Multi-tower combined wind measurement data cleaning method based on SA-TCN
By using SA-TCN data interpolation network in air measurement data interpolation, integrating multi-tower data and taking into account multi-sequence correlation, the problem of low accuracy of air measurement data interpolation in the prior art and difficulty in adapting to non-random missing mode is solved, and more efficient and accurate air measurement data processing is achieved.
Patent Information
- Application Number
- CN202411618055.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-13
- Publication Date
- 2025-03-07
AI Technical Summary
The prior art fails to effectively consider the autocorrelation and periodicity of the wind measurement data during the interpolation process of wind measurement data, as well as the multi-sequence correlation between multi-swing towers and multi-height layers, resulting in limited interpolation accuracy and difficulty in adapting to non-random missing mode.
The multi-tower joint wind measurement data cleaning method based on SA-TCN is adopted to construct a feature matrix by fusing the wind speed, wind direction, temperature and air pressure data of the multi-tower, and data interpolation is performed using the SA-TCN data interpolation network. The method includes a self-attention mechanism and a time convolutional neural network, which can consider multivariate timing correlations and process missing and abnormal data through linear interpolation and 3sigma anomaly detection methods.
It improves the accuracy of interpolation of air measurement data, can effectively identify and process abnormal data, and is suitable for multi-tower and multi-height layers of air measurement data processing, especially in non-random missing mode, showing high interpolation quality.
Smart Images
Figure CN119577327A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of data processing, and in particular relates to a multi-tower joint wind measurement data cleaning method based on SA-TCN. Background Art
[0002] Due to the uncertainty and randomness of wind energy resources, before the development of wind power, it is necessary to conduct in-depth analysis of the long-term wind measurement data in the region and objectively understand the wind resources to control the feasibility and economy of wind power projects. In wind resource assessment, wind measurement data is often affected by freezing, lightning, complex terrain, uncontrollable human factors, etc., and wind measurement data often has missing measurements and abnormalities. Therefore, it is necessary to use data cleaning methods such as quality control and data interpolation to meet the requirements of integrity, rationality, and trend of wind measurement data.
[0003] Faced with the lack of wind measurement data, domestic and foreign scholars have also studied the interpolation of wind measurement data, which is mainly divided into statistical fitting method interpolation and time series prediction method interpolation. In terms of wind measurement data interpolation based on statistical fitting methods, engineering applications are relatively extensive. At present, the MCP (Measure-Correlate-Predict) interpolation method is mainly used. The existing technology is based on the wind measurement data interpolation method of reanalysis data [J]. Journal of Huazhong University of Science and Technology (Natural Science Edition) analyzed the correlation between mesoscale data and wind measurement data, and proposed an interpolation method based on MERRA reanalysis data; Research on the uncertainty of wind farm wind measurement data interpolation [J]. Renewable Energy, using the mesoscale data source of MERRA2 as the reference wind speed for the missing wind measurement data, studied the least squares method (LLS) and total least squares method (TLS) and variance ratio (VR) interpolation accuracy comparison; Research on the identification and completion method of abnormal wind speed data of wind towers [J]. Renewable Energy, a data completion method based on attribute importance and similar fragments is proposed to complete the missing wind measurement data from the perspective of clustering; In terms of time series prediction, the existing technology studies the interpolation accuracy of ARIMA model with Kalman smoothing and multivariate linear model in meteorological data; In order to fill the missing wind measurement data, the interpolation performance of two neural networks, LSTM and ARIMA, is compared; In view of the nonlinear correlation between the data of the wind tower to be interpolated and the reference tower or meteorological station, a GA-ELM model for wind farm wind measurement data interpolation is proposed.
[0004] In existing studies, statistical fitting methods reconstruct wind data based on the overall correlation between wind data and reference data to interpolate missing data. The autocorrelation and periodicity of wind data are not considered, and the accuracy of interpolation is limited. In terms of time series prediction, although the autocorrelation and periodicity of data are considered, the multivariate series correlation between multiple wind towers and multiple height layers is not considered. The interpolation of wind data should be carried out through the same-tower correlation interpolation, different-tower correlation interpolation, and reanalysis data correlation interpolation. After long-term meteorological station data correction, representative annual wind data is obtained. In the process of wind data processing, interpolation needs to be performed for multivariate time series. The existing technology summarizes the types of multivariate time series interpolation methods, but the interpolation mode is for the missing completely at random mode (MCAR), which is not applicable to the non-random missing mode (MNAR) that often appears in wind data. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a multi-tower joint wind measurement data cleaning method based on SA-TCN to solve the abnormal recognition problem of wind measurement data existing in the prior art and improve the accuracy of missing data interpolation.
