A power weather-oriented cross-modal data processing method and related device

Abstract

The application discloses a kind of power meteorology-oriented cross-modal data processing method and related equipment, comprising: in response to data access request, the cross-modal original data of accessed power meteorology is standardized, and cross-modal standard data is obtained, and according to cross-modal standard data, standard intermediate data model is constructed;According to target object to be processed, target anchor point time, target spatial position, target height, target physical variable and statistical window, construct space-time semantic anchor point;Based on space-time semantic anchor point, the cross-modal standard data in standard intermediate data model is processed to cross-modal alignment, and cross-modal alignment data is obtained;Cross-modal alignment data is quality evaluated, and comprehensive quality score is obtained;According to comprehensive quality score, cross-modal alignment data is classified and processed to abnormal difference, and target data is generated.The application can improve the processing precision of power meteorology cross-modal data, and can be widely applied in power meteorology data processing technical field.

Description

Technical Field

[0001] This invention relates to the field of power meteorological data processing technology, and in particular to a cross-modal data processing method and related equipment for power meteorology. Background Technology

[0002] In the field of power meteorology, particularly in wind power forecasting and grid safety dispatching, the fusion of multi-source meteorological data is crucial for improving accuracy. These data sources include numerical weather prediction (NWP), reanalysis data, observations from surface / marine weather stations, and wind turbine SCADA data. However, these data differ significantly in spatiotemporal resolution, physical definitions, and coordinate systems. For example, NWP uses regular grid point data, while weather stations use discrete point observations; SCADA records wind speed at the turbine hub height, while NWP provides wind speed at a specific pressure layer. Existing data fusion methods typically involve simple nearest-neighbor interpolation or temporal linear matching, which suffer from problems such as introducing noise due to spatiotemporal misalignment, ignoring operational timing leading to future information leakage, failing to effectively identify and repair anomalies caused by the turbine's own operating conditions, and lacking reliable quantification of the processing results. This severely restricts the construction of high-precision power meteorological models. Summary of the Invention

[0003] In view of this, the main objective of the embodiments of the present invention is to provide a method and related equipment for cross-modal data processing of power meteorology, in order to solve at least one of the problems of the prior art and improve the processing accuracy of cross-modal data of power meteorology.

[0004] To achieve the above objectives, one aspect of the present invention provides a cross-modal data processing method for power meteorology, the method comprising: In response to a data access request, the cross-modal raw data of the accessed power meteorology is standardized to obtain cross-modal standard data, and a standard intermediate data model is constructed based on the cross-modal standard data. Based on the target object to be processed, the target anchor time, the target spatial location, the target height, the target physical variables, and the statistical window, construct spatiotemporal semantic anchors; Based on the spatiotemporal semantic anchor points, cross-modal alignment processing is performed on the cross-modal standard data in the standard intermediate data model to obtain cross-modal aligned data; The quality of the cross-modal alignment data is evaluated to obtain a comprehensive quality score; Based on the comprehensive quality score, the cross-modal aligned data is classified as abnormal and processed for abnormal differentiation to generate target data.

[0005] In some embodiments, the process of performing cross-modal alignment processing on the cross-modal standard data in the standard intermediate data model based on the spatiotemporal semantic anchor points to obtain cross-modal aligned data includes the following steps: Based on the spatiotemporal semantic anchors and preset dual time tags, the cross-modal standard data is subjected to availability filtering and physical time series matching to obtain time series aligned data. Based on the spatiotemporal semantic anchors and atmospheric transport direction characteristics, the mapping weights from the temporal alignment data to the target spatial location and the target altitude are obtained, and the temporal alignment data is directionally weighted according to the mapping weights to obtain spatial mapping data. Based on the spatiotemporal semantic anchors and variable semantic dictionary, variable semantic alignment processing is performed on the spatial mapping data to obtain the cross-modal aligned data.

[0006] In some embodiments, the step of performing availability filtering and physical time-series matching on the cross-modal standard data based on the spatiotemporal semantic anchor points and preset dual time tags to obtain time-series aligned data includes the following steps: Based on the dual time tags, each piece of cross-modal standard data is labeled with physical time and available time; the physical time represents the observation time or forecast time described by the cross-modal standard data; the available time represents the earliest time when the cross-modal standard data is allowed to be called in the business system. Based on the spatiotemporal semantic anchor points, the target anchor point time is determined, and a candidate data window is constructed based on the target anchor point time. The cross-modal standard data whose physical time is within the candidate data window and whose available time is before the preset business processing time are selected to obtain cross-modal available data. Based on the physical time of the cross-modal available data, a physical correspondence between the cross-modal available data and the target anchor time is established through one or more of the following methods: nearest neighbor matching, fixed time interval matching, statistical window aggregation matching, or prediction step size matching. By using sliding time window correlation analysis, the dynamic lag of the cross-modal available data relative to the target anchor time is estimated, and the physical correspondence is corrected based on the dynamic lag. The timing alignment data is generated based on the available cross-modal data and the corrected physical correspondence.

[0007] In some embodiments, obtaining the mapping weights from the temporal alignment data to the target spatial location and the target altitude based on the spatiotemporal semantic anchor points and atmospheric transport direction characteristics includes the following steps: Obtain the geographic location corresponding to the time-series aligned data, and obtain the geometric distance between the geographic location and the target spatial location; Obtain the wind direction information corresponding to the time-series aligned data, and obtain the wind direction sector factor based on the geographical location, the wind direction information, and the target spatial location; Obtain the height difference between the height of the time-aligned data and the target height; Based on the first land-sea attribute corresponding to the time-series aligned data and the second land-sea attribute of the target spatial location, obtain the land-sea attribute consistency factor; Obtain the historical correlation coefficient between the temporal alignment data and the spatiotemporal semantic anchor point; Obtain the wake influence factor at the target spatial location; The geometric distance, wind direction sector factor, height difference, land-sea property consistency factor, historical correlation coefficient, and wake influence factor are input into a preset weighted mapping function to calculate the mapping weight.

[0008] In some embodiments, the quality assessment of the cross-modal alignment data to obtain a comprehensive quality score includes the following steps: Obtain the quantization metrics of the cross-modal alignment data in each quantization dimension; the quantization dimensions include integrity dimension, continuity dimension, physical consistency dimension, cross-modal consistency dimension, operational validity dimension, and source credibility dimension; The comprehensive quality score is obtained by weighting and summing the quantitative indicators according to the preset adjustable weight coefficients.

