Wind speed missing value completion method for wind farm based on mode meteorological data

Abstract

The application provides a wind speed missing value completion method based on mode meteorological data, and comprises the following steps: based on the obtained wind speed time series of a wind farm and corresponding mode meteorological multi-element data, a multivariate time series tensor is formed to construct sample data; based on the sample data, a time series interpolation model is trained by using a weighted loss function; the measured wind speed time series with missing values and corresponding mode meteorological multi-element data are input into the time series interpolation model, the missing position is predicted and completed, and through post-processing operation, the complete wind speed time series after completion is obtained. The beneficial effects of the application are as follows: by constructing an observation missing mask, designing a reasonable training strategy, and introducing a deep learning time series interpolation model, high-credibility wind speed completion can be realized under different missing scenarios, thereby effectively improving the accuracy and reliability of wind power prediction, wind resource assessment and wind farm operation management.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy data processing and machine learning, and particularly relates to a wind speed missing value completion method for a wind farm based on numerical weather prediction (NWP) data. BACKGROUND

[0002] In the operation process of a wind farm, wind speed is an important basic parameter for wind power prediction, equipment control and operation and maintenance. However, due to sensor failure, communication interruption or data acquisition anomaly, wind speed observation data inevitably has missing values, and the missing mode usually has irregularity and may have continuous missing for a long period of time. Such missing not only affects the accuracy of wind power prediction, but also brings potential risks to the operation scheduling and equipment maintenance of the wind farm.

[0003] Existing wind speed missing value completion methods mainly include linear interpolation, polynomial interpolation, mean filling and K-Nearest Neighbor (KNN) interpolation based on similar samples. In the processing of short-time and regular missing, they have certain effect, but in the face of long-time continuous missing or multi-variable simultaneous missing, the completion process often fails, and there is obvious deviation between the interpolation result and the true observation. In addition, they generally fail to fully utilize external meteorological data, and are difficult to reflect the real dynamics of the wind farm.

[0004] Numerical Weather Prediction (NWP) and reanalysis data can provide multi-element and multi-height meteorological information, and have the advantages of strong time continuity and wide spatial coverage. Using such model data can make up for the lack of information caused by observation missing to a certain extent. However, there is generally a systematic deviation between the model data and the actual observation, and if the observation data is directly replaced by the model data, error accumulation will often be introduced, which is difficult to meet the high-precision wind speed completion and subsequent business needs.

[0005] Therefore, there is an urgent need for a systematic method to use NWP data as external prior information and jointly model with limited measured observation data. SUMMARY

[0006] Therefore, the application aims to provide a wind speed missing value completion method for a wind farm based on NWP data, to solve the problems in the prior art that the completion process often fails in the face of long-time continuous missing or multi-variable simultaneous missing, and there is obvious deviation between the interpolation result and the true observation, and that external meteorological data is generally not fully utilized, and it is difficult to reflect the real dynamics of the wind farm.

[0007] To achieve the above-mentioned purpose, the technical scheme of the application is as follows:

[0008] The wind speed missing value completion method of a wind farm based on mode meteorological data comprises the following steps:

[0009] S1, based on the obtained wind farm measured wind speed time series and its corresponding mode meteorological multi-element data, a multivariate time series tensor is formed, and sample data is constructed based on the multivariate time series tensor;

[0010] S2, based on the sample data, a time series interpolation model is trained using a weighted loss function;

[0011] S3, the measured wind speed time series with missing values and its corresponding mode meteorological multi-element data are input into the time series interpolation model, the missing positions are predicted and completed, and through post-processing operation, the completed complete wind speed time series is obtained;

[0012] In step S1, based on the obtained wind farm measured wind speed time series and its corresponding mode meteorological multi-element data, a multivariate time series tensor is formed, comprising:

[0013] The mode meteorological multi-element data is time-aligned by resampling;

[0014] Based on the wind farm measured wind speed time series, a missing mask matrix is constructed;

[0015] The wind farm measured wind speed time series and the time-aligned mode meteorological multi-element data are normalized or standardized;

[0016] Time encoding is constructed, and based on the time encoding, the missing mask matrix, the normalized or standardized wind farm measured wind speed time series and the mode meteorological multi-element data, a multivariate time series tensor is formed.