[0006] Technical solution of the present invention:
[0007] A SA-TCN-based multi-tower joint wind measurement data cleaning method, the method comprising: combining the data of each wind measurement tower, and fusing the wind speed, wind direction, temperature and air pressure data types of the wind measurement towers, fusing the 10-minute resolution data of each wind measurement tower into a feature matrix, using linear interpolation to process the mesoscale reanalysis data of each wind measurement tower from a 1-hour resolution to a 10-minute resolution, and inputting the two data into a SA-TCN data interpolation network as feature matrices, and outputting reconstructed complete wind measurement data.
[0008] The method also includes: using a 3sigma anomaly test method to identify discrete data anomalies on the original data, where μ is the average value of the reconstructed data, τ is the standard deviation between the original data and the reconstructed data, and when the original data exceeds the 3sigma distribution, it is determined to be abnormal data, and finally the reconstructed data is used to recover the abnormal and missing data to obtain complete representative annual wind measurement data of multiple towers.
[0009] The SA-TCN data interpolation network includes two SA-TCN modules and a self-attention weight layer. The input of the SA-TCN interpolation model is the feature matrix and the missing position matrix, and the interpolated feature matrix is output. Each SA-TCN module contains a self-attention mechanism and a TCN convolution layer, and outputs the calculated feature matrix through a linear layer, and inputs the feature matrix to be interpolated. and the masked missing position matrix In the first SA-TCN module it is represented as:
[0010]
[0011] Z2={TCN(Attention mask (Z1)⊙Z1)} N (9)
[0012]
[0013] In the second SA-TCN module, the result X1 of the previous interpolation and the masked missing position matrix are input The calculation process is expressed as:
[0014]
[0015] S2={TCN(Attention mask (S1)⊙S1)} N (13)
[0016]
[0017] In the self-attention weight layer, the attention coefficient is the average of the multi-head attention coefficients in the second SA-TCN module, and introduces Calculate the self-attention weighted coefficient matrix η, the calculation process is expressed as:
[0018]
[0019] In the self-attention weight layer, the output X2 in the second SA-TCN module and the self-attention weight coefficient matrix η are input, and the interpolation result X3 in the self-attention weight layer is output. The calculation process is expressed as:
[0020] X3=(1-η)⊙X1+η⊙X2(17)
[0021] In the SA-TCN data interpolation network, the final interpolation result is expressed as
[0022] During the training process, two loss functions are set to control the model reconstruction performance and interpolation performance. The loss function is expressed as:
[0023]
[0024] ζ MIT = l MAE (X c ,X,I) (20)
[0025] ζ=ζ ORT+λζ MIT (twenty one).
[0026] The reconstruction data performance and interpolation missing performance of the training model are improved by using the joint optimization method according to the missing position matrix and the mask position matrix. The two training strategies of interpolation missing data MIT and reconstruction observed data ORT are used to train the model. MIT and ORT Two loss functions, the loss function is constructed using the mean absolute error function MAE, and the two loss functions constitute the total loss function l for model training. λ is the weight parameter for adjusting the two learning modes.
[0027]
[0028] l = ζ MIT +λζ ORT (twenty four).
[0029] When establishing the model, the effective data measured by the wind tower is used as the output data. Then, according to the anomalies and missing types of the wind measurement data, the output data is non-randomly contaminated and randomly contaminated, and used as the input data. This kind of mapping data is used as the training data set and the test data set, with a ratio of 6:4.
[0030] Beneficial effects of the present invention:
[0031] The wind data interpolation method considering multivariate time series correlation of the present invention combines the multi-height layer data of the wind tower to be interpolated, the surrounding wind tower data with good correlation and the mesoscale reanalysis data, and constructs a wind data interpolation method for the same-tower interpolation, different-tower interpolation and mesoscale interpolation in parallel based on the self-attention multivariate time series prediction model, which has the following advantages:
[0032] The multivariate self-attention mechanism of the present invention can take into account multi-dimensional wind measurement data and distribute weights according to the strength of correlation, so as to effectively restore long-term wind measurement data.
