Method, device and equipment for repairing wind power prediction input data and medium

Abstract

The application relates to a wind power prediction input data repairing method, device, equipment and medium. The method comprises the following steps: determining to-be-repaired data in an original data sequence; wherein the original data sequence comprises time sequence data sequences of various candidate types associated with a wind power scene; analyzing the original data sequence to obtain a data correlation relationship between the various candidate types and a change trend parameter corresponding to the time sequence data sequences of the various candidate types; for each candidate type, using an attention head bound with a variable channel of the candidate type to process the time sequence data sequence of the candidate type, type coding and physical attribute information to obtain an attention output of the candidate type; and repairing the to-be-repaired data according to the change trend parameter and the attention output corresponding to each candidate type and the data correlation relationship between the various candidate types to obtain a data repairing result. The method can improve the accuracy of wind power data repairing.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, and in particular to a method, apparatus, equipment and medium for repairing wind power prediction input data. Background Technology

[0002] With the continuous development of the wind power forecasting field, in order to improve the accuracy of wind power forecasting results, the wind power data on which the forecasting depends is generally repaired before the forecasting is conducted. Existing wind power data processing methods generally use methods such as mean imputation, linear interpolation, and moving average to handle missing values ​​in the wind power data.

[0003] However, while the above method can fill in missing values, it ignores the changing trends in wind power data, which reduces the accuracy of wind power data repair. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, equipment, and medium for repairing wind power forecast input data to address the aforementioned technical problems, thereby improving the accuracy of wind power data repair.

[0005] Firstly, this application provides a method for repairing wind power forecast input data, including:

[0006] In response to a repair request for the original data sequence, the data to be repaired in the original data sequence is identified; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios;

[0007] The original data sequence is analyzed to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type;

[0008] For each candidate type, an attention head bound to the variable channel of the candidate type is used to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type;

[0009] Based on the trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, the data to be repaired is repaired to obtain the data repair result.

[0010] In one embodiment, determining the data to be repaired in the original data sequence includes: determining the observation mask corresponding to each time series data in the original data sequence based on the validity of each time series data in the original data sequence; determining the anomaly mask corresponding to each time series data in the original data sequence based on the input data constraints of the wind power prediction model and the spatial consistency corresponding to each time series data in the original data sequence; determining the comprehensive repair mask corresponding to each time series data based on the observation mask and anomaly mask corresponding to each time series data; and determining the data to be repaired in the original data sequence based on the comprehensive repair mask corresponding to each time series data.

[0011] In one embodiment, the anomaly mask corresponding to each time series data in the original data sequence is determined based on the input data constraints of the wind power prediction model and the spatial consistency of each time series data in the original data sequence. This includes: performing anomaly evaluation on each time series data in the original data sequence based on the input data constraints of the wind power prediction model to obtain a basic anomaly score for each time series data; wherein, the input data constraints include standard value range constraints associated with standard input data, standard ramp rate constraints, standard power curve constraints, and standard variable linkage information constraints between each candidate type; determining a target anomaly score for each time series data based on the basic anomaly score of each time series data and the spatial consistency of each time series data in the original data sequence; and determining the anomaly mask corresponding to each time series data based on the relationship between the target anomaly score of each time series data and the score threshold.

[0012] In one embodiment, the original data sequence is analyzed to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type. This includes: determining the data correlation between each candidate type based on the time series data sequence of each candidate type; for each candidate type, analyzing the time series data sequence of the candidate type based on each preset trend dimension to obtain the change trend parameters corresponding to the time series data sequence of the candidate type; wherein each preset trend dimension includes a trend dimension under a preset duration, a periodic trend dimension, a local dynamic dimension, and a local deviation dimension.

[0013] In one embodiment, an attention head bound to a variable channel of a candidate type is used to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type. This includes: constructing the input information of the variable channel corresponding to the candidate type based on the temporal data sequence of the candidate type, the observation mask and the comprehensive repair mask corresponding to each temporal data in the temporal data sequence of the candidate type, and the type encoding and physical attribute information corresponding to the candidate type; and processing the input information using an attention head bound to the variable channel of the candidate type to obtain the attention output of the candidate type.

[0014] In one embodiment, an attention head bound to the variable channel of the candidate type is used to process the input information to obtain the attention output of the candidate type. This includes: processing the input information using an attention head bound to the variable channel of the candidate type to obtain the query matrix, key matrix, and value matrix corresponding to the candidate type; and performing attention processing on the prior bias matrix associated with the wind power scenario, as well as the query matrix, key matrix, and value matrix corresponding to the candidate type, to obtain the attention output of the candidate type; wherein the prior bias matrix includes the variable weights of each variable under the candidate type.

[0015] In one embodiment, the data to be repaired is processed according to the trend parameters and attention output corresponding to each candidate type, as well as the data association between each candidate type, to obtain the data repair result. This includes: decoding the trend parameters and attention output corresponding to each candidate type, as well as the data association between each candidate type, to obtain the initial repair value corresponding to the data to be repaired; and correcting the initial repair value based on the physical constraints under the preset wind power scenario to obtain the data repair result. The physical constraints include spatial consistency constraints and input data constraints of the wind power prediction model.

[0016] Secondly, this application also provides a device for repairing wind power forecast input data, comprising:

[0017] The data selection module is used to identify the data to be repaired in the original data sequence in response to a repair request for the original data sequence; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with the wind power scenario;

[0018] The data analysis module is used to analyze the original data sequence to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type;

[0019] The data processing module is used to process the temporal data sequence, type encoding, and physical attribute information of each candidate type using an attention head bound to the variable channel of the candidate type, and to obtain the attention output of the candidate type.