[0017] Further, in step S1, based on the multivariate time series tensor, sample data is constructed, comprising:

[0018] The multivariate time series tensor is cut into sample data by using a sliding window.

[0019] Further, in step S2, based on the sample data, a time series interpolation model is trained using a weighted loss function, comprising:

[0020] By randomly simulating short-time random missing and long-time continuous missing for the training samples, a training target mask is constructed;

[0021] Based on the training target mask, a weighted loss function is constructed;

[0022] The weighted loss function is used to minimize the weighted reconstruction error.

[0023] Further, the expression of the weighted loss function is as follows:

[0024] ;

[0025] wherein, is a predicted wind speed, is a measured wind speed, is a mean square error (MSE) or a mean absolute error (MAE), , is a reconstruction weight, is a training target mask, t is a time step.

[0026] Further, in step S3, the post-processing operation includes:

[0027] a smoothing process;

[0028] a physical constraint correction;

[0029] a fusion check with observed short-time available values.

[0030] Further, the mode meteorological multi-element data includes at least one of the following elements: 2-meter temperature, 2-meter relative humidity, sea level pressure, 100-meter wind speed, 100-meter wind direction, 10-meter wind speed, 10-meter wind direction and 10-meter gust.

[0031] Further, the time series interpolation model is set as a time series interpolation model based on a self-attention mechanism or other deep learning time series interpolation model.

[0032] Compared with the prior art, the wind speed missing value completion method based on mode meteorological data has the following beneficial effects:

[0033] (1) The multi-element, multi-height mode meteorological multi-element data is standardized and systematized and introduced into the completion process, and the time continuity and spatial coverage characteristics thereof are utilized to provide reliable prior information for long missing sections, thereby significantly improving the completion accuracy.

[0034] (2) Through missing simulation and mask weighted loss in the training stage, the adaptability of the model to diversified missing modes (including long-time continuous missing) is enhanced.

[0035] (3) It has good expansibility, supports multi-site joint training, online incremental updating and access of mode data with different spatio-temporal resolutions.

[0036] (4) It can be directly and seamlessly embedded into the wind power prediction business process, reducing the prediction performance fluctuations caused by data missing. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiments of the application and assist in the explanation of the application. In the drawings:

[0038] Figure 1 The whole flowchart schematic diagram of the embodiment of the present application is shown in the figure;

[0039] Figure 2 The flowchart schematic diagram of the data organization and sample construction of the embodiment of the present application is shown in the figure;

[0040] Figure 3 The flowchart schematic diagram of the complete wind speed filling sequence of the embodiment of the present application is shown in the figure;

[0041] Figure 4 The completion effect schematic diagram of the wind speed missing segment of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0042] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0043] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" and the like are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features limited by "first", "second" and the like can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0044] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood through specific circumstances.

[0045] The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] The wind speed missing value completion method based on mode meteorological data includes the following steps:

[0047] A1, obtaining a measured wind speed time series of a wind farm and mode meteorological multi-element data corresponding to the time series;

[0048] A2, time aligning and preprocessing the mode meteorological multi-element data to obtain a time-consistent multivariate sequence, and normalizing or standardizing the multivariate sequence;

[0049] A3, constructing an observation availability mask to identify missing positions of the measured wind speed, and merging the measured wind speed, the multi-element sequence, and the mask in time sequence to form an input tensor;

[0050] A4, inputting the input tensor into a time series interpolation model, the time series interpolation model being a time series interpolation model based on a self-attention mechanism or other deep learning time series interpolation model, obtaining trained model parameters by simulating missing observation data in a training stage and optimizing using a mask-based weighted loss function;

[0051] A5, predicting and completing the missing positions of the measured wind speed based on the trained model, and obtaining a complete wind speed time series after completion through physical constraints and post-processing operations.

[0052] The mode meteorological multi-element includes at least one of the following elements: 2-meter temperature, 2-meter relative humidity, sea level pressure, 100-meter wind speed, 100-meter wind direction, 10-meter wind speed, 10-meter wind direction, and 10-meter gust.