[0009] In some embodiments, the step of classifying and differentiating the cross-modal alignment data based on the comprehensive quality score to generate target data includes the following steps: Obtain abnormal data from the cross-modal alignment data, classify the abnormal data, and obtain the abnormal type of the abnormal data; Based on the anomaly type, a differentiated processing strategy is adopted to perform anomaly differentiation processing on the anomaly data; the differentiated processing strategy includes removing the anomaly data, retaining and marking the anomaly data, repairing the anomaly data, or only marking the anomaly data; When the cross-modal alignment data contains missing values, and the missing values ​​are data to be repaired, candidate repair values ​​are searched according to a preset priority. Obtain the target repair value from the candidate repair values, and fill in the missing values ​​according to the target repair value to obtain the target data.

[0010] In some embodiments, the step of performing anomaly classification and anomaly differentiation processing on the cross-modal alignment data based on the comprehensive quality score to generate target data further includes the following steps: When multiple candidate repair values ​​are found within the same priority range, or when the overall quality score of a lower priority candidate repair value is greater than the overall quality score of a higher priority candidate repair value, multi-source fusion processing is triggered. In response to multi-source fusion processing, the comprehensive weight of each candidate repair value is obtained based on the mapping weight of each candidate repair value, the comprehensive quality score, the time distance weight, the high consistency weight, and the source credibility weight. When the candidate repair value is a continuous variable, the target repair value is obtained by weighting each candidate repair value according to the comprehensive weight. When the candidate repair value belongs to the wind direction variable, the sine and cosine components of each candidate repair value are weighted and averaged, and the arctangent function is used to calculate the target repair value. When the candidate repair value is a discrete variable, the target repair value is obtained by prioritizing high-confidence sources, majority voting, or determining by preset business rules.

[0011] To achieve the above objectives, another aspect of the present invention proposes a cross-modal data processing device for power meteorology, the device comprising: The data standardization processing module is used to respond to data access requests, standardize the accessed cross-modal raw power meteorological data to obtain cross-modal standard data, and construct a standard intermediate data model based on the cross-modal standard data. The anchor point construction module is used to construct spatiotemporal semantic anchor points based on the target object to be processed, the target anchor point time, the target spatial location, the target height, the target physical variables, and the statistical window. The data alignment module is used to perform cross-modal alignment processing on the cross-modal standard data in the standard intermediate data model based on the spatiotemporal semantic anchor points to obtain cross-modal aligned data. The quality assessment module is used to assess the quality of the cross-modal alignment data and obtain a comprehensive quality score. The anomaly handling module is used to classify anomalies and perform anomaly differentiation processing on the cross-modal aligned data based on the comprehensive quality score, and generate target data.

[0012] To achieve the above objectives, another aspect of the present invention provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described above.

[0013] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0014] To achieve the above objectives, another aspect of the present invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions to cause the computer device to perform the aforementioned method.

[0015] The embodiments of the present invention include at least the following beneficial effects: The present invention provides a method and related equipment for cross-modal data processing of power meteorology. This solution, in response to a data access request, standardizes the accessed cross-modal raw power meteorology data to obtain cross-modal standard data. Based on the cross-modal standard data, a standard intermediate data model is constructed to provide a unified input benchmark for subsequent cross-modal processing. Furthermore, based on the target object to be processed, target anchor time, target spatial location, target height, target physical variables, and statistical windows, spatiotemporal semantic anchors are constructed, which can improve the accuracy of subsequent cross-modal data alignment. To avoid redundant computation of invalid data, the system accurately processes data boundaries, reducing business resource consumption. Based on spatiotemporal semantic anchors, it performs cross-modal alignment processing on cross-modal standard data in the standard intermediate data model to obtain cross-modal aligned data, which improves the spatiotemporal matching accuracy of cross-modal data. Quality assessment of the cross-modal aligned data yields a comprehensive quality score, enhancing the objectivity of data quality judgment and quantifying data credibility. Based on the comprehensive quality score, it performs anomaly classification and anomaly differentiation processing on the cross-modal aligned data to generate target data, improving the processing accuracy and business availability of power and meteorological cross-modal data. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a flowchart of a cross-modal data processing method for power meteorology provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of various cross-modal data processing stages for power meteorology provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0019] It should be noted that although functional modules are divided in the system diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first / S100" and "second / S200" in the specification, claims, and the foregoing drawings may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of the embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words "if" or "when" as used herein may be interpreted as "when," "in response to a determination," or "in the event of a determination."

[0020] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0022] Before providing a detailed description of the embodiments of the present invention, some of the nouns and terms involved in the embodiments of the present invention will be explained first. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations.

[0023] Cross-modal data refers to multiple types of data that differ fundamentally in their source, format, structure, spatiotemporal resolution, and physical meaning, such as numerical weather prediction data, ground meteorological station observation data, and wind farm SCADA data.

[0024] Spatiotemporal semantic anchors: a standardized and unified "bullseye" target defined for data alignment and cleaning, including a clear target object, time, spatial location, altitude, meteorological variable semantics, and statistical window.

[0025] Physical Time and Available Time: This invention proposes a dual-time system. Physical time refers to the actual time when the atmospheric phenomena described by the data occur; available time refers to the time when the data can be acquired and used for tasks such as prediction within the business process. Distinguishing between the two aims to prevent future information leakage.

[0026] Sliding window correlation analysis is a method for calculating the correlation coefficient between variables within a fixed-length window in time series or continuous data. It can not only capture local dynamic changes but also reveal differences in correlation across different time periods.

[0027] Directional weighted mapping: a spatial interpolation or mapping method whose weights not only consider distance, but also integrate multidirectional factors such as wind sector, land-sea surface properties, height differences, historical correlation, and wake effects.

[0028] Non-meteorological anomalies: These refer to data anomalies that are not caused by actual weather events, but by human factors such as power outages, shutdowns, equipment malfunctions, or sensor icing.

[0029] Quality Profile: A multi-dimensional data quality assessment vector that comprehensively scores each piece of data from six dimensions: completeness, continuity, physical consistency, cross-source consistency, operational effectiveness, and source credibility.