[0020] The result determination module is used to repair the data to be repaired based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, and to obtain the data repair result.

[0021] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0022] In response to a repair request for the original data sequence, the data to be repaired in the original data sequence is identified; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios;

[0023] The original data sequence is analyzed to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type;

[0024] For each candidate type, an attention head bound to the variable channel of the candidate type is used to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type;

[0025] Based on the trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, the data to be repaired is repaired to obtain the data repair result.

[0026] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0027] In response to a repair request for the original data sequence, the data to be repaired in the original data sequence is identified; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios;

[0028] The original data sequence is analyzed to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type;

[0029] For each candidate type, an attention head bound to the variable channel of the candidate type is used to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type;

[0030] Based on the trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, the data to be repaired is repaired to obtain the data repair result.

[0031] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0032] In response to a repair request for the original data sequence, the data to be repaired in the original data sequence is identified; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios;

[0033] The original data sequence is analyzed to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type;

[0034] For each candidate type, an attention head bound to the variable channel of the candidate type is used to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type;

[0035] Based on the trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, the data to be repaired is repaired to obtain the data repair result.

[0036] The aforementioned method, apparatus, equipment, and medium for repairing wind power forecast input data, on the one hand, by identifying the data to be repaired in the original data sequence and analyzing the original data sequence to obtain the data correlation between each candidate type and the trend parameters corresponding to the time series data sequences of each candidate type, can better extract the overall data correlation and data change trend of the wind power forecast input data; and, by using an attention head bound to the variable channel of the candidate type for each candidate type, processing the time series data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type, it can avoid mutual interference between the features of different variables and ensure the reliability of the attention output; on the other hand, by combining the trend parameters and attention output corresponding to the candidate type, as well as the data correlation between each candidate type, to repair the data to be repaired and obtain the data repair result, it can ensure that the data repair result is compatible with the change trend and data characteristics of the wind power forecast input data, thereby improving the accuracy of wind power data repair. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart illustrating a method for repairing wind power forecast input data in one embodiment;

[0039] Figure 2 This is a flowchart illustrating the process of determining the data to be repaired in one embodiment;

[0040] Figure 3This is a flowchart illustrating the process of determining an anomaly mask in one embodiment;

[0041] Figure 4 This is a flowchart illustrating the analysis of the original data sequence in one embodiment;

[0042] Figure 5 This is a schematic diagram of the process for determining attention output in one embodiment;

[0043] Figure 6 This is a flowchart illustrating the process of determining the data repair result in one embodiment;

[0044] Figure 7 This is a flowchart illustrating a method for repairing wind power forecast input data in another embodiment;

[0045] Figure 8 This is a structural block diagram of a wind power prediction input data repair device in one embodiment;

[0046] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] With the continuous development of wind power forecasting, in order to improve the accuracy of wind power forecasting results, the wind power data on which the forecasting relies (such as historical wind power, wind speed, wind direction, temperature, humidity, air pressure, pitch angle, yaw angle, speed, turbine status, adjacent turbine operating data, meteorological tower data, and numerical weather prediction data) are generally repaired before the forecasting is conducted. Existing wind power data processing methods generally employ mean imputation, linear interpolation, and moving average to handle missing values ​​in the wind power data.

[0049] However, while the above method can fill in missing values, it ignores the changing trends in wind power data, which reduces the accuracy of wind power data repair.

[0050] Based on this, in an exemplary embodiment, a method for repairing wind power forecast input data is provided. The method is illustrated using an application to a server as an example. Figure 1 As shown, the specific steps include:

[0051] S101, in response to a repair request for the original data sequence, determine the data to be repaired in the original data sequence.

[0052] The original data sequence serves as the input data for the wind power prediction model. It includes time-series data sequences of various candidate types associated with the wind power scenario. These time-series data sequences can be understood as data in chronological order, with timestamp alignment between them. Candidate types can be understood as data types, such as power type and wind speed type. Data to be repaired can be understood as data requiring repair processing, such as missing values ​​or outliers.

[0053] Optionally, if a wind power forecasting requirement is detected at the operation and maintenance end, the time series data sequences of each candidate type associated with the wind power scenario within the same time period can be obtained from the historical database as the original data sequence based on the wind power forecasting requirement, and a repair request for the original data sequence can be automatically initiated.

[0054] After detecting a repair request for the original data sequence, the data to be repaired in the original data sequence can be determined based on the data integrity and validity in the original data sequence. For example, missing values ​​and obvious outliers in the original data sequence can be directly used as the data to be repaired. Among them, obvious outliers can be understood as maximum or minimum values ​​that do not conform to the wind power operation scenario.

[0055] Alternatively, the original data sequence can be verified based on a preset data range, and data in the original data sequence that falls outside the preset data range can be used as data to be repaired.

[0056] S102, Analyze the original data sequence to obtain the data correlation between each candidate type, as well as the change trend parameters corresponding to the time series data sequences of each candidate type.

[0057] Among these, data correlation can be understood as the logical relationship between different candidate data types, such as the relationship between power and wind speed at the same time. Trend parameters can be understood as the data change trend of the time series itself.

[0058] Optionally, a dynamic decoupling approach can be used to analyze the original data sequence, extracting the data correlation between each candidate type and the trend parameters corresponding to the time series data sequences of each candidate type.