[0053] The mode meteorological multi-element data is time-aligned by a resampling method to compensate for differences in time resolution and data integrity of the mode meteorological multi-element data.

[0054] In the training stage, short random missing and long continuous missing are simulated on training samples to construct a training target mask, so as to enhance the generalization ability of the model under different missing conditions.

[0055] A loss function based on missing position weighting is used in training.

[0056] The complete wind speed time series after completion can be applied to downstream businesses such as wind power prediction, wind resource assessment, or wind farm operation and maintenance, and can be post-processed, including but not limited to smoothing, physical constraint correction, and fusion verification with observed short-time available values, to improve prediction accuracy and application reliability.

[0057] The specific implementation is as follows:

[0058] B1, data acquisition and preprocessing

[0059] B11, obtaining a measured wind speed time series V of a wind farm and mode meteorological multi-element data M corresponding to the time, expressed as:

[0060] ;

[0061] ;

[0062] In the formula, This is a time series of measured wind speeds at a wind farm. For model meteorological multi-element data, The measured wind speed at the wind farm at time step t. For model meteorological elements at time step t Data, This represents the total time step. =8 indicates the total number of meteorological elements in the model, including: 2-meter temperature, 2-meter relative humidity, sea level pressure, 100-meter wind speed, 100-meter wind direction, 10-meter gust, 10-meter wind speed, and 10-meter wind direction. If other meteorological layer data or reanalysis data are available, the element dimensions can be flexibly expanded.

[0063] B12. Considering that the sampling interval of model meteorological multi-element data is usually different from that of observation data, and there may be occasional missing data, a resampling method is used to align the model meteorological multi-element data in time, resulting in a completed time series, as shown in the following expression:

[0064] ;

[0065] In the formula, For time-aligned meteorological elements at time step t Data, For model meteorological elements at time step t Data, For time steps, For model meteorological elements, The total time step, This represents the total number of meteorological elements in the model.

[0066] B13. Construct a missing mask matrix for observed wind speeds. The missing mask matrix is ​​obtained directly from the integrity of the original data: when the observed wind speed at time step t exists and passes quality control (non-empty, non-anomaly), let When the wind speed at that moment is missing or determined to be an invalid value, let To indicate the absence of a target object, and based on this, in order to achieve self-supervised reconstruction learning during the training phase, a training target mask is constructed. The training target mask does not come from the actual missing values, but rather from... At the effective observation locations, a subset of time steps are selected as "pseudo-missing" locations according to a preset masking strategy (such as random sampling, continuous time-time masking, or proportional masking), and the corresponding... This indicates that the given location is the training target (and needs to be reconstructed), and the remaining locations are set as follows: .

[0067] B14. To improve model convergence and training stability, the observed wind speed and model meteorological elements are standardized, such as using Z-score standardization, as shown in the following expression:

[0068] ;

[0069] In the formula, The measured wind speed at the wind farm at time step t is the standardized value. For the standardized model meteorological elements at time step t Data, and Let these represent the mean and standard deviation of the observed wind speed on the training dataset, respectively. and and are the mean and standard deviation of the k-th meteorological element on the training set, respectively.

[0070] To improve the convergence of model training.

[0071] B15. Considering that wind speed observations may be affected by sensor errors or environmental interference, outlier detection and denoising steps are introduced. Thresholding methods, seasonal quantile tests, and Median Absolute Deviation (MAD) methods can be used to mark outliers, and wavelet denoising or moving median filtering can be combined to suppress short-term noise, improving the stability and accuracy of the input data.

[0072] B16. To improve the model's ability to perceive time periodicity and temporal location information, a time encoding PE is added, with the following expression:

[0073] ;

[0074] In the formula, This indicates that time step t is the th position in the position encoding vector. Sine positional encoding values ​​in each dimension This indicates that time step t is the th position in the position encoding vector. Cosine position encoding values ​​in each dimension; =0, 1, ..., d / 2-1 are the coding dimension indices, d is the total coding dimension, and 10000 is the scaling cardinality (empirical value). This coding can be concatenated with observed wind speed and model meteorological multi-element data features as model input.