[0033] According to the certain correlation between wind measurement data from different wind measurement towers and at different heights, the present invention extracts the features between these wind measurement data, can effectively solve the problem of abnormal identification of wind measurement data, and improve the accuracy of missing data interpolation.
[0034] The multivariate time series wind measurement data cleaning method proposed in the present invention can extract relevant features of multi-tower wind measurement data. Compared with the MCP method that considers single data in engineering, it can perform anomaly detection on wind measurement data, and the accuracy of data interpolation is higher.
[0035] After combining the self-attention mechanism and the temporal convolutional neural network, the present invention can improve the ability to extract spatiotemporal features of data and show good interpolation performance in data structures with different completeness rates.
[0036] The invention solves the problem of abnormal identification of wind measurement data existing in the prior art and improves the accuracy of missing data interpolation. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is a schematic diagram of the causal expansion convolutional network structure in a specific implementation manner;
[0038] Figure 2 It is a schematic diagram of the residual connection network structure in a specific implementation manner;
[0039] Figure 3 It is a schematic diagram of the SA-TCN data cleaning network in a specific implementation mode;
[0040] Figure 4 It is a schematic diagram of the multi-wind tower joint data cleaning process in a specific implementation method;
[0041] Figure 5 It is a schematic diagram of data pollution in a specific implementation mode;
[0042] Figure 6 A location diagram of a wind tower in a specific implementation manner;
[0043] Figure 7 This is a schematic diagram of abnormal wind measurement data identification in a specific implementation manner;
[0044] Figure 8 This is a schematic diagram of identifying randomly missing wind measurement data in a specific implementation manner;
[0045] Fig. 9 This is a schematic diagram of identifying non-randomly missing wind measurement data in a specific implementation manner. DETAILED DESCRIPTION
[0046] Multivariate Time Series Definition of Missing States
[0047] Assume that the multivariate time series contains two dimensions: feature and time. D features can be represented as a time series {x 1 ,x 2 ,...,x t When there are missing data in a multivariate time series, the specific location of the missing data in the multivariate time series can be represented by the Mask matrix M:
[0048]
[0049] There are missing data in the original data. It is impossible to directly obtain the missing pattern and then interpolate and evaluate the model performance. Therefore, it is necessary to preprocess the Mask matrix of the original data. The Mask matrix M is mainly based on the original Mask matrix M, and the non-missing data is randomly masked according to a certain ratio to obtain Get the missing position matrix that mixes the original missing information and the mask information And the mask position matrix I, where the mask position matrix can be expressed as:
[0050]
[0051] Self-Attention
[0052] The attention mechanism is a method to quickly and reasonably assign weights to feature information under specific conditions. In the process of wind data interpolation, due to the existence of multivariate time series data from multiple data sources and multiple height layers, and the existence of inconsistent correlations, the self-attention mechanism is used to assign reference weights to multivariate data. The self-attention mechanism can be expressed as:
[0053]
[0054] The self-attention mechanism function calculates the attention relationship between Q and V when inputting the query element, key value, and the vector Q, K, and V corresponding to the key value. Q represents the multivariate data dimension d k The query element,
[0055] In order to improve the performance of the interpolation model, SAITS introduced a diagonal mask matrix in the self-attention mechanism, ignoring the feature's own attention coefficient and improving the performance of data interpolation. The mask matrix in the attention mechanism and the improved attention mechanism can be expressed as:
[0056]
[0057] Temporal Convolutional Neural Network (TCN)
[0058] Wind measurement data is a kind of time series data, so there is autocorrelation in the time dimension. In order to capture the time autocorrelation, it is necessary to consider the time series related neural network architecture. The recursive neural network architecture in deep learning can effectively obtain the long-term dependencies of the time series, but it is highly dependent on the forward time information. When dealing with large-scale tasks such as interpolating missing long-term wind measurement data, it has low efficiency and poor reliability. Therefore, the present invention introduces TCN, which adopts the convolution method, provides efficient parallel computing, is suitable for large-scale data processing, and can increase the receptive field of the convolution kernel, and can capture the long-term and short-term dependencies in the time series data. TCN consists of causal dilation convolution and residual module. Causal dilation convolution is as follows: Figure 1 shown.