[0059] For example, the original data sequence can be decoupled and dynamically represented by the following formula (1). Where X is the original data sequence; U a The low-rank representation of the variable channels can be understood as the basis vectors of the variable channels, used to characterize the inherent correlation structure between variables, that is, the data association relationship between each candidate type in this embodiment; V aFor time-side dynamic representation, low-rank decomposition is performed in the time dimension to characterize the evolution law of time series itself, time-varying patterns, trend-cycle-abrupt dynamics and other information, that is, the change trend parameters corresponding to the time series data sequences of each candidate type in this embodiment.

[0060] (1)

[0061] S103. For each candidate type, an attention head bound to the variable channel of the candidate type is used to process the temporal data sequence, type encoding and physical attribute information of the candidate type to obtain the attention output of the candidate type.

[0062] Here, type encoding can be understood as the encoding corresponding to the candidate type, and physical attribute information can be understood as the physical attributes of the variables corresponding to the candidate type, such as the variable's physical role, unit, constraint attributes, and equipment source. Attention output can be understood as the result obtained after processing the relevant information of the input using an attention head, and can be understood as the overall condensed features of the data under the candidate type.

[0063] Understandably, assigning different attention heads to the variable channels corresponding to each candidate type can avoid interference between the features of different variables. In addition, it can specifically learn the unique fluctuation patterns and physical constraints of the variable, and will not be contaminated by other variable patterns, thus adapting to variable heterogeneity.

[0064] Based on this, for each candidate type, the type code of the candidate type is first retrieved from the variable type number table. Then, an attention head bound to the variable channel of the candidate type can be used to perform attention processing on the temporal data sequence, type code, and physical attribute information of the candidate type, thereby obtaining the attention output of the candidate type.

[0065] S104. Based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, the data to be repaired is repaired to obtain the data repair result.

[0066] The data repair result can be understood as the data repair result corresponding to the data to be repaired, or it can be the repaired data corresponding to the data to be repaired.

[0067] Optionally, a global repair data sequence containing each candidate type can be constructed based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type. Then, the repaired data corresponding to the data to be repaired can be extracted from the global repair data sequence according to the position of the data to be repaired in the original data sequence, thereby obtaining the data repair result.

[0068] Furthermore, the repaired data can be written into the corresponding position of the data to be repaired to obtain the repaired data sequence; then, subsequent wind power prediction processing can be performed based on the repaired data sequence.

[0069] In the aforementioned method for repairing wind power forecast input data, on the one hand, by identifying the data to be repaired in the original data sequence and analyzing the original data sequence to obtain the data correlation between each candidate type and the trend parameters corresponding to the time series data sequences of each candidate type, the overall data correlation and data trend of the wind power forecast input data can be better extracted. Furthermore, by using an attention head bound to the variable channel of each candidate type to process the time series data sequence, type encoding, and physical attribute information of the candidate type for each candidate type, the attention output of the candidate type can be obtained, avoiding mutual interference between the features of different variables and ensuring the reliability of the attention output. On the other hand, by combining the trend parameters and attention output corresponding to the candidate type, as well as the data correlation between each candidate type, the data to be repaired is processed to obtain the data repair result. This ensures that the data repair result is compatible with the trend and data characteristics of the wind power forecast input data, thereby improving the accuracy of wind power data repair.

[0070] Based on the above embodiments, this application provides an optional method for determining the data to be repaired, such as... Figure 2 As shown, the specific steps include:

[0071] S201, Based on the validity of each time series data in the original data sequence, determine the observation mask corresponding to each time series data in the original data sequence.

[0072] In this context, validity can be understood as the effective status of the data. Invalid data can include both missing data and invalid data. An observation mask can be understood as a character representing the validity of the data. For example, the observation mask can be 1 if the time series data is valid, and 0 if the time series data is invalid.

[0073] Optionally, for each time series data in the original data sequence, an observation mask can be selected from observation masks representing valid observations and observation masks representing invalid observations, based on the validity of the time series data.

[0074] For example, the original data sequence can be referenced by the following formula (2), where X is the original data sequence; C represents the number of candidate types; T represents the time length; x c,t This represents the time series data at time t in the time series data sequence of the c-th candidate type.

[0075] (2)

[0076] Furthermore, in x c,t When the value is valid, the corresponding observation mask Let it be 1, in x c,t When the value is invalid or missing, the corresponding observation mask Record it as 0.

[0077] S202, based on the input data constraints of the wind power prediction model and the spatial consistency of each time series data in the original data sequence, determine the anomaly mask corresponding to each time series data in the original data sequence.

[0078] Input data constraints can be understood as constraints that limit the reasonableness of input data; for example, they could be a standard data range for the input data. Spatial consistency can be understood as the synchronous similarity of data collected by power equipment that are geographically close and have similar terrain and meteorological conditions. Anomaly masks can be understood as characters that characterize abnormal data conditions.

[0079] Optionally, the input data constraints of the wind power prediction model can be used to perform anomaly detection on each time series data in the original data sequence, and the spatial consistency of each time series data in the original data sequence can be determined based on the location of the equipment that collected each time series data. Then, for each time series data, the anomaly situation of the time series data can be determined based on the anomaly detection results of the time series data and its spatial consistency with other data.

[0080] For example, when time-series data is outside the range of standard data, and / or when the spatial consistency of the time-series data indicates a different trend of change between the time-series data and the reference data, the abnormality of the time-series data is determined as a data anomaly. In this case, the device acquiring the time-series data is adjacent to the device acquiring the reference data.

[0081] Furthermore, based on the anomalies in the time series data, an anomaly mask can be selected from an anomaly mask representing anomalies and an anomaly mask representing normality.