[0030] SCADA (Supervisory Control And Data Acquisition) system: a data acquisition and monitoring control system used in wind farms to collect data on the operating status of wind turbines and the environment.

[0031] Numerical weather prediction (NWP): Provides a three-dimensional physical field forecast of the future atmosphere and is the basis of forecasting.

[0032] Reanalysis of data: High-precision long-sequence gridded data from historical observations were assimilated for model training and climate assessment.

[0033] Surface / Marine meteorological station observations: Provides real meteorological element observations for the station locations.

[0034] Wind turbine SCADA data: Provides real-time operating conditions and environmental monitoring data for wind turbines.

[0035] In the field of power meteorology, multi-source meteorological data exhibit significant differences in spatiotemporal resolution, physical definitions, and coordinate systems. Related technologies include data preprocessing pipelines based on nearest neighbor and linear interpolation. In spatial matching, for any target wind turbine location, the nearest grid point in the NWP or reanalysis data grid is found by calculating the Euclidean distance, and the meteorological element value of that grid point is directly assigned to the wind turbine. In temporal matching, timestamped data recorded by the SCADA system is directly paired with the nearest NWP forecast lead time data for that timestamped time. For example, wind turbine observations at 12:00 on January 1, 2023, directly correspond to the NWP forecast values ​​reported starting at 12:00 on that day. In variable conversion, variables from different sources are unified using fixed formulas and coefficients, such as using a fixed power law formula to extrapolate wind speed at the measured height to the hub height. For outlier handling, out-of-bounds data is detected and removed based on a preset physical threshold range (e.g., wind speed 0-75 m / s). For missing values, simple linear interpolation or mean imputation is typically used. However, the simple nearest neighbor method ignores the significant impact of topography, underlying surface type (such as land-sea differences), wind direction, and other factors on local meteorological fields. A land station downwind of the sea breeze and the nearest upwind offshore grid point may represent completely different air masses; direct matching introduces serious errors and fails to account for micro-effects such as wakes. There is a risk of future information leakage in time matching. In operational forecasting, data from "future" observations or analyses that correspond physically but have not yet arrived on the operational timeline may be used, leading to inflated model evaluations that fail to reflect actual operational capabilities. Different data sources (such as NWPs from different institutions) or their products with different forecast lead times may exhibit systematic time lead or lag due to model errors or transmission delays. Existing methods cannot automatically estimate and correct this dynamic shift. Furthermore, relying solely on physical thresholds can misclassify many reasonable extreme weather events (such as typhoons) as anomalous, while failing to identify "non-meteorological anomalies." For example, a sudden drop in wind turbine power due to power curtailment orders manifests as an anomaly on the wind speed-power curve. However, screening methods using a single wind speed / power threshold cannot distinguish this from anomalies caused by turbulence or icing, leading to the inaccurate cleaning of valid data. Current data restoration methods are crude and lack uncertainty assessment; replacements or interpolations do not follow the priority of the data source, potentially substituting high-quality data with low-quality data. Using the restored data directly as "true values" without any quantitative information about its reliability introduces unknown risks to downstream tasks.

[0036] In view of this, this invention provides a method and related equipment for cross-modal data processing in power meteorology. This solution addresses the cross-modal data processing scenario in power meteorology by constructing a standardized end-to-end process from data access to cleaning and output. First, standardization unifies the metadata format, naming rules, and physical units of data from different sources, constructing a standard intermediate data model to eliminate underlying compatibility barriers between multi-source data. Second, by constructing spatiotemporal semantic anchors, the target object, time, space, altitude, and variable boundaries to be processed are clearly defined, avoiding invalid data from participating in redundant calculations. Then, based on the spatiotemporal semantic anchors, dual-time label filtering, directional spatial mapping, and three-dimensional alignment with unified variable semantics are performed to solve the spatiotemporal misalignment and expression differences in cross-modal data. Next, a comprehensive quality score is generated through a six-dimensional weighted calculation based on integrity, continuity, etc., providing a quantitative basis for anomaly identification. Finally, differentiated processing is performed on anomalous data according to different anomaly types. This solution improves the processing accuracy of cross-modal data in power meteorology, effectively reduces the interference of data noise on downstream power prediction models, and improves model training accuracy and online deployment stability.

[0037] This invention provides a cross-modal data processing method for power meteorology, relating to the field of power meteorological data processing technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle-mounted terminal, but is not limited thereto. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing a cross-modal data processing method for power meteorology, but is not limited to the above forms.

[0038] Figure 1 This is an optional flowchart of a cross-modal data processing method for power meteorology provided in an embodiment of the present invention. Figure 1 The method may include, but is not limited to, steps S100 to S500: Step S100: In response to the data access request, the cross-modal raw data of the accessed power meteorology is standardized to obtain cross-modal standard data, and a standard intermediate data model is constructed based on the cross-modal standard data. Step S200: Construct spatiotemporal semantic anchors based on the target object to be processed, target anchor time, target spatial location, target height, target physical variables, and statistical window; Step S300: Based on the spatiotemporal semantic anchor points, perform cross-modal alignment processing on the cross-modal standard data in the standard intermediate data model to obtain cross-modal aligned data; Step S400: Perform a quality assessment on the cross-modal alignment data to obtain a comprehensive quality score; Step S500: Based on the comprehensive quality score, perform anomaly classification and anomaly differentiation processing on the cross-modal aligned data to generate target data.

[0039] In step S100 of some embodiments, at least two types of raw power meteorological data of different modalities are first accessed. Cross-modal raw data may include, but is not limited to, numerical weather prediction, reanalysis data, ground meteorological station data, offshore wind measurement tower data, and wind turbine SCADA data. Subsequently, the metadata of various types of cross-modal raw data undergoes unified standardization processing. The timestamp, spatial latitude / longitude coordinates, variable naming, physical units of measurement, and missing value marking rules for each data point are parsed one by one. The metadata of all accessed cross-modal raw data is converted into a format conforming to a unified description specification, resulting in cross-modal standard data. Finally, all cross-modal standard data is encapsulated into a standard intermediate data model according to a predefined unified data architecture. This model serves as the data carrier for all subsequent data processing stages, eliminating format conflicts, naming ambiguities, and dimensional differences among multi-source heterogeneous data, providing a standardized input basis for subsequent cross-modal alignment and cleaning operations.