[0059] As the number of hidden layers in the TCN network increases, the computational complexity of the network increases. In order to avoid the increase in computational complexity and the explosion of historical memory information, dilated convolution is introduced into the network. The formula is expressed as:
[0060]
[0061] in, For each time series input, is the convolution kernel, where k is the size of the convolution kernel. The input time series data is s, and the output result of the dilated convolution layer is F(s). s-di is the result of the current dilated convolution layer obtaining the next layer of the network based on the previous layer of the network. d is the dilation factor, which is expanded exponentially by 2, d = 1, 2, 4, ..., and the problem of computational redundancy in the complex structure of causal convolution can be improved by changing the dilation factor of different network layers. By controlling the size of the dilation factor d, the TCN network can control the depth of the neural network and obtain a larger receptive field, where the receptive field is defined as field = (k-1)d.
[0062] After the residual connection passes through a neural network layer of a certain depth, a better calculation result is obtained. In order to prevent overfitting, an identity mapping is performed in the subsequent network to keep the output result of the entire neural network at the optimal solution. The formula for the residual connection is:
[0063] TCN(s)=σ(s+F(s)) (7)
[0064] Where σ is the activation function, the output of the first layer of the network is set as the input of the residual module, and the input of the residual module is defined as s, F(s) is the residual, which means the difference between the output and input of the network in the residual connection. As shown in the figure, the application of the residual module in the TCN network is added to each convolution layer. A residual connection is added until all residual modules are calculated and the final result is output. Figure 2 It is the structure of the residual connection layer.
[0065] SA-TCN data interpolation network structure
[0066] The SA-TCN interpolation model consists of two SA-TCN modules and a self-attention weight layer. The input of the SA-TCN interpolation model is the feature matrix and the missing position matrix, and the output is the interpolated feature matrix. Each SA-TCN module contains a self-attention mechanism and a TCN convolution layer, and outputs the calculated feature matrix through a linear layer. The model structure is as follows: Figure 3 shown.
[0067] Input the feature matrix to be interpolated and the masked missing position matrix In the first SA-TCN module, it can be expressed as:
[0068]
[0069] Z2={TCN(Attentionmask (Z1)⊙Z1)} N (9)
[0070]
[0071] in p 1 is the first layer weight parameter of the first SA-TCN module, is the weight parameter of the second layer.
[0072] In the second SA-TCN module, the result X1 of the previous interpolation and the masked missing position matrix are input The calculation process can be expressed as:
[0073]
[0074] S2={TCN(Attention mask (S1)⊙S1)} N (13)
[0075]
[0076] in p 2 is the first layer weight parameter of the second SA-TCN module, is the weight parameter of the second layer, w γ 、b γ is the weight parameter of the output layer. In the self-attention weight layer, the attention coefficient is the average of the multi-head attention coefficients in the second SA-TCN module, and introduces Calculate the self-attention weighted coefficient matrix η. The calculation process can be expressed as:
[0077]
[0078] where w η 、b η is the weight parameter of the self-attention weight layer.
[0079] Therefore, in the self-attention weight layer, the output X2 in the second SA-TCN module and the self-attention weight coefficient matrix η are input, and the interpolation result X3 in the self-attention weight layer is output. The calculation process can be expressed as:
[0080] X3=(1-η)⊙X1+η⊙X2(17)
[0081] In the SA-TCN data interpolation network, the final interpolation result can be expressed as
[0082]
[0083] During the training process, two loss functions are set to control the model reconstruction performance and interpolation performance. The loss function can be expressed as:
[0084]
[0085] ζ MIT = l MAE (X c ,X,I) (20)
[0086] ζ=ζ ORT +λζ MIT (twenty one)
[0087] Among them l MAE is the loss function composed of variance.
[0088] Aiming at the missing data in the wind tower, the data of multiple wind towers are combined, and the wind speed, wind direction, temperature, and air pressure data of the wind towers are integrated. Combined with the data of mesoscale reanalysis, a wind data cleaning method considering multivariate time series correlation is proposed, including: first, the 10-minute resolution data of multiple wind towers are integrated into a feature matrix, and the mesoscale reanalysis data of each wind tower is processed from 1h resolution to 10min resolution using linear interpolation, and the two are input into the SA-TCN data interpolation network as feature matrices, and the reconstructed complete wind data is output. The 3sigma anomaly test method is used to identify the anomalies of discrete data on the original data, where μ is the mean value of the reconstructed data, and τ is the standard deviation of the original data and the reconstructed data. When the original data exceeds the 3sigma distribution, it is judged as abnormal data. Finally, the reconstructed data is used to restore the abnormal and missing data, and the complete representative annual wind data of multiple towers is obtained.