[0082] S203. Based on the observation mask and anomaly mask corresponding to each time series data, determine the comprehensive repair mask corresponding to each time series data.

[0083] Among them, the comprehensive repair mask can characterize whether time-series data needs to be repaired.

[0084] Optionally, for each time series data, the comprehensive repair mask corresponding to that time series data can be determined by combining the meanings represented by the observation mask and the anomaly mask of that time series data.

[0085] For example, if the observation mask representation data is valid and the abnormal mask representation is normal, the time series data of the comprehensive repair mask representation does not need to be repaired; if the observation mask representation data is invalid or the abnormal mask representation is abnormal, the time series data of the comprehensive repair mask representation needs to be repaired.

[0086] For example, if the observation mask is 1 when the time series data is valid and 0 when the time series data is invalid, and the anomaly mask is 0 when the time series data is normal and 1 when the anomaly mask is abnormal, the following formula (3) can be used to calculate the observation mask and anomaly mask of the time series data to obtain the comprehensive repair mask corresponding to the time series data.

[0087] (3)

[0088] in, This is an anomaly mask; Represents a logical OR operation; This is the comprehensive repair mask; when the comprehensive repair mask is 1, the time series data represents that repair processing is required; when the comprehensive repair mask is 0, the time series data represents that repair processing is not required.

[0089] S204. Based on the comprehensive repair mask corresponding to each time series data, determine the data to be repaired in the original data sequence.

[0090] Optionally, for each time series data, when the comprehensive repair mask corresponding to the time series data indicates that the time series data needs to be repaired, the time series data is regarded as the data to be repaired.

[0091] In this embodiment of the application, by combining the validity of each time series data, input data constraints, and spatial consistency, the data to be repaired in the original data sequence is determined, which can ensure the rationality of the determination of the data to be repaired.

[0092] Based on the above embodiments, this application provides an optional method for determining the anomaly mask, such as... Figure 3 As shown, the specific steps include:

[0093] S301, based on the input data constraints of the wind power prediction model, performs anomaly assessment on each time series data in the original data sequence to obtain the basic anomaly score of each time series data.

[0094] The basic anomaly score can be understood as an anomaly score determined in conjunction with data constraints. The anomaly score is used to characterize the anomalies in time series data; the higher the anomaly score, the more anomalous the time series data.

[0095] Input data constraints include standard value range constraints, standard ramp rate constraints, standard power curve constraints, and standard variable linkage information constraints between candidate types. Standard value range constraints can include reasonable value ranges for each candidate type of data under normal wind power scenarios. Standard ramp rate constraints can include limits on the ramp rate of data under normal wind power scenarios, such as the maximum ramp rate. Standard power curve constraints can include the power curve corresponding to the data under normal wind power scenarios. Standard variable linkage information constraints between candidate types can include the inherent logical relationships between candidate type data, such as predefined normal matching relationships (linkage rules) between variables.

[0096] Optionally, for each time series data, anomaly assessment can be performed by combining standard value range constraints, standard ramp rate constraints, standard power curve constraints, and standard variable linkage information constraints between candidate types. Based on the matching between the time series data and each constraint, the basic anomaly score of the time series data can be determined.

[0097] S302, Based on the basic anomaly scores of each time series data and the spatial consistency of each time series data in the original data sequence, determine the target anomaly score of each time series data.

[0098] Among them, the target anomaly score can be understood as the anomaly score under the overall evaluation dimension.

[0099] Optionally, for each time series data point, the spatial consistency corresponding to that time series data point can be converted into a spatial anomaly score. A higher spatial anomaly score indicates that the changing trends of the data points, as represented by spatial consistency, are more distinct. Then, by weighting the basic anomaly score and the spatial anomaly score of the time series data, the target anomaly score can be obtained.

[0100] For example, the following formula (4) can be used to apply a preset weighting coefficient to the anomaly scoring under the standard value range constraint. Anomaly scoring under standard ramp rate constraints Anomaly scoring under standard power curve constraints Abnormal scoring under the constraint of standard variable linkage information and spatial anomaly scoring Weighted processing is performed to obtain the target anomaly score. .in, , , , and The weighting coefficients can be obtained through network self-learning or determined by those skilled in the art based on historical experience; there are no restrictions on this.

[0101] (4)

[0102] S303, determine the anomaly mask corresponding to each time series data based on the relationship between the target anomaly score and the score threshold of each time series data.

[0103] The scoring threshold is used to measure the magnitude of the target anomaly score and to assess the anomalies in the time series data. It can be determined by those skilled in the art based on historical experience or by a large amount of experimental data; there are no restrictions on this.

[0104] Optionally, for each time series data, the target anomaly score corresponding to the time series data can be compared with the score threshold. If the target anomaly score is greater than the score threshold, the anomaly mask corresponding to the time series data indicates that the data is abnormal; if the target anomaly score is less than or equal to the score threshold, the anomaly mask corresponding to the time series data indicates that the data is abnormal.

[0105] In this embodiment of the application, the anomaly of time series data is evaluated from multiple dimensions to obtain a target anomaly score, and the target anomaly score is compared with a score threshold to obtain an anomaly mask, which can ensure the accuracy of the anomaly mask determination.

[0106] Based on the above embodiments, this application provides an optional method for analyzing raw data sequences, such as... Figure 4 As shown, the specific steps include:

[0107] S401, determine the data association relationship between each candidate type based on the time series data sequence of each candidate type.

[0108] Optionally, after aligning the time series data sequences of each candidate type with timestamps, the time series data of each candidate type at the same time can be analyzed separately to obtain the data correlation between each candidate type.