[0040] In step S200 of some embodiments, the business system first clarifies the data processing requirements based on the currently triggered business task type (such as short-term power prediction, wind turbine anomaly cleaning, wind resource assessment, etc.), thereby determining the target object to be processed (such as a wind turbine with a specific number, a designated offshore wind farm), selecting the target anchor point time (such as the start time of the power prediction task, the target verification time of the data cleaning task), defining the target spatial location (such as the latitude and longitude coordinates of the wind turbine, the center coordinates of the wind farm), the target height (such as the wind turbine hub height, the characteristic wind speed observation height), and specifying the target physical variables (such as the 10-minute average wind speed, active power, and pitch angle) and statistical window (such as the 15-minute moving average window). Subsequently, the above six types of elements are encapsulated into structured spatiotemporal semantic anchor points, which serve as the benchmark for subsequent cross-modal data alignment. This not only avoids irrelevant and redundant data from participating in the calculation and reduces the computing power consumption of the business system, but also provides a clear semantic reference for the association of different modal data, ensuring the accuracy of subsequent alignment operations.

[0041] In step S300 of some embodiments, based on spatiotemporal semantic anchors, time dimension alignment based on dual time label filtering, directional spatial mapping, and variable semantic unification are performed sequentially. This systematically solves problems such as spatiotemporal misalignment, inconsistent expression forms, and mismatched physical meanings in cross-modal data such as numerical forecasts, measured data, and SCADA data. The output cross-modal aligned data simultaneously meets all constraints defined by the spatiotemporal semantic anchors, providing a unified and accurately correlated input foundation for subsequent quality assessment and anomaly cleaning.

[0042] In some embodiments, step S300 may include, but is not limited to, steps S310 to S330: Step S310: Based on the spatiotemporal semantic anchor points and the preset dual time labels, perform availability filtering and physical time series matching on the cross-modal standard data to obtain time series aligned data; Step S320: Based on the spatiotemporal semantic anchor points and atmospheric transmission direction characteristics, obtain the mapping weights from the temporal alignment data to the target spatial location and target altitude, and perform directional weighted mapping on the temporal alignment data according to the mapping weights to obtain spatial mapping data; Step S330: Based on the spatiotemporal semantic anchors and variable semantic dictionary, perform variable semantic alignment processing on the spatial mapping data to obtain cross-modal aligned data.

[0043] In step S310 of some embodiments, with dual time labels as constraints and spatiotemporal semantic anchors as the benchmark, invalid data that cannot be actually called in the business is first excluded through availability filtering, and then the correspondence between the data and the target anchor time is established through physical time series matching. At the same time, the systematic lag of different data sources is corrected to avoid information leakage problems (such as using future data that has not been entered into the database). The output time series aligned data also carries metadata such as physical time, available time, and dynamic lag, providing accurate time dimension input for subsequent spatial mapping.

[0044] In some embodiments, step S310 may include, but is not limited to, steps S311 to S316: Step S311: Based on dual time labels, label each piece of cross-modal standard data with physical time and available time; physical time represents the observation time or forecast time described by the cross-modal standard data; available time represents the earliest time when the cross-modal standard data is allowed to be called in the business system; Step S312: Determine the target anchor time based on the spatiotemporal semantic anchor, and construct a candidate data window based on the target anchor time; Step S313: Filter out cross-modal standard data whose physical time is within the candidate data window and whose time is before the preset business processing time to obtain cross-modal available data; Step S314: Based on the physical time of cross-modal available data, establish a physical correspondence between cross-modal available data and target anchor time through one or more of the following methods: nearest neighbor matching, fixed time interval matching, statistical window aggregation matching, or forecast step size matching. Step S315: Through sliding time window correlation analysis, estimate the dynamic lag of cross-modal available data relative to the target anchor point time, and correct the physical correspondence based on the dynamic lag. Step S316: Generate timing alignment data based on available cross-modal data and the corrected physical correspondence.

[0045] In step S311 of some embodiments, a dual-time tag field is added to each cross-modal standard data in the standard intermediate data model. The dual-time tags include two tags: physical time and available time. For example, for measured data, the physical time can be the observation or acquisition time, and the available time is determined based on the data entry time, interface arrival time, or preset acquisition delay (e.g., if SCADA data is acquired at 10:00 and entered into the database at 10:05, the physical time is 10:00, and the available time is 10:05). For forecast data, the physical time is the forecast validity time, and the available time is determined based on the reporting start time, product release time, file arrival time, or business readable time. For reanalysis data or offline supplementary data, the available time is marked according to the time it can actually be called in the target business scenario, avoiding the misinterpretation of post-event data as real-time available data. The dual-time tag annotation provides a core time basis for subsequent availability verification and time series matching, which is a fundamental prerequisite for avoiding information leakage.

[0046] The available time refers to the earliest time that a piece of data can be used by models, algorithms, or cleaning processes in the actual business system. This time depends not only on the physical time corresponding to the data but also on factors such as data release time, transmission latency, file generation cycle, interface synchronization latency, quality control time, and data entry time. For example, if the effective forecast time of a numerical weather prediction product is 10:00, but the product is released and entered into the system at 08:45, then 10:00 is its physical time, and 08:45, or the actual time it is entered into the system, is its available time. As another example, if the acquisition time of a wind turbine SCADA record is 10:00, but the system can only read the record after 10:05, then 10:05 is its available time.

[0047] In step S312 of some embodiments, the business system determines the target object to be processed (such as a wind turbine) and the corresponding target anchor time (such as the target verification time for wind turbine power cleaning and the start time for power prediction) based on the spatiotemporal semantic anchor point, and constructs a candidate data window around the target anchor time. Optionally, if it is a real-time cleaning task, a short window of 15 minutes before and after the target anchor time is constructed; if it is a prediction task, an input window before the prediction start time and a verification window after the prediction target time are constructed respectively.

[0048] For example, for a power cleaning task of a wind turbine at a target time T, a certain range around the target time T can be selected as a candidate data window. Then, the meteorological, SCADA, or wind measurement data within this range around the target time T are the candidate data. For a forecasting task, an input window and a target window can be constructed around the forecast start time and forecast target time, respectively. By constructing candidate data windows, the scope of subsequent data filtering can be significantly narrowed, the computational cost of invalid data can be reduced, and the alignment efficiency can be improved.

