence of one factor on the change of the value of another factor. The influence degree between the multiple factors is proportional to the temporal coupling, that is, the greater the temporal coupling, the greater the influence degree. Based on the above formula (1), in the time dimension, the more stable the influence of the change of the Uth factor on the value of the Ath factor, the higher the similarity between the contribution feature sequences in different clusters, The larger the value of , the more significant the impact of the change of the Uth factor on the Ath factor. Compared with the other factors, the contribution of the change of the Uth factor to the predicted value of the Ath factor is greater. The larger the value of The larger the value of , the greater the impact of the change in the Uth factor on the change in the value of the Ath factor.

[0086] Among them, the degree of influence between multiple factors is proportional to the correlation between the multiple factors; that is, the greater the degree of influence between multiple factors, the greater the value of the correlation between the multiple factors. For example, the higher the degree of influence of the change of the Uth factor on the change of the value of the Ath factor, the greater the influence of the Uth factor on the analysis of the correlation between the Ath factor and the wind turbine icing shutdown phenomenon, that is, the greater the value corresponding to the correlation between the Uth factor and the Ath factor, the closer the correlation between the Uth factor and the Ath factor. In this way, by analyzing the coupling relationship between various factors, the degree of influence between multiple factors can be accurately evaluated through the coupling relationship.

[0087] In step 103, the time-series coupling between the various factors is used to reflect the closeness of the correlation between the various factors, so as to achieve data cleaning of the samples to be cleaned.

[0088] In some possible implementations, the samples in the training set are used as input, and the VAE model is used to obtain the corresponding generated samples. The variational autoencoder model includes two parts: an encoder and a decoder. The process of obtaining the generated samples is as follows: Figure 3 As shown, first, the sample 30 is input to the encoder 31, and the encoder 31 outputs multiple means and variances, for example, mean 1, mean 2..., mean n; variance 1, variance 2..., variance n. Afterwards, the normal distribution of each mean and variance is calculated to obtain normal distribution 1, normal distribution 2..., normal distribution n. Normal distribution 1, normal distribution 2..., normal distribution n are input to the decoder 32 to obtain the generated sample 33 output by the decoder 32. The encoder maps the input data to the latent space and obtains multiple groups of normal distribution results. The decoder maps the latent representation back to the original data space. The loss function of the model consists of two parts: reconstruction loss and KL divergence. The optimizer is the (Adaptive Moment Estimation, Adam) optimizer. The above step 103 can be achieved by Figure 4 The steps shown achieve:

[0089] 401 : Determine edge samples among the samples to be cleaned.

[0090] Here, clustering results are obtained by clustering the samples to be cleaned, and edge samples in the samples to be cleaned are determined based on the clustering results. In this way, by analyzing the clustering results of the samples to be cleaned, edge samples on the boundary of each cluster can be quickly and accurately marked.

[0091] In some possible implementations, the local density of the generated samples in each cluster in the clustering results is statistically calculated, and the generated samples with the smallest local density are marked as edge samples on the boundary of each cluster. Afterwards, the mean of the sample size in all clusters is calculated, and each cluster with a sample size less than the mean is marked as a small sample cluster (i.e., a preset cluster); the purpose of marking the small sample cluster is that the sample size under the wind turbine icing shutdown is usually smaller than the sampling number when the wind turbine is not icing shutdown, which facilitates the subsequent interpolation to use the samples under the wind turbine icing shutdown with a greater probability, providing more samples of wind turbines in different states for the icing prediction model, and reducing the impact of sample imbalance.

[0092] 402 : Determine a neighboring sample corresponding to the edge sample based on the temporal coupling between multiple factors in the edge sample.

[0093] In some possible implementations, the above step 402 may be implemented by the following steps:

[0094] In the first step, the decoding neighbor distance between the edge sample and the candidate sample is determined through the temporal coupling between multiple factors in the edge sample; wherein the candidate sample is a generated sample corresponding to a preset cluster in the clustering result.

[0095] Here, since the VAE model generates an output for each input, the sampling time of each edge sample corresponding to the input can be obtained; secondly, the sequence of the same factors in all samples in the training set according to the sampling time is used as the input of the Standard Template Library (STL) algorithm for time series analysis to obtain the trend value of each element in each sample. The decoding neighbor distance between the edge sample q and the generated sample j in all small sample clusters As shown in formula (2):

[0096] (2);

[0097] in, is the normalized result of the absolute value of the difference between the sampling time of the edge sample q and the generated sample j. For example, the standard score (z-score) normalization or K-means normalization is used to normalize the absolute value of the difference between the sampling time of the edge sample q and the generated sample j, and the result is . is the absolute value of the difference between the time series trend feature values ​​of the edge sample q and the generated sample j, is the Euclidean distance between the edge sample q and the generated sample j in the VAE decoding space.