[0075] B2. Sample Construction and Enhancement

[0076] B21. Combine the normalized observed wind speed, model meteorological elements, missing mask, and time code to form a multivariate time series tensor, as shown in the following expression:

[0077] ;

[0078] In the formula, Let be a multivariate temporal tensor at time step t. This is the position encoding at time step t.

[0079] B22. Construct training samples using the sliding window method, with each window having a length of [length missing]. time step

[0080] ;

[0081] In the formula, For the i-th training sample window, Training sample window The sequence starting tensor in the sequence, Training sample window The tensor at the end of the sequence in the sequence, For window length, This is the sliding step size, which can be adjusted according to the number of samples. This is the starting position index of the sliding window.

[0082] B23. During training, a simulated missing mask is generated based on the empirical distribution. (i.e., the training target mask), including short-term random missing data and long-term continuous missing data, is expressed as follows:

[0083] ;

[0084] In the formula, This represents the minimum interval for wind farm data. This represents the maximum length of missing segments supported by the device and can be modified according to actual business needs. For each sample, the training objective is to reconstruct the masked portion while retaining some true observation locations to supervise model generalization.

[0085] B24. For long missing samples, oversampling can be used, or the long missing position can be weighted in the loss function to avoid the model training being biased towards short missing cases.

[0086] B3. Interpolation Implementation

[0087] B31, will , , Input a temporal interpolation model and train it. The model can be a self-attention type temporal interpolation model or other existing deep learning temporal interpolation models.

[0088] B32、In the attention or input process, an attention mask mechanism should be used to prevent the model from obtaining the masked real values (i.e. mask in the attention logits for the positions set to ).

[0089] B33、The model output is the predicted sequence within the window.

[0090] B34、The model training objective is to minimize the weighted reconstruction error, expressed as follows:

[0091] ;

[0092] where, is the predicted wind speed, is the measured wind speed, is the mean square error (MSE) or mean absolute error (MAE), ∈{0,1} is the training objective mask, , is the reconstruction weight, and t is the time step.

[0093] At the same time, a smoothing regularizer and a parameter regularizer can be added to control the output smoothness and prevent overfitting. Gradient clipping, learning rate decay, or optimizers (such as Adam) can also be used during training to stabilize the training.

[0094] B4、Inference completion and post-processing

[0095] B41、In the inference stage, multiple overlapping predictions output by the sliding window are fused into a sequence using window weighting (triangular window or linear weight), expressed as follows:

[0096] ;

[0097] where, is the predicted wind speed at time step in the th window, is the predicted wind speed obtained by weighting multiple overlapping predictions, is the set of all windows covering time step t, is the fusion weight (the weighting coefficient of the th window at time step , commonly using a triangular window or linear weight, so that the prediction weight of the window center is high and the edge is low), and finally fused into a complete wind speed sequence, expressed as follows:

[0098] ;

[0099] where, is the interpolated complete wind speed sequence.

[0100] B42. By using methods such as non-negative truncation, consistency correction with multi-height wind speed (power law or logarithmic wind profile), and extreme value truncation, ensure that the completion results conform to the physical laws of wind speed.

[0101] B5. Downstream Applications and Extended Scenarios

[0102] B51. Input the completed wind speed sequence into the downstream model, including wind power prediction, wind resource assessment or wind farm operation and maintenance, to improve prediction accuracy and decision quality.

[0103] B52 can be expanded to multi-wind farm joint training, multi-regional model meteorological multi-element data input, multi-step prediction or iterative autoregressive prediction scenarios.

[0104] B53. It can support online incremental training or real-time updates of multi-element meteorological data input for the model, enabling the model to adapt to long-term operation of wind farms and seasonal meteorological changes.