[0049] The target anchor time refers to the baseline time of the target object in this alignment task; it is the "bullseye time" around which the cross-modal data alignment and cleaning process revolves. The target anchor time is usually determined by specific business tasks, such as power cleaning of a wind turbine at a certain moment, feature construction of a wind farm at a certain forecast start time, or generation of meteorological-power samples at a certain forecast target time. The target anchor time, together with the target object, target spatial location, target altitude, target physical variables, and statistical windows, constitutes the spatiotemporal semantic anchor.

[0050] Physical time, available time, and target anchor time are distinct yet interconnected. Physical time emphasizes the actual occurrence, observation, or forecast validity time described by the data; available time emphasizes when the data can be used in real business processes; and the target anchor time is used to determine the baseline time and sample window for this alignment task. Physical time is used to determine the correspondence of data in natural time series, available time is used to determine the availability of data in business time series, and the target anchor time is used to determine the center of data alignment. For the same data, physical time and available time can be the same or different; for the same target anchor time, data from multiple data sources and multiple physical times can be associated, but all selected data must satisfy the constraint that the available time is no later than the corresponding business processing time.

[0051] In step S313 of some embodiments, cross-modal standard data whose physical time falls within the candidate data window is first screened to ensure the physical time sequence correlation between the data and the target anchor time; then, cross-modal standard data whose available time is before the preset business processing time (such as the prediction start time or the cleaning execution time) is further screened to exclude data that has not yet been entered into the database and cannot be actually called in the business. The final cross-modal available data not only meets the physical time sequence requirements, but can also be actually obtained in the actual business.

[0052] For example, the business system first uses cross-modal standard data whose physical time falls within the candidate data window as candidate data. It then determines whether the availability time of the candidate data meets business constraints. Specifically, if the availability time of a piece of data is later than the time allowed by the current business task, that data cannot be included in subsequent feature construction, model training, or cleaning and discrimination processes. For prediction tasks, the data used as input features must have an availability time no later than the prediction start time; for real-time cleaning tasks, the data used must have an availability time no later than the business-processable time corresponding to the cleaning execution time or the target anchor time.

[0053] Business constraints based on available time can effectively prevent information leakage. Specifically, during model training and rule cleaning, data that has not yet been generated, published, or stored at the actual business time will not be used as input features; the samples constructed during the training phase are consistent with the actual available data conditions during the deployment phase; for numerical weather forecasts, reanalysis data, post-correction data, and delayed SCADA data, the system can distinguish "which time the data describes" and "when the data can be used," thereby avoiding the implicit introduction of future or post-hoc information into the model and improving the authenticity of offline evaluation results and the stability of online deployment.

[0054] In step S314 of some embodiments, the business system performs time matching based on the relationship between physical time and target anchor time, including nearest neighbor matching, fixed time interval matching, statistical window aggregation matching, or forecast step size matching. For example, for high-frequency SCADA data, nearest neighbor matching at 15-minute intervals is used to associate SCADA data closest to the target anchor time; for ground meteorological station and wind tower data, statistical window aggregation matching is used to associate data within windows before and after the target anchor time after calculating the mean / extreme values; for numerical weather prediction data, forecast step size matching is used to associate data with a forecast effective time equal to the target anchor time. The established physical correspondence clarifies the target anchor time corresponding to each piece of cross-modal available data, providing a basis for subsequent lag correction.

[0055] In step S315 of some embodiments, to address the systematic lag problem in cross-modal data (such as the 5-minute delay in wind transmission from the wind measurement tower to the wind turbine, and the delay in the transmission of meteorological changes to the power response), a candidate lag range (such as ±30 minutes) is set in the historical samples. The correlation coefficient, error index, or mutual information index between the available cross-modal data and the target variable are calculated for different lag values. The lag value with the highest correlation or the smallest error is selected as the dynamic lag value. Based on this lag value, the previously established physical correspondence is corrected (such as correcting the original matching wind speed at the same moment to the wind speed 5 minutes earlier), but the time constraint of data availability is not changed, so that the time-series matching is more in line with the physical laws of atmospheric transmission and equipment response.

[0056] In step S316 of some embodiments, the cross-modal available data and the corrected physical correspondence obtained from the previous steps are integrated, and a complete metadata identifier is attached to each time-series aligned data. This may include, but is not limited to, data source, physical time, available time, dynamic hysteresis correction amount, matching method, preliminary quality identifier, etc., to ensure that the source and matching logic of each time-series aligned data are traceable, and the final generated time-series aligned data has achieved accurate alignment in the time dimension.

[0057] In step S320 of some embodiments, physical characteristics such as the directionality of atmospheric transmission, altitude attenuation, land-sea differences, and wake effects are considered. Multi-dimensional factors are used to quantify the reference value of different time-series aligned data for the target spatial location. Through weighted mapping, time-series aligned data from different geographical locations and different altitude layers are uniformly mapped to the target spatial location and target altitude defined by the spatiotemporal semantic anchor point, solving the spatial mismatch problem of cross-location and cross-altitude data. The output spatial mapping data has been adapted to the physical environmental characteristics of the target spatial location.

[0058] In some embodiments, step S320 may include, but is not limited to, steps S321 to S327: Step S321: Obtain the geographic location corresponding to the time-series aligned data, and obtain the geometric distance between the geographic location and the target spatial location; Step S322: Obtain wind direction information corresponding to the time-series aligned data; and obtain wind direction sector factor based on geographical location, wind direction information, and target spatial location. Step S323: Obtain the height difference between the height of the timing alignment data and the target height; Step S324: Obtain the land-sea attribute consistency factor based on the first land-sea attribute corresponding to the time-series aligned data and the second land-sea attribute of the target spatial location; Step S325: Obtain the historical correlation coefficient between the temporal alignment data and the spatiotemporal semantic anchor point; Step S326: Obtain the wake influence factor at the target spatial location; Step S327: Input the geometric distance, wind direction sector factor, height difference, land-sea property consistency factor, historical correlation coefficient and wake influence factor into the preset weighted mapping function to calculate the mapping weight.