[0098] The second step is to determine the neighboring sample corresponding to the edge sample based on the decoded neighbor distance. In this way, by analyzing the decoded neighbor distance between the edge sample and the generated sample, the generated sample with the closest decoded neighbor distance is used as the neighboring sample corresponding to the edge sample.

[0099] Here, the mean of the trend values ​​of all elements in the input corresponding to each edge sample is taken as the temporal trend feature value of each edge sample. The decoding neighbor distance D is calculated by combining the spatial distance between each edge sample and the generated sample in the small sample cluster, and k generated samples with the smallest decoding neighbor distances are selected as the neighboring samples of each edge sample. The empirical value of k is 10.

[0100] 403 : Based on the neighboring samples, perform interpolation synthesis on the edge samples to obtain synthesized samples.

[0101] Here, the neighboring samples are used to interpolate multiple factors in the edge samples to obtain interpolation results, and the interpolation results are synthesized to obtain the synthesized samples. In this way, the edge samples are interpolated and synthesized by neighboring samples to obtain synthesized samples, thereby providing more samples of wind turbines in different states for the icing prediction model, reducing the impact of sample imbalance.

[0102] In some possible implementations, based on the temporal coupling of each factor in the input sample corresponding to each edge sample, each edge sample and its k neighboring samples are interpolated to correct the generated sample, as shown in formula (3):

[0103] (3);

[0104] in, is the interpolation result of the synthetic sample V on the Uth factor, is the value of the marginal sample q on the Uth factor, is the mean of the temporal coupling between the Uth factor and all other factors, It is the value of the rth neighboring sample of the edge sample q on the Uth factor.

[0105] Traverse all factors and calculate the interpolation results of the synthetic sample V on all factors to obtain the synthetic sample V, such as Figure 5As shown in the figure, the black triangle is the edge sample q, the dark gray triangle connected to the black triangle is the neighboring sample of the edge sample q, the white triangle on the connecting line is the synthetic sample, and the circle is the remaining generated samples. The purpose of this is to adjust the generated samples of the decoder by adding temporal features to ensure that the distribution of the cleaned samples has temporal features; on the other hand, by interpolating the distribution of the edge samples of each cluster, the influence of the noise value can be eliminated, further improving the sample quality.

[0106] 404 : Clean the sample to be cleaned based on the synthetic sample.

[0107] Here, the decoder's generated samples are adjusted by adding the timing features of the generated samples and using neighbor interpolation to ensure that the cleaned sample distribution has timing features. At the same time, the edge sample distribution of each cluster is adjusted by interpolation to eliminate the influence of noise values ​​and further improve the quality of the cleaned samples.

[0108] In some possible implementations, the Euclidean distance D1 between each synthetic sample and any generated sample x in a non-small sample cluster is calculated, and the Euclidean distance D2 between the generated sample x and any other generated sample is calculated. If D1 is greater than D2, it is considered that there are generated samples of similar or same category around the synthetic sample, and the synthetic sample is retained; if it is not satisfied, it is considered that the generated samples around the synthetic sample are all samples of other categories, and the synthetic sample is a noise point, and the synthetic sample is deleted. Secondly, for the generated sample y in each small sample cluster, the k neighboring samples of the generated sample y are obtained based on the Euclidean distance using k neighboring samples, and the proportion of the generated samples in the non-small sample cluster in the k neighboring samples is calculated. If the proportion is equal to 1, it is considered that the generated samples around the generated sample y are all samples of other categories, and the generated sample y is also a noise point, and the generated sample is deleted. After traversing all synthetic samples and generated samples in the small sample cluster, and completing sample cleaning, the prediction model (i.e., prediction model 2 for predicting wind turbine icing phenomenon in the above-mentioned wind turbine icing prediction model) can be trained with the corrected samples for prediction of wind turbine icing phenomenon, and the prediction model includes but is not limited to the autoregressive moving average (AMRIA) model, random forest, and prediction model based on neural network. In this way, by analyzing the edge samples in the samples to be cleaned and the neighboring samples of the edge samples, the edge samples are interpolated and synthesized, and then the corrected generated samples are obtained, and the abnormal samples are cleaned, that is, the data cleaning of the samples to be cleaned is achieved.

[0109] An embodiment of the present invention provides a sample cleaning device for training a fan icing prediction model. The fan icing prediction model used by the device is composed of a timing prediction model and a prediction model for predicting fan icing phenomena. The device includes: an acquisition module, used to acquire samples to be cleaned during the operation of the fan; wherein the samples to be cleaned include at least two factors; a determination module, used to use the timing prediction model to determine the timing coupling between different factors; and a cleaning module, used to clean the samples to be cleaned based on the timing coupling between the different factors.