[0109] For example, the Pearson correlation coefficient can be used to measure the strength of linear synchronization between candidate types; or, a method based on Singular Value Decomposition (SVD) / low-rank decomposition can be used to mine the global implicit association between candidate types of data from the perspective of channel coupling.

[0110] S402, for each candidate type, analyze the time series data sequence of the candidate type based on each preset trend dimension to obtain the change trend parameters corresponding to the time series data sequence of the candidate type.

[0111] The preset trend dimensions include trend dimensions under preset duration, periodic trend dimensions, local dynamic dimensions, and local deviation dimensions. Preset duration can be understood as long-term duration; trend dimensions under preset duration are used to depict the long-term, gradual, unidirectional / gradual overall trend of the sequence, belonging to low-frequency, slow-changing components with the longest change cycle. Periodic trend can be understood as a trend within a certain period, such as daily, weekly, or seasonal cycles, used to extract repetitive, regular oscillating components in the time series. Local dynamics are short-term, sudden, non-periodic, locally rapid fluctuations in medium-frequency / short-term dynamics, such as gusts and short-term fluctuations. Local deviation can be understood as abnormal residuals or short-term disturbances.

[0112] Optionally, for each candidate type, feature extraction processing can be performed on the time series data sequence of the candidate type based on each preset trend dimension to obtain the change trend parameters of the time series data sequence of the candidate type under each preset trend dimension.

[0113] For example, in the case of decoupled dynamic representation of the original data sequence, V can be represented by the following formula (5). a That is, the trend parameters corresponding to the time series data sequences of each candidate type are represented.

[0114] (5)

[0115] Among them, V t For the trend item, that is, the changing trend parameter under the trend dimension for the preset duration; V s For multi-period terms, that is, the changing trend parameters under the cyclical trend dimension; V l This refers to the local dynamic term, i.e., the parameter representing the trend of change within the local dynamic dimension; V b This refers to the local deviation term, i.e., the parameter representing the trend of change under the local deviation dimension; , , For dynamically decoupled adaptive learning weights.

[0116] Furthermore, the multi-period term V s It can be expressed using the following formula (6), where, Number the cycle type, for example, type 1 is daily cycle, type 2 is weekly cycle, type 3 is seasonal cycle; N p The number of candidate cycle types; For the first The length of the periodicity is expressed in the number of sampling points; t is the current time index; k is the harmonic order of this period. For the first The maximum harmonic order used in quasi-periodic systems; For the first The importance weight of cyclical patterns; , For the first The coefficients of the k-th order sine and cosine basis functions of the periodic class.

[0117] (6)

[0118] In this embodiment of the application, by performing an overall analysis of the original data sequence to obtain the data correlation relationship, and by performing individual analysis of each time series data sequence to extract the trend parameters, the rationality and accuracy of the data analysis can be improved.

[0119] Based on the above embodiments, this application provides an optional method for determining attention output, such as... Figure 5 As shown, the specific steps include:

[0120] S501, based on the time series data sequence of the candidate type, the observation mask and integrated repair mask corresponding to each time series data in the time series data sequence of the candidate type, and the type code and physical attribute information corresponding to the candidate type, construct the input information of the variable channel corresponding to the candidate type.

[0121] The input information can be understood as the information that needs to be input into the variable channel, that is, the initial input features of the variable channel.

[0122] Optionally, for each candidate type, the time series data sequence of the candidate type, the observation mask and the comprehensive repair mask corresponding to each time series data in the time series data sequence can be convolved to obtain the convolution processing result; then, the convolution processing result, the type code and physical attribute information corresponding to the candidate type are concatenated to obtain the input information of the variable channel corresponding to the candidate type.

[0123] For example, the input information z of the variable channel corresponding to candidate type c can be constructed by referring to the following formula (7). c , z c It contains an observation mask and a comprehensive repair mask.

[0124] (7)

[0125] in, Indicates the convolution operation; This is the normalized variable sequence of the time series data sequence of candidate type c; Type encoding for candidate type c; The physical attribute information for candidate type c is used to distinguish between power channels, meteorological channels, unit control channels, and spatial proximity channels.

[0126] S502 uses an attention head bound to the variable channel of the candidate type to process the input information and obtain the attention output of the candidate type.

[0127] Optionally, for each candidate type, the attention head bound to that candidate type can be used to perform attention processing on the input information, thereby obtaining the attention output of that candidate type.

[0128] For example, an attention head bound to the variable channel of the candidate type can be used to process the input information to obtain the query matrix, key matrix and value matrix corresponding to the candidate type; and attention processing can be performed on the prior bias matrix associated with the wind power scenario, as well as the query matrix, key matrix and value matrix corresponding to the candidate type to obtain the attention output of the candidate type.

[0129] The query matrix is ​​used to actively retrieve matching relevance during the attention processing; the key matrix serves as a reference index for matching during the attention processing; and the value matrix is ​​the original information carrier for weighted fusion during the attention processing. Each candidate type can contain multiple variable data representing the same type, and the prior bias matrix includes the variable weights of each variable under the candidate type. The wind power business prior bias matrix is ​​used to increase the attention weights between physically related variables. For example, in the power channel, wind speed, pitch angle, speed, unit status, adjacent unit power, and historical cycle power will have higher weights; in the wind speed channel, the wind speed of adjacent meteorological towers, the wind speed of adjacent units, wind direction, air pressure, and historical cycle wind speed will have higher weights.