[0059] In step S321 of some embodiments, the geographic coordinates (such as the latitude and longitude of the wind tower and the station coordinates of the weather station) corresponding to each time-series aligned data are extracted from the standard intermediate data model. At the same time, the coordinates of the target spatial location (such as the latitude and longitude of the target wind turbine) are extracted from the spatiotemporal semantic anchor points. The geometric distance between the two is calculated using the spherical distance formula. The geometric distance is the basic calculation factor for the spatial mapping weight. Generally, the closer the distance, the higher the reference value of the time-series aligned data to the target location, and the greater the weight assigned.

[0060] In step S322 of some embodiments, the wind direction value at the current moment is first extracted from the time-aligned data. Then, the azimuth angle between the target spatial location and the geographical location of the time-aligned data is calculated. The wind direction is then used to determine whether the direction of atmospheric transmission is consistent with the direction of the line connecting the two locations. For example, if the wind direction is from the time-aligned data location to the target spatial location, the wind direction sector factor value is higher; if the wind direction is opposite or perpendicular, the wind direction sector factor value is lower. The wind direction sector factor reflects the directional characteristics of atmospheric transmission, avoiding the incorrect use of downwind or unrelated direction data in the calculation of the target location.

[0061] In step S323 of some embodiments, the observation or forecast height corresponding to each time-series aligned data (such as the wind speed at the 100-meter level of the meteorological tower or the wind speed at the 10-meter level of the numerical forecast) is extracted from the standard intermediate data model. At the same time, the target height (such as the wind turbine hub height of 120 meters) is extracted from the spatiotemporal semantic anchor point. The height difference between the two is calculated. The height difference is the core indicator for measuring the degree of vertical attenuation of wind speed. Generally, the smaller the height difference, the higher the correlation between the wind speed of the time-series aligned data and the wind speed at the target height, and the greater the weight assignment, which can effectively reduce the mapping error caused by wind shear.

[0062] In step S324 of some embodiments, a first land-sea attribute (e.g., offshore, coastal, inland) of the region where the time-series aligned data is located is extracted from the standard intermediate data model. Simultaneously, a second land-sea attribute of the target spatial location is extracted from the spatiotemporal semantic anchor point. If the land-sea attributes are consistent (e.g., both are offshore), the land-sea attribute consistency factor has a higher value; if the attributes differ significantly (e.g., the data source is an inland meteorological station, and the target is an offshore wind turbine), the factor value is lower. The land-sea attribute consistency factor avoids using data with excessively large underlying surface differences for target location calculations, improving the physical rationality of spatial mapping.

[0063] In step S325 of some embodiments, synchronous observation sequences of the time-aligned data location and the target physical variable (e.g., wind speed) at the target spatial location are extracted within a historical statistical window period (e.g., the past 3 months), and the Pearson correlation coefficient or mutual information value between the two is calculated as the historical correlation coefficient. The historical correlation coefficient reflects the data association strength between the data source and the target location within a historical period. The higher the historical correlation, the greater the long-term reference value of the data source, and the higher the weight assigned, which can effectively distinguish the reliability differences of different data sources.

[0064] In step S326 of some embodiments, the wake shading effect is calculated based on the target wind farm layout and current wind direction information defined by the spatiotemporal semantic anchor point. For example, if the time-series aligned data is located within the upstream wake region of the target wind turbine, the wake influence factor value is lower, reducing the weight of the data; if the time-series aligned data is located outside the wake region, the factor value is higher. The wake influence factor can mitigate data bias caused by upstream wind turbine wake shading, and is particularly suitable for spatial mapping scenarios of clustered wind farms, improving the accuracy of the mapping results.

[0065] In step S327 of some embodiments, the six types of factors acquired—geometric distance, wind direction sector factor, height difference, land-sea property consistency factor, historical correlation coefficient, and wake influence factor—are used as inputs and substituted into a predefined weighted mapping function (which can be a linear weighted model or a nonlinear mapping model trained based on historical data) to calculate the mapping weight of each time-series aligned data point to the target spatial location and target height. The weight values ​​range from [0,1], with higher values ​​indicating greater reference value of the time-series aligned data to the target location. Subsequent weighted mapping steps will fuse the results from different data sources based on this weight, ultimately outputting spatial mapping data adapted to the physical characteristics of the target location. For example, a directional weighted mapping algorithm is used for any source data point. Calculate its application to the target anchor point Mapping weights Its formula is: ; In the formula, For a composite function, For geometric distance, For wind direction sector factors, As a land-sea property consistency factor, For height difference, Historical correlation coefficient This represents the wake influence factor. This weight models the directional characteristics of atmospheric transport, replacing the simple nearest neighbor method.

[0066] In step S330 of some embodiments, to address issues such as inconsistent variable naming, inconsistent units, differences in expression forms (e.g., differences between U / V components and wind speed and direction), and inconsistent height layers in cross-modal data, variable name normalization and unit conversion are performed based on a predefined variable semantic dictionary. Through dedicated processing modules such as wind vector conversion, wind direction circumferential averaging, and height layer conversion, spatial mapping data of different formats are uniformly converted into the standard expression form of the target physical variables defined by spatiotemporal semantic anchors. The final output cross-modal aligned data conforms to the semantic specifications required by the business.

[0067] For example, the specific steps for variable semantic alignment are as follows: T1. Construct a variable semantic dictionary: The business system first establishes a unified variable semantic dictionary, mapping synonymous fields from different data sources to standard variable names. For example, fields such as "wind speed," "ws," "wind_speed," and "10m wind speed" are uniformly mapped to wind speed variables, fields such as "wind direction," "wd," and "wind_direction" are uniformly mapped to wind direction variables, and fields such as "u10," "v10," "u100," and "v100" are identified as east-west and north-south wind vector components at different altitude levels. Simultaneously, the physical unit, time statistical caliber, spatial level, altitude level information, and data source of each variable are recorded.

[0068] T2. Unit conversion and dimensional unification: The business system standardizes variables from different data sources to standard units based on the unit rules in the variable semantic dictionary. For example, wind speed is standardized to m / s, power to kW or MW, temperature to ℃ or K, air pressure to Pa or hPa, and precipitation to mm or m. For data with inconsistent units, the system converts them according to preset conversion relationships and retains the original unit, the converted unit, and the conversion identifier for easy traceability later.