[0110] In some possible implementations, the acquisition module is also used to obtain geographic location information data, meteorological element data, and icing shutdown data of the wind turbine during operation; and determine at least two factors based on the geographic location information data, the meteorological element data, and the icing shutdown data.

[0111] In some possible implementations, the acquisition module is further used to perform data cleaning on the geographic location information data, the meteorological element data, and the icing shutdown data to determine at least two factors.

[0112] In some possible implementations, the determination module is further used to remove any row element in the training set of the samples to be cleaned to obtain a training sequence; based on the training sequence, the prediction model to be trained is trained to obtain a time series prediction model for any row element.

[0113] In some possible implementations, the determination module is also used to input each factor in the training set into the time series prediction model to obtain a predicted value; determine the contribution of each factor to the predicted value; based on the contribution of each factor to the predicted value, determine the similarity between the contributions corresponding to different factors; and fuse the similarity and the contribution to obtain the time series coupling between different factors.

[0114] In some possible implementations, the determination module is further used to determine the similarity between the contribution degrees corresponding to any two factors based on the contribution degree of each factor to the predicted value; and to fuse the similarity between the contribution degrees corresponding to any two factors to obtain the similarity between the contribution degrees corresponding to the different factors.

[0115] In some possible implementations, the cleaning module is further used to determine edge samples in the samples to be cleaned; determine neighboring samples corresponding to the edge samples based on the timing coupling between different factors in the edge samples; interpolate and synthesize the edge samples based on the neighboring samples to obtain synthetic samples; and clean the samples to be cleaned based on the synthetic samples.

[0116] In some possible implementations, the cleaning module is further configured to cluster the samples to be cleaned to obtain a clustering result; and determine edge samples in the samples to be cleaned based on the clustering result.

[0117] In some possible implementations, the cleaning module is further used to determine the decoded neighbor distance between the edge sample and the candidate sample based on the temporal coupling between different factors in the edge sample; wherein the candidate sample is a generated sample corresponding to a preset cluster in the clustering result; based on the decoded neighbor distance, determine the neighboring sample corresponding to the edge sample.

[0118] In some possible implementations, the cleaning module is further configured to use the neighboring samples to perform interpolation on different factors in the edge samples to obtain interpolation results; and synthesize the interpolation results to obtain the synthesized samples.

[0119] Optionally, the transmission medium may be a wired link (such as but not limited to coaxial cable, optical fiber and digital subscriber line (DSL), etc.) or a wireless link (such as but not limited to wireless Fidelity (WIFI), Bluetooth and mobile device network, etc.). It should be noted that the apparatus provided in the above embodiment is only illustrated by the division of the above functional modules. In actual application, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device is divided into different functional modules to complete all or part of the functions described above.

[0120] Figure 6 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 6As shown, the electronic device 600 includes: a memory 601, a processor 602, and a computer program 603 stored in the memory 601 and running on the processor 602, wherein when the processor 602 executes the computer program 603, the device can execute any sample cleaning method for wind turbine icing prediction model training described above. In addition, the embodiment of the present invention also protects a device, which may include a memory and a processor, wherein an executable program code is stored in the memory, and the processor is used to call and execute the executable program code to execute the sample cleaning method for wind turbine icing prediction model training provided by the embodiment of the present invention. In this embodiment, the functional modules of the device can be divided according to the above method example, for example, each functional module can be corresponded, or two or more functions can be integrated into one processing module, and the above integrated module can be implemented in the form of hardware. It should be noted that the division of modules in this embodiment is schematic, which is only a logical function division, and there may be other division methods in actual implementation. It should be noted that all relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module, and will not be repeated here.

[0121] It should be understood that the device provided in this embodiment is used to execute the above-mentioned sample cleaning method for training the wind turbine icing prediction model, so the same effect as the above-mentioned implementation method can be achieved. In the case of an integrated unit, the device may include a processing module and a storage module. Among them, when the device is applied to a device, the processing module can be used to control and manage the actions of the device. The storage module can be used to support the device to execute mutual program codes, etc. Among them, the processing module can be a processor or a controller, which can implement or execute various exemplary logic boxes, modules and circuits described in combination with the contents disclosed in the present invention. The processor can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc., and the storage module can be a memory.

[0122] In addition, the device provided in the embodiment of the present invention may be specifically a chip, a component or a module, and the chip may include a connected processor and a memory; wherein the memory is used to store instructions, and when the processor calls and executes the instructions, the chip may execute the sample cleaning method for wind turbine icing prediction model training provided in the above embodiment. This embodiment also provides a computer-readable storage medium, in which a computer program code is stored, and when the computer program code is run on a computer, the computer executes the above-mentioned related method steps to implement the sample cleaning method for wind turbine icing prediction model training provided in the above embodiment.