[0089]
[0090] In order to train the model's reconstruction data performance and interpolation missing performance, a joint optimization method is used to train the model based on the missing position matrix and mask position matrix mentioned above. Through the two training strategies of interpolation missing data (MIT) and reconstruction observed data (ORT), corresponding to ζ MIT and ORT Two loss functions,The loss function is constructed using the mean absolute error function (MAE), and the two loss functions constitute the total loss function l of model training, and λ is the weight parameter for adjusting the two learning modes.
[0091]
[0092] l = ζ MIT +λζ ORT (twenty four)
[0093] When establishing the model, the effective data measured by multiple wind towers are used as the output data. Then, according to the abnormal and missing types that often appear in the wind measurement data, the output data is subjected to non-random and random pollution to form a polluted data set, such as Figure 5 As shown, the contaminated dataset is used as input data, and the two datasets are divided into training dataset and test dataset with a ratio of 6:4.
[0094] Case analysis:
[0095] The present invention takes the wind measurement data in a mountain wind farm in southwest China as an example. There are two wind measurement towers in the wind farm. The wind measurement time periods and locations are shown in Table 1 and Table 2. Figure 6 shown.
[0096] Table 1 On-site wind tower observation schedule
[0097]
[0098]
[0099] Correlation analysis of multivariate wind measurement data
[0100] Taking the 1# wind tower as the target tower, the wind speed correlation analysis at different height layers of the wind tower, the wind speed correlation analysis at the height of the pre-installed hub of different towers and the wind speed correlation analysis of the mesoscale reanalysis data were carried out.
[0101] As the height spacing increases, the correlation between different height layers gradually weakens. At each observation point, the mesoscale reanalysis data of that point can be obtained. The data sources include MERRA2 and ERA5 data sources. According to the degree of correlation with the wind measurement point, the data source with higher correlation is selected as the reference data, and the mesoscale wind speed correlation is the smallest. Therefore, in the calculation process, it is necessary to assign weights to the data of different height layers according to the correlation.
[0102] Table 21# Wind speed autocorrelation analysis of wind tower
[0103]
[0104] Comparing the correlation of wind speed at 120m of 1#, 2# and 3# wind towers, the correlation of wind speed does not decrease with increasing distance, but is related to local terrain conditions and altitude.
[0105] Table 3 Correlation analysis between wind speeds of 3 wind towers
[0106]
[0107] According to the above correlation analysis, the data between the same towers usually have the highest correlation because they have the same wind measurement environment and instruments; the data between different towers are second only to the measured data; the mesoscale reanalysis data is simulated data, and the correlation is lower, but when the measured data quality is poor, the mesoscale data will also show a good correlation. Therefore, in the interpolation process, it is necessary to assign weights to different data sources and data at different altitudes according to the correlation.
[0108] Analysis of wind measurement data cleaning results
[0109] By jointly modeling the three wind towers, after obtaining the model through the training set, the test set was used to verify the effectiveness of the model. The data in the test set was cleaned through data reconstruction, anomaly identification, and data interpolation. In the result analysis, only the pollution data was compared, and the RMSE (root mean square error) was used for evaluation.
[0110] Data anomaly identification
[0111] By jointly modeling the three wind towers, the original data of the wind towers were cleaned through data reconstruction, anomaly identification, and data interpolation. In the abnormal data identification, the wind data can be effectively identified, such as the wind data stiffness caused by freezing, which obviously exceeds the reasonable threshold, and the large difference in wind speed between adjacent height layers.
[0112] Missing data interpolation
[0113] The present invention sets up a control experimental group for the two missing modes of MCAR and MNAR, and compares them with the least squares (LLS) interpolation method in the MCP method, and uses the root mean square error to evaluate the accuracy of wind speed data interpolation.
[0114] Table 4: Comparison table of the present invention and the LLS interpolation method
[0115]
[0116] When data is randomly missing, the LLS method fits the missing data by comparing the correlation between two variables. Although it can restore the statistical value of the data, it is difficult to capture the changes in time series. The TCN network can capture the time series rules of wind speed and improve the accuracy of the recovered data.