[0069] T3, Wind Vector Conversion Module: For wind field data provided in U / V component form, such as numerical weather prediction and reanalysis data, the system performs a conversion of wind vectors to wind speed and direction. First, it reads the east-west wind component U and the north-south wind component V at a specified altitude and checks whether their time, spatial grid, and altitude are consistent. Then, it calculates the wind speed based on the vector composition relationship, i.e., the wind speed at the corresponding altitude is obtained by taking the square root of the sum of the squares of the U and V components. Next, it calculates the wind direction according to the meteorological wind direction definition, converting the vector angle into a 0° to 360° wind direction angle indicating the direction from which the wind is blowing. Finally, it checks the reasonableness of the conversion results to ensure that the wind speed is non-negative and the wind direction falls within the 0° to 360° range, and then writes the converted wind speed and direction into the standard intermediate data model.

[0070] T4, Wind Direction Circular Average Module: Because wind direction has a periodicity with 0° and 360° coinciding, a simple arithmetic mean cannot be directly applied. When performing temporal aggregation, spatial interpolation, or multi-source fusion of wind direction, the system first converts each wind direction angle into sine and cosine components on a unit circle. Then, it performs a weighted average of the sine and cosine components, with weights derived from time window weights, spatial mapping weights, or data quality weights. Next, it calculates the circumferential average wind direction based on the averaged sine and cosine components. Finally, it normalizes the result to the range of 0° to 360°. This process avoids erroneous results such as directly averaging 350° and 10° to obtain 180°, which is not physically meaningful.

[0071] T5, Height Conversion Module: To address the issue of inconsistent wind speed height layers across different data sources, the system employs power-law or logarithmic wind profiles to convert source height wind speeds to target anchor point heights. First, it identifies the observation or forecast height of the source data, such as 10m, 100m, 200m, or different tower heights, and reads the target height defined in the target anchor point, such as the wind turbine hub height. Then, it determines the number of available height layers and auxiliary information such as surface roughness and stability. When two or more height layers exist, the system can estimate the wind shear index or wind profile parameters based on the multi-layer wind speeds; when only a single height layer exists, the system can use regional empirical parameters, historical statistical parameters, or determine default parameters based on topographic surface classification.

[0072] T6. Height conversion based on power-law profile: When using a power-law model, the system converts the source height wind speed to the target height wind speed based on the source height, target height, and wind shear index. The wind shear index can be determined from historical data, real-time multi-level wind speeds, or empirical parameters from the site. If multiple wind speed levels (10m, 100m, 200m, etc.) exist, the system can dynamically estimate the wind shear index for the current moment or current wind conditions; if multi-level data is missing or of poor quality, historical statistical wind shear indices are used as a substitute. After the conversion, the system performs range checks and abrupt change checks on the target height wind speed to prevent abnormal height conversion results from entering subsequent processes.

[0073] T7. Height conversion based on logarithmic wind profile: When auxiliary information such as surface roughness, near-surface stability, or land-sea properties is available, the system can use logarithmic wind profiles for height conversion. The system first determines the roughness parameters based on the underlying surface type of the target area, or retrieves the roughness parameters from historical data; then, it combines the source height, target height, and roughness parameters to convert the source height wind speed to the target height. For offshore wind farms, wind farms in complex terrain, or data sources with large height differences, the system can adaptively select between power-law and logarithmic methods, taking into account land-sea properties, terrain type, and historical error performance.

[0074] T8. Integration and Quality Marking of Height Conversion Results: When the same target height can be obtained from multiple source heights or multiple conversion methods, the system performs weighted fusion based on factors such as data quality, historical relevance, spatial mapping weights, and height differences to generate the final target height wind speed. Simultaneously, the system attaches a quality identifier to each conversion result, including source height, target height, conversion method, parameters used, parameter source, and anomaly flag. If source data is missing, parameters are unavailable, or the conversion result exceeds a reasonable range, the system marks the variable as low confidence or missing, and it will not be directly used for subsequent critical cleaning judgments.

[0075] Through variable semantic alignment, the system can unify meteorological variables from different data sources, different field names, different physical units, different wind vector expressions, and different altitude layers into the same semantic space, providing consistent, traceable, and physically constrained standardized inputs for subsequent spatial mapping, anomaly identification, data cleaning, and model training.

[0076] In step S400 of some embodiments, a quality profile of the cross-modal alignment data is constructed, which evaluates the quality of the cross-modal alignment data from multiple dimensions by weighted summation to obtain a comprehensive quality score.

[0077] In some embodiments, step S400 may include, but is not limited to, steps S410 to S420: Step S410: Obtain the quantization metrics of cross-modal alignment data in each quantization dimension; the quantization dimensions include integrity dimension, continuity dimension, physical consistency dimension, cross-modal consistency dimension, operating condition validity dimension, and source credibility dimension; Step S420: Based on the preset adjustable weight coefficients, the quantitative indicators are weighted and summed to obtain the comprehensive quality score.

[0078] In step S410 of some embodiments, quantifiable feature values ​​of six dimensions are extracted from the cross-modal alignment data. Optionally, the completeness dimension counts the proportion of non-empty fields and the missing rate of required variables to reflect the completeness of the data; the continuity dimension counts the continuity rate of timestamps and the amplitude of data jumps between adjacent times to reflect the temporal continuity of the data; the physical consistency dimension counts the proportion of samples that exceed preset physical limits (such as negative wind speed, exceeding rated power, and wind direction exceeding the 0°-360° range) to reflect whether the data conforms to natural physical laws; the cross-modal consistency dimension counts the relative deviation rate of data from different sources (such as wind speed collected by SCADA and wind speed observed by anemometer towers) under the same spatiotemporal conditions to reflect the matching degree of cross-source data; the operating condition validity dimension counts the proportion of samples under non-meteorological operating conditions such as power rationing, shutdown, faults, and maintenance to reflect whether the data meets the business premise of normal wind turbine operation; and the source credibility dimension quantifies the calibration level of the data source, the accuracy of historical data, and the authority of the publishing agency to reflect the inherent reliability of the data source.

[0079] In step S420 of some embodiments, adjustable weight coefficients matching the current business scenario are first loaded. These weight coefficients support dynamic adjustment; for example, in short-term power prediction scenarios, weights for physical consistency and cross-modal consistency can be increased, while in wind farm operation and maintenance diagnosis scenarios, weights for operational effectiveness and source credibility can be increased. Subsequently, the quantitative indicators of each dimension are normalized to the 0-1 range, multiplied by the corresponding weight coefficients, and summed to obtain a comprehensive quality score with a value of 0-1. The closer the score is to 1, the higher the overall credibility of the cross-modal aligned data. This score can be used as a threshold for subsequent anomaly classification and as input for weight calculation in the multi-source fusion stage, enabling the full-process reuse of quality assessment results and improving the automation level of the overall processing flow.