[0105] Example 1:

[0106] like Figure 1 As shown, the overall process of the wind speed missing value completion method for wind farms based on model meteorological data is as follows:

[0107] This study uses partial data from a 15-minute interval measured wind speed dataset of a coastal wind farm from January 2023 to June 2025. The external model meteorological multi-element data is sourced from the Global Forecast System (GFS) provided by Open-Meteo, which offers wide spatial coverage and strong temporal continuity, providing reliable external references for missing observational data. The model meteorological elements selected in this embodiment include the following eight types:

[0108] (a) Temperature at 2 meters (T2m);

[0109] (b) 2m relative humidity RH;

[0110] (c) Sea level pressure Psl;

[0111] (d) Wind speed at 100 meters: WS100m;

[0112] (e) 100m wind direction WD100m;

[0113] (f) 10-meter gust GF10m;

[0114] (g) Wind speed at 10 meters WS10m;

[0115] (h) 10m wind direction WD10m;

[0116] These elements can characterize, to some extent, various physical processes that affect wind speed changes, such as thermal circulation driven by temperature gradients, large-scale wind fields caused by pressure field distribution, and near-surface turbulence and surface friction effects.

[0117] C1. Data Sample Construction

[0118] C11, such as Figure 2 As shown, the data organization and sample construction process strictly aligns the model meteorological multi-element data with the observed wind speed data according to timestamps, ensuring a one-to-one correspondence between the two at each time point. If the time resolution (e.g., 1 hour) of the model meteorological multi-element data is inconsistent with the observation resolution (15 minutes), cubic spline interpolation is used for resampling to obtain a sequence with consistent time intervals Δt=15 minutes, thereby eliminating the error caused by the inconsistent time resolution.

[0119] C12. Constructing a missing mask matrix The expression is as follows:

[0120] ;

[0121] This matrix is ​​used not only to distinguish between available and missing locations when inputting the model, but also as a supervision signal during training to ensure that the model only makes reconstruction predictions at missing locations, thus avoiding information leakage.

[0122] C13. Use Z-score normalization to normalize the observed wind speed and model meteorological multi-element data. The expression is as follows:

[0123] ;

[0124] In the formula, The mean, The standard deviation is the standard deviation. The purpose of standardization is to eliminate the dimensional differences between different physical quantities, ensuring that the model can learn features at a uniform scale during training.

[0125] C14, Add time location coding Then, the input variables are combined in chronological order into a multivariate temporal tensor. The expression is as follows:

[0126] ;

[0127] ;

[0128] C15. The long sequence is divided into training samples using a sliding window method. Each window is set to a length of N = 4096 time steps (approximately 42 days), and the step size S is 2048. These values ​​can be adjusted based on the number of samples. The window sample format is as follows:

[0129] ;

[0130] In the formula, For the i-th training sample window, For training samples The sequence starting tensor in the sequence, For training samples The tensor at the end of the sequence, For window length, This is the starting position index of the sliding window.

[0131] This construction method ensures that the model can capture long-term dependencies while avoiding excessive memory consumption caused by excessively long sequences.

[0132] C16. During training, additional simulated missing masks are generated. (i.e., training target mask), with a missing length of 1-1344 (corresponding to 15 minutes to 14 days) time steps. This strategy effectively improves the model's generalization ability under different missing modes, especially for scenarios such as long-term communication interruptions and equipment shutdowns common in wind farms.

[0133] C2. Model Training and Testing

[0134] C21. It adopts a time-series interpolation model based on self-attention mechanism. Its core advantage is that it can dynamically capture the correlation between multiple factors and the long-range dependence across time steps through attention weights, thereby avoiding the gradient decay problem of traditional RNN / LSTM models under long sequences.

[0135] C22. During training, the initial learning rate is set to 0.001, and cosine annealing is used to bring the learning rate to 1×10⁻⁶. -6 The optimizer is AdamW, with a total of 500 training epochs and an early stopping threshold of 20 epochs. The weight parameter is 5 (reconstruction term weights), and the regularization parameter is 1e. -5 The smoothing term coefficient γ=1e -3 This configuration ensures stable model convergence while effectively avoiding overfitting and improving recovery accuracy for scenarios with long missing values.