[0130] Optionally, the trainable mapping parameters associated with the attention head bound to the variable channel of the candidate type can be used to apply the input information Z, referring to the following formulas (8)-(10). c After processing, the query matrix Q corresponding to candidate type c is obtained. c Key matrix K c Sum matrix V c .in, , and These are the trainable mapping parameters associated with the attention head corresponding to candidate type c.

[0131] (8)

[0132] (9)

[0133] (10)

[0134] Furthermore, referring to the following formula (11), for each candidate type, combined with the variable weights of each variable in the prior bias matrix associated with the wind power scenario, attention processing is performed on the query matrix, key matrix, and value matrix corresponding to each variable under that candidate type to obtain the attention output for that candidate type. Wherein, H is the prior bias matrix corresponding to candidate type c; c The attention output for candidate type c; d c The channel represents the dimension.

[0135] (11)

[0136] In this embodiment, by constructing input information and using an attention head bound to the variable channel of the candidate type to process the input information, the reliability of the attention output can be guaranteed.

[0137] Based on the above embodiments, this application provides an optional method for determining the data repair result, such as... Figure 6 As shown, the specific steps include:

[0138] S601 decodes the trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, to obtain the initial repair value corresponding to the data to be repaired.

[0139] The initial repair value can be understood as the data obtained after the data to be repaired has been repaired.

[0140] Optionally, the trend parameters and attention outputs corresponding to each candidate type, as well as the data correlations between each candidate type, can be concatenated, and the concatenation result can be decoded to obtain the initial repair sequence. The initial repair sequence contains the initial repair values ​​corresponding to each time series data point in the original data sequence.

[0141] For example, you can refer to formula (12) to represent the low-rank U on the variable channel side. a (Data relationships between candidate types), dynamic representation V on the time side a The initial repair sequence is obtained by decoding the trend parameters corresponding to each candidate type and the attention output corresponding to each candidate type. Among them, O1-O c Output the sum and attention for each candidate type; This is the decoding function.

[0142] (12)

[0143] Furthermore, based on the position of the data to be repaired in the original data sequence, the initial repair value corresponding to the data to be repaired can be extracted from the initial repair sequence.

[0144] S602, based on the physical constraints of the preset wind power scenario, corrects the initial repair value to obtain the data repair result.

[0145] The physical constraints include spatial consistency constraints and input data constraints for the wind power prediction model. Spatial consistency constraints are used to ensure that each time series data satisfies spatial consistency.

[0146] Optionally, after determining the initial repair value, it can be combined with other real time series data to determine whether the initial repair value meets the spatial consistency constraints and the input data constraints of the wind power prediction model. If it meets the constraints, the initial repair value is directly used as the data repair result; if it does not meet the constraints, the initial repair value is corrected by combining the spatial consistency constraints and the input data constraints to obtain the data repair result.

[0147] For example, the initial repair value can be corrected by referring to the following formula (13), combining spatial consistency constraints and input data constraints, to obtain the data repair result. Wherein, M r To comprehensively repair the mask; The corrected data repair result is shown in Figure 1; Y represents the data to be repaired. The standard power curve constraint for the data Y to be repaired; To constrain the standard value range of the data Y to be repaired; The standard ramp rate constraint for the data Y to be repaired; Constraints on the linkage information of standard variables among the candidate types of the data Y to be repaired; Spatial consistency constraints for the data Y to be repaired; Represents the square of the L2 norm; , , , , These are the weighting coefficients corresponding to each physical constraint.

[0148] (13)

[0149] After determining the data repair results, in order to ensure the accuracy of the data repair results, the confidence level of the data repair results can be calculated. For example, referring to the following formula (14), the confidence level of the data repair process can be calculated based on the local missing density, the attention intensity of the effective reference variable, the degree of violation of physical constraints, and the degree of residual abnormality during the data repair process.

[0150] (14)

[0151] in, Confidence level for data repair processing; This represents the density of localized defects. Attention intensity is a valid reference variable; The degree of violation of physical constraints; b represents the degree of residual abnormality; b is the bias term.

[0152] If the confidence level is greater than the confidence threshold, the data repair result can be written to the corresponding position of the data to be repaired to obtain the repaired data sequence; then, subsequent wind power prediction processing is performed based on the repaired data sequence. If the confidence level is less than or equal to the confidence threshold, a repair anomaly prompt can be sent to the operation and maintenance department to prompt intervention.

[0153] In this embodiment, the initial repair value is obtained by decoding the trend parameters, attention output, and data correlation. Then, the initial repair value is corrected based on physical constraints to obtain the data repair result. This ensures that the data repair result version meets the actual requirements, thereby improving the accuracy of the data repair result.

[0154] Figure 7 This is a flowchart illustrating a method for repairing wind power forecast input data in another embodiment. Based on the above embodiments, this embodiment provides an optional example of a method for repairing wind power forecast input data. (Combined with...) Figure 7 The specific implementation process is as follows:

[0155] S701, in response to a repair request for the original data sequence, determines the observation mask corresponding to each time series data in the original data sequence based on the validity of each time series data in the original data sequence.

[0156] The original data sequence is the input data for the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios.

[0157] S702, based on the input data constraints of the wind power prediction model and the spatial consistency of each time series data in the original data sequence, determine the anomaly mask corresponding to each time series data in the original data sequence.

[0158] Optionally, based on the input data constraints of the wind power prediction model, anomaly assessment is performed on each time series data in the original data sequence to obtain a basic anomaly score for each time series data. The input data constraints include standard value range constraints associated with standard input data, standard ramp rate constraints, standard power curve constraints, and standard variable linkage information constraints between candidate types. Based on the basic anomaly scores of each time series data and the spatial consistency of each time series data in the original data sequence, a target anomaly score for each time series data is determined. Based on the relationship between the target anomaly score and the score threshold, an anomaly mask corresponding to each time series data is determined.