[0080] For example, for each piece of data Generate a comprehensive quality score It is a weighted sum of six dimensions: ; In the formula, For each quantitative dimension, {Integrity, Continuity, Physical Consistency, Cross-Source Consistency, Operating Condition Validity, Source Trustworthiness}; Index for quantization dimensions; Adjustable weights; For data In quantitative dimensions The normalized individual quality score is obtained by normalizing the quantitative indicators of each quantitative dimension.

[0081] In step S500 of some embodiments, abnormal samples in cross-modal alignment data are first identified based on a threshold determination of the comprehensive quality score. Then, processing strategies are adapted to match different types of anomalies to avoid the problems of accidental deletion of extreme meteorological samples and incorrect inclusion of non-meteorological conditions in meteorological-power modeling caused by traditional indiscriminate cleaning. The final generated target data is accompanied by complete quality identification, anomaly tags and operating condition metadata, which can directly support downstream business scenarios such as power prediction model training and wind farm operation and maintenance analysis, thereby improving the reliability of business applications.

[0082] In some embodiments, step S500 may include, but is not limited to, steps S510 to S540: Step S510: Obtain abnormal data from the cross-modal alignment data, classify the abnormal data, and obtain the abnormal type of the abnormal data; Step S520: Based on the anomaly type, adopt a differentiated processing strategy to perform anomaly differentiation processing on the anomaly data; the differentiated processing strategy includes removing anomaly data, retaining and marking anomaly data, repairing anomaly data, or only marking anomaly data; Step S530: When there are missing values ​​in the cross-modal alignment data, and the missing values ​​are data to be repaired, candidate repair values ​​are searched according to the preset priority. Step S540: Obtain the target repair value from the candidate repair values, and fill in the missing values ​​according to the target repair value to obtain the target data.

[0083] In step S510 of some embodiments, anomaly data is identified by using a comprehensive quality score lower than a preset threshold as an initial screening condition, combined with multivariate joint analysis rules. For example, firstly, suspected anomalous samples with comprehensive quality scores lower than the preset threshold are screened, and then three types of anomalies are distinguished by variable correlation verification: if the sample has features such as missing, duplicate, timestamp errors, or exceeds physical limits (e.g., negative wind speed, exceeding rated power), it is determined to be a normal collection anomaly; if the sample shows high wind speed, low power, large propeller pitch angle, or features such as wind speed, zero power, or non-operational status code, it is determined to be a non-meteorological anomaly (including subcategories such as power rationing, shutdown, fault, maintenance, etc.); if the sample is an outlier caused by real weather processes such as gusts, sudden changes in wind direction, extreme low temperatures, typhoons, etc., it is determined to be a meteorological anomaly. The classification results serve as the core basis for subsequent differentiated processing, ensuring that the handling logic of each type of anomaly matches its generation mechanism.

[0084] The reason for anomaly classification is that different types of anomalies have different generation mechanisms, and the subsequent repair, removal, replacement, and fusion strategies also differ. If the causes of anomalies are not distinguished and all anomalies are deleted or interpolated uniformly, real extreme weather samples may be mistakenly deleted, or non-meteorological conditions such as power outages and shutdowns may be incorrectly learned as meteorological-power relationships, thereby affecting the data cleaning results and model training effects.

[0085] In step S520 of some embodiments, specific handling logic is matched to abnormal data of different abnormal types to perform differentiated abnormal processing. Optionally, four differentiated processing strategies are adopted for abnormal data: directly removing abnormal data, retaining and marking abnormal data, repairing abnormal data, and marking only abnormal data. For example, for normal collection anomalies, short-term missing data and local breakpoints are repaired by nearest neighbor interpolation, duplicate and conflicting records are retained with the highest quality version, and unrepairable out-of-limit data are directly removed; for non-meteorological anomalies, if used for meteorological-power relationship modeling, samples are removed or their training weights are reduced, if used for operational status analysis, samples are retained and labeled with operating condition tags such as power rationing and shutdown; for meteorological anomalies, they are not directly deleted or smoothed, but their authenticity is first confirmed through cross-validation of multi-source data, samples are retained and labeled with meteorological event tags to avoid misjudging real extreme weather samples as erroneous data, thus balancing data quality and the integrity of meteorological characteristics.

[0086] In some optional embodiments, routine data acquisition anomalies mainly originate from quality issues during data acquisition, transmission, storage, or format conversion, such as missing data, duplicates, incorrect timestamps, incorrect field formats, or exceeding physical limits. These anomalies typically lack valid physical meaning; therefore, the repair and fusion process primarily focuses on error correction, completion, and removal. For short-term missing data or local breakpoints, interpolation can be used to complete the data based on adjacent time points, historical patterns from the same source, or cross-source data. For duplicate or conflicting records, the most reliable record can be retained based on quality scores, entry time, source credibility, or business priority. Data that clearly exceeds physical boundaries and cannot be reliably repaired is marked as unusable or directly removed and does not participate in subsequent fusion and model training.

[0087] For anomalies not caused by meteorological factors, these are mainly caused by factors such as equipment control, power grid dispatching, power curtailment, shutdown, maintenance, faults, communication anomalies, or human intervention. This type of data is not due to abnormal meteorological conditions themselves, but rather changes in operating conditions leading to abnormal performance of variables such as power, pitch angle, and status codes. Therefore, during the repair and fusion process, it is not advisable to simply interpolate or directly use it as a normal sample; instead, it should be differentiated according to the business objective. If used for meteorological-power relationship modeling, theoretical power modeling, or wind resource assessment, this type of sample should be removed or downweighted to prevent the model from learning non-meteorological control behaviors such as power curtailment and shutdown. If used for actual output statistics or operating status analysis, the original data and its status labels should be retained. If used for generating power or theoretical power estimation, the actual power affected by power curtailment or shutdown can be replaced or repaired by combining power curves, adjacent normal wind turbines, available power data, or meteorological driving models. During the multi-source fusion stage, this type of data needs to retain