[0136] C3. Wind farm missing value completion based on model meteorological multi-element data

[0137] C31. Input the missing observed wind speed sequences and model meteorological multi-element data into the trained model to obtain the prediction results. To ensure physical plausibility, the completed values ​​need to undergo non-negative truncation, extreme value correction, and wind profile consistency correction to obtain the final predicted wind speed. ;

[0138] C32, such asFigure 3 As shown in the mask matrix , the complete wind speed imputation sequence can be obtained after replacing the missing positions with the predicted values , and the expression is as follows:

[0139] ;

[0140] As shown in the mask matrix Figure 4 , the wind speed missing segment completion effect, wherein the abscissa is the time date, the ordinate is the wind speed value (m / s), the blue line segment wind speed (Original) is the original missing wind speed data, and the yellow dashed line wind speed (Imputed-connected) is the imputed data. In the missing value completion process, the originally broken wind speed sequence can be effectively restored to a complete and continuous time sequence.

[0141] The completed wind speed sequence can be directly used as input for a wind power prediction model, a wind resource assessment platform, or a wind farm operation and maintenance system, thereby significantly improving the accuracy, stability, and reliability of downstream businesses.

[0142] Advantages and beneficial effects of the present application:

[0143] (1) The multi-element and multi-height mode meteorological multi-element data is standardized and systematized into the completion process, and the time continuity and spatial coverage characteristics thereof are utilized to provide reliable prior information for long missing segments, thereby significantly improving the completion accuracy.

[0144] (2) Through missing simulation and mask weighted loss in the training stage, the adaptability of the model to diversified missing modes (including long-time continuous missing) is enhanced.

[0145] (3) It has good expansibility, supports multi-site joint training, online incremental updating, and access of mode data with different spatial and temporal resolutions.

[0146] (4) It can be directly and seamlessly embedded into the wind power prediction business process, reducing the prediction performance fluctuations caused by data missing.

[0147] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for completing missing wind speed values ​​in wind farms based on model meteorological data, characterized in that: Includes the following steps: S1. Based on the obtained wind speed time series of wind farms and the corresponding model meteorological multi-element data, a multivariate time series tensor is formed, and sample data is constructed based on the multivariate time series tensor. S2. Based on the sample data, train the time series interpolation model using a weighted loss function; S3. Input the missing measured wind speed time series and its corresponding model meteorological multi-element data into the time series interpolation model, predict and fill in the missing positions, and obtain the complete wind speed time series after post-processing. In step S1, based on the acquired measured wind speed time series of the wind farm and its corresponding model meteorological multi-element data, a multivariate time series tensor is formed, including: A resampling method was used to align the multi-element meteorological data of the model over time. A missing mask matrix is ​​constructed based on the measured wind speed time series of wind farms. Normalize or standardize the measured wind speed time series and time-aligned model meteorological multi-element data of the wind farm; Construct a time code, and based on the time code, the missing mask matrix, the normalized or standardized wind speed time series of wind farms and model meteorological multi-element data, form a multivariate time series tensor; In step S2, based on the sample data, a time-series interpolation model is trained using a weighted loss function, including: A training target mask is constructed by randomly simulating short-term random missing data and long-term continuous missing data in the training samples. Construct a weighted loss function based on the training target mask; Minimize the weighted reconstruction error using a weighted loss function; The expression for the weighted loss function is as follows: ； In the formula, To predict wind speed, To measure the wind speed, Mean square error or mean absolute error , To rebuild the weights, ∈{0,1} is the training target mask, and t is the time step.

2. The method for completing missing wind speed values ​​in wind farms based on model meteorological data according to claim 1, characterized in that: In step S1, sample data is constructed based on multivariate time series tensors, including: A sliding window is used to segment the multivariate time series tensor into sample data.

3. The method for completing missing wind speed values ​​in wind farms based on model meteorological data according to claim 1, characterized in that: In step S3, the post-processing operations include: Smoothing; Physical constraint correction; Fusion verification with short-term available observation values.

4. The method for completing missing wind speed values ​​in wind farms based on model meteorological data according to claim 1, characterized in that: The model's multi-element meteorological data includes at least one of the following elements: 2-meter temperature, 2-meter relative humidity, sea level pressure, 100-meter wind speed, 100-meter wind direction, 10-meter wind speed, 10-meter wind direction, and 10-meter gust.

5. The method for completing missing wind speed values ​​in wind farms based on model meteorological data according to claim 1, characterized in that: The temporal interpolation model is set as a self-attention mechanism-based temporal interpolation model or other deep learning temporal interpolation models.

Images