[0117] When data is missing for a long time, the LLS method only refers to single-dimensional wind measurement data to address the strong uncertainty of wind speed. The multi-dimensional self-attention mechanism can consider multi-dimensional wind measurement data and assign weights based on the strength of the correlation, which can effectively restore long-term wind measurement data.
[0118] Ablation experiment
[0119] In order to verify the effect of the improved model, commonly used neural network models and data sets with different completeness rates were used for verification.
[0120] Table 5: Data integrity rate comparison table
[0121]
[0122]
[0123] As the data completeness rate decreases, the calculation errors of the three models increase. This is because the model relies on the correlation between data, and large data missing will directly affect the accuracy of the calculation. Although the TCN network can extract the time series characteristics of wind measurement data, it ignores the correlation between features. Therefore, after introducing the self-attention mechanism, it can effectively improve the feature extraction ability of multivariate data, thereby improving the calculation accuracy.
Claims
1. A multi-tower joint wind measurement data cleaning method based on SA-TCN, characterized by: The method comprises: collecting data from each wind tower, fusing the wind speed, wind direction, temperature and air pressure data types of the wind towers, fusing the 10-minute resolution data of each wind tower into a feature matrix, using linear interpolation to process the mesoscale reanalysis data of each wind tower from a 1-hour resolution to a 10-minute resolution, and inputting the data into the SA-TCN data interpolation network as the feature matrix, and outputting the reconstructed complete wind measurement data.
2. The SA-TCN-based multi-tower joint wind measurement data cleaning method according to claim 1 is characterized in that: The method also includes: using a 3sigma anomaly test method to identify discrete data anomalies on the original data, where μ is the average value of the reconstructed data, τ is the standard deviation between the original data and the reconstructed data, and when the original data exceeds the 3sigma distribution, it is determined to be abnormal data, and finally the reconstructed data is used to recover the abnormal and missing data to obtain complete representative annual wind measurement data of multiple towers.
3. The SA-TCN-based multi-tower joint wind measurement data cleaning method according to claim 1 is characterized in that: The SA-TCN data interpolation network includes two SA-TCN modules and a self-attention weight layer. The input of the SA-TCN interpolation model is the feature matrix and the missing position matrix, and the interpolated feature matrix is output. Each SA-TCN module contains a self-attention mechanism and a TCN convolution layer, and outputs the calculated feature matrix through a linear layer, and inputs the feature matrix to be interpolated. and the masked missing position matrix In the first SA-TCN module it is represented as: Z2={TCN(Attention mask (Z1)⊙Z1)} N (9) In the second SA-TCN module, the result X1 of the previous interpolation and the masked missing position matrix are input The calculation process is expressed as: S2={TCN(Attention mask (S1)⊙S1)} N (13) In the self-attention weight layer, the attention coefficient is the average of the multi-head attention coefficients in the second SA-TCN module, and introduces Calculate the self-attention weighted coefficient matrix η, the calculation process is expressed as: In the self-attention weight layer, the output X2 in the second SA-TCN module and the self-attention weight coefficient matrix η are input, and the interpolation result X3 in the self-attention weight layer is output. The calculation process is expressed as: X3=(1-η)⊙X1+η⊙X2 (17) In the SA-TCN data interpolation network, the final interpolation result is expressed as During the training process, two loss functions are set to control the model reconstruction performance and interpolation performance. The loss function is expressed as: g=g ORT +lz MIT (21).
4. The SA-TCN-based multi-tower joint wind measurement data cleaning method according to claim 3 is characterized in that: The reconstruction data performance and interpolation missing performance of the training model are improved by using the joint optimization method according to the missing position matrix and the mask position matrix. The two training strategies of interpolation missing data MIT and reconstruction observed data ORT are used to train the model. MIT and ORT Two loss functions, the loss function is constructed using the mean absolute error function MAE, and the two loss functions constitute the total loss function l for model training. λ is the weight parameter for adjusting the two learning modes. l=ζ MIT +lz ORT (24).
5. The SA-TCN-based multi-tower joint wind measurement data cleaning method according to claim 4 is characterized in that: When establishing the model, the effective data measured by the wind tower is used as the output data. Then, according to the anomalies and missing types of the wind measurement data, the output data is non-randomly contaminated and randomly contaminated, and used as the input data. This kind of mapping data is used as the training data set and the test data set, with a ratio of 6:4.