[0123] This embodiment also provides a computer program product. When the computer program product is run on a computer, the computer executes the above-mentioned related steps to implement the sample cleaning method for training the wind turbine icing prediction model provided in the above embodiment. Among them, the device, computer-readable storage medium, computer program product or chip provided in this embodiment are all used to execute the corresponding method provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method provided above, and will not be repeated here. Through the description of the above implementation mode, the technicians in the relevant field can understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In practical applications, the above-mentioned function allocation can be completed by different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In the embodiments provided by the present invention, it should be understood that the disclosed device and method can be implemented in other ways. For example, the device embodiment described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices or units, which may be electrical, mechanical or other forms. It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-tasking and parallel processing are also possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same and similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. The above content is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention.

Claims

1. A sample cleaning method for training a wind turbine icing prediction model, characterized in that: The wind turbine icing prediction model used in the method is composed of a time series prediction model and a prediction model for predicting wind turbine icing phenomena, and the method comprises: Acquire a sample to be cleaned during the operation of the fan; wherein the sample to be cleaned includes at least two factors; Using the time series prediction model, determining the time series coupling between different factors; Based on the temporal coupling between the different factors, cleaning the sample to be cleaned; Obtain at least two factors during the operation of the fan, including: Acquiring geographic location information data, meteorological element data, and icing shutdown data of the wind turbine during operation; Determining at least two factors based on the geographic location information data, the meteorological element data, and the icing shutdown data; The method further comprises: Eliminate any row element in the training set of the sample to be cleaned to obtain a training sequence; Based on the training sequence, training the prediction model to be trained to obtain a time series prediction model for any row element; The adopting of the time series prediction model to determine the time series coupling between different factors includes: Input each factor in the training set into the time series prediction model to obtain a predicted value; Determining the contribution of each of the factors to the predicted value; Based on the contribution of each factor to the predicted value, determining the similarity between the contribution of different factors; The similarity and the contribution are combined to obtain the temporal coupling between different factors.

2. The sample cleaning method for wind turbine icing prediction model training according to claim 1 is characterized in that: The determining of at least two factors based on the geographic location information data, the meteorological element data, and the icing shutdown data includes: The geographic location information data, the meteorological element data and the icing shutdown data are cleaned to obtain at least two factors.

3. The sample cleaning method for wind turbine icing prediction model training according to claim 1, characterized in that: The determining, based on the contribution of each factor to the predicted value, the similarity between the contribution corresponding to different factors includes: Based on the contribution of each factor to the predicted value, determining the similarity between the contribution corresponding to any two factors; The similarities between the contribution degrees corresponding to the arbitrary two factors are merged to obtain the similarities between the contribution degrees corresponding to the different factors; The step of fusing the similarities between the contribution degrees corresponding to any two factors comprises: Use clustering algorithm to divide all samples in the training set into multiple clusters; The sequence consisting of the maximum value, minimum value, mean value, and third quartile of all contributions of any two factors in each cluster is used as the contribution feature sequence of any two factors in each cluster; Determine the sum of the similarity measurement results between the contribution feature sequences of any two factors among all clusters, and obtain the similarity between the contributions corresponding to different factors.

4. The sample cleaning method for wind turbine icing prediction model training according to claim 1, characterized in that: The step of cleaning the sample to be cleaned based on the timing coupling between the different factors includes: Determining edge samples among the samples to be cleaned; Determining adjacent samples corresponding to the edge samples based on the temporal coupling between different factors in the edge samples; Based on the neighboring samples, interpolating and synthesizing the edge samples to obtain synthesized samples; Based on the synthetic sample, cleaning the sample to be cleaned; Determining edge samples among the samples to be cleaned includes: Clustering the samples to be cleaned to obtain clustering results; Based on the clustering result, edge samples in the samples to be cleaned are determined.

5. The sample cleaning method for wind turbine icing prediction model training according to claim 4 is characterized in that: The determining, based on the temporal coupling between different factors in the edge sample, the neighboring sample corresponding to the edge sample comprises: Based on the temporal coupling between different factors in the edge sample, determining the decoding neighbor distance between the edge sample and the candidate sample; wherein the candidate sample is a generated sample corresponding to a preset cluster in the clustering result; Based on the decoded neighbor distance, a neighboring sample corresponding to the edge sample is determined.

6. The sample cleaning method for wind turbine icing prediction model training according to claim 4, characterized in that: The interpolating and synthesizing the edge samples based on the neighboring samples to obtain synthesized samples includes: Using the neighboring samples to perform interpolation on different factors in the edge samples to obtain an interpolation result; The interpolation results are synthesized to obtain the synthesized samples.

Images