[0159] S703. Based on the observation mask and anomaly mask corresponding to each time series data, determine the comprehensive repair mask corresponding to each time series data, and based on the comprehensive repair mask corresponding to each time series data, determine the data to be repaired in the original data sequence.

[0160] S704, determine the data association relationship between each candidate type based on the time series data sequence of each candidate type.

[0161] S705: For each candidate type, the time series data sequence of the candidate type is analyzed based on each preset trend dimension to obtain the change trend parameters corresponding to the time series data sequence of the candidate type.

[0162] The preset trend dimensions include the trend dimension under the preset duration, the periodic trend dimension, the local dynamic dimension, and the local deviation dimension.

[0163] S706: For each candidate type, based on the time series data sequence of the candidate type, the observation mask and comprehensive repair mask corresponding to each time series data in the time series data sequence of the candidate type, and the type code and physical attribute information corresponding to the candidate type, the input information of the variable channel corresponding to the candidate type is constructed, and the attention head bound to the variable channel of the candidate type is used to process the input information to obtain the attention output of the candidate type.

[0164] Optionally, an attention head bound to the variable channel of the candidate type is used to process the input information to obtain the query matrix, key matrix and value matrix corresponding to the candidate type; and attention processing is performed on the prior bias matrix associated with the wind power scenario, as well as the query matrix, key matrix and value matrix corresponding to the candidate type to obtain the attention output of the candidate type; wherein, the prior bias matrix includes the variable weights of each variable under the candidate type.

[0165] S707 decodes the trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, to obtain the initial repair value corresponding to the data to be repaired.

[0166] S708, based on the physical constraints of the preset wind power scenario, corrects the initial repair value to obtain the data repair result.

[0167] The physical constraints include spatial consistency constraints and input data constraints for the wind power prediction model.

[0168] The specific processes of S701-S708 described above can be found in the description of the above method embodiments. Their implementation principles and technical effects are similar, and will not be repeated here.

[0169] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0170] Based on the same inventive concept, this application also provides a wind power forecast input data repair device for implementing the wind power forecast input data repair method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more wind power forecast input data repair device embodiments provided below can be found in the limitations of the wind power forecast input data repair method described above, and will not be repeated here.

[0171] In one exemplary embodiment, such as Figure 8 As shown, a wind power forecast input data repair device 1 is provided, comprising: a data selection module 10, a data analysis module 20, a data processing module 30, and a result determination module 40, wherein:

[0172] The data selection module 10 is used to determine the data to be repaired in the original data sequence in response to a repair request for the original data sequence; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with the wind power scenario;

[0173] The data analysis module 20 is used to analyze the original data sequence to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type.

[0174] The data processing module 30 is used to process the temporal data sequence, type encoding, and physical attribute information of each candidate type using an attention head bound to the variable channel of the candidate type, and to obtain the attention output of the candidate type.

[0175] The result determination module 40 is used to repair the data to be repaired based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, and to obtain the data repair result.

[0176] In one exemplary embodiment, the data selection module 10 is specifically used for:

[0177] Based on the validity of each time series data in the original data sequence, determine the observation mask corresponding to each time series data in the original data sequence; based on the input data constraints of the wind power prediction model and the spatial consistency corresponding to each time series data in the original data sequence, determine the anomaly mask corresponding to each time series data in the original data sequence; based on the observation mask and anomaly mask corresponding to each time series data, determine the comprehensive repair mask corresponding to each time series data; based on the comprehensive repair mask corresponding to each time series data, determine the data to be repaired in the original data sequence.

[0178] In one exemplary embodiment, the data selection module 10 is further configured to:

[0179] Based on the input data constraints of the wind power prediction model, anomaly assessment is performed on each time series data in the original data sequence to obtain a basic anomaly score for each time series data. The input data constraints include standard value range constraints associated with standard input data, standard ramp rate constraints, standard power curve constraints, and standard variable linkage information constraints between candidate types. Based on the basic anomaly scores of each time series data and the spatial consistency of each time series data in the original data sequence, a target anomaly score for each time series data is determined. Finally, based on the relationship between the target anomaly score and the score threshold, an anomaly mask corresponding to each time series data is determined.

[0180] In one exemplary embodiment, the data analysis module 20 is specifically used for:

[0181] Based on the time series data sequences of each candidate type, the data correlation between each candidate type is determined; for each candidate type, the time series data sequences of the candidate type are analyzed based on each preset trend dimension to obtain the change trend parameters corresponding to the time series data sequences of the candidate type; among them, each preset trend dimension includes the trend dimension under preset duration, the periodic trend dimension, the local dynamic dimension, and the local deviation dimension.

[0182] In one exemplary embodiment, the data processing module 30 is specifically used for:

[0183] Based on the time series data sequence of the candidate type, the observation mask and integrated repair mask corresponding to each time series data in the time series data sequence of the candidate type, and the type encoding and physical attribute information corresponding to the candidate type, the input information of the variable channel corresponding to the candidate type is constructed; the input information is processed by the attention head bound to the variable channel of the candidate type to obtain the attention output of the candidate type.

[0184] In one exemplary embodiment, the data processing module 30 is further configured to:

[0185] An attention head bound to the variable channel of the candidate type is used to process the input information to obtain the query matrix, key matrix and value matrix corresponding to the candidate type; and attention processing is performed on the prior bias matrix associated with the wind power scenario, as well as the query matrix, key matrix and value matrix corresponding to the candidate type to obtain the attention output of the candidate type; wherein, the prior bias matrix includes the variable weights of each variable under the candidate type.

[0186] In one exemplary embodiment, the result determination module 40 is specifically used for:

[0187] The change trend parameters and attention outputs corresponding to each candidate type, as well as the data correlation between each candidate type, are decoded to obtain the initial repair value corresponding to the data to be repaired; based on the preset physical constraints under the wind power scenario, the initial repair value is corrected to obtain the data repair result; among which, the physical constraints include spatial consistency constraints and input data constraints of the wind power prediction model.

[0188] Each module in the aforementioned wind power forecast input data repair device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0189] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores wind power forecasting input data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for repairing wind power forecasting input data.

[0190] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0191] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0192] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0193] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0194] It should be noted that the data involved in this application (including but not limited to wind power forecast input data) are all data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0195] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0196] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0197] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for repairing wind power forecast input data, characterized in that, The method includes: In response to a repair request for the original data sequence, data to be repaired in the original data sequence is determined; wherein, the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios; The original data sequence is analyzed to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type; For each candidate type, an attention head bound to the variable channel of the candidate type is used to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type; Based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, the data to be repaired is repaired to obtain the data repair result.

2. The method according to claim 1, characterized in that, The step of determining the data to be repaired in the original data sequence includes: Based on the validity of each time series data in the original data sequence, determine the observation mask corresponding to each time series data in the original data sequence; Based on the input data constraints of the wind power prediction model and the spatial consistency of each time series data in the original data sequence, the anomaly mask corresponding to each time series data in the original data sequence is determined. Based on the observation mask and anomaly mask corresponding to each time series data, determine the comprehensive repair mask corresponding to each time series data; Based on the comprehensive repair mask corresponding to each time series data, the data to be repaired in the original data sequence is determined.

3. The method according to claim 2, characterized in that, The step of determining the anomaly mask corresponding to each time series data in the original data sequence based on the input data constraints of the wind power prediction model and the spatial consistency of each time series data in the original data sequence includes: Based on the input data constraints of the wind power prediction model, anomaly assessment is performed on each time series data in the original data sequence to obtain the basic anomaly score of each time series data; wherein, the input data constraints include standard value range constraints associated with standard input data, standard ramp rate constraints, standard power curve constraints, and standard variable linkage information constraints between each candidate type; Based on the basic anomaly score of each time series data and the spatial consistency of each time series data in the original data sequence, the target anomaly score of each time series data is determined. Based on the relationship between the target anomaly score and the score threshold of each time series data, the anomaly mask corresponding to each time series data is determined.

4. The method according to any one of claims 1-3, characterized in that, The analysis of the original data sequence to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequences of each candidate type includes: Based on the time series data sequences of each candidate type, determine the data correlation between each candidate type; For each candidate type, the time series data sequence of the candidate type is analyzed based on each preset trend dimension to obtain the change trend parameters corresponding to the time series data sequence of the candidate type; wherein, each preset trend dimension includes the trend dimension under a preset duration, the periodic trend dimension, the local dynamic dimension, and the local deviation dimension.

5. The method according to claim 2, characterized in that, The step of using an attention head bound to the variable channel of the candidate type to process the temporal data sequence, type encoding, and physical attribute information of the candidate type to obtain the attention output of the candidate type includes: Based on the time series data sequence of the candidate type, the observation mask and integrated repair mask corresponding to each time series data in the time series data sequence of the candidate type, and the type encoding and physical attribute information corresponding to the candidate type, the input information of the variable channel corresponding to the candidate type is constructed. The input information is processed using an attention head bound to the variable channel of the candidate type to obtain the attention output of the candidate type.

6. The method according to claim 5, characterized in that, The step of using an attention head bound to the variable channel of the candidate type to process the input information and obtain the attention output of the candidate type includes: The input information is processed using an attention head bound to the variable channel of the candidate type to obtain the query matrix, key matrix, and value matrix corresponding to the candidate type; and, Attention processing is performed on the prior bias matrix associated with the wind power scenario, as well as the query matrix, key matrix, and value matrix corresponding to the candidate type, to obtain the attention output of the candidate type; wherein, the prior bias matrix includes the variable weights of each variable under the candidate type.

7. The method according to any one of claims 1-3, characterized in that, The process of repairing the data to be repaired based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, to obtain the data repair result includes: The change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, are decoded to obtain the initial repair value corresponding to the data to be repaired; Based on the physical constraints of the preset wind power scenario, the initial repair value is corrected to obtain the data repair result; wherein, the physical constraints include spatial consistency constraints and input data constraints of the wind power prediction model.

8. A device for repairing wind power forecast input data, characterized in that, The device includes: A data selection module is used to determine the data to be repaired in the original data sequence in response to a repair request for the original data sequence; wherein the original data sequence is the input data of the wind power prediction model; the original data sequence includes time series data sequences of various candidate types associated with wind power scenarios; The data analysis module is used to analyze the original data sequence to obtain the data correlation between each candidate type and the change trend parameters corresponding to the time series data sequence of each candidate type. The data processing module is used to process the temporal data sequence, type encoding, and physical attribute information of each candidate type using an attention head bound to the variable channel of the candidate type, and obtain the attention output of the candidate type. The result determination module is used to repair the data to be repaired based on the change trend parameters and attention output corresponding to each candidate type, as well as the data correlation between each candidate type, to obtain the data repair result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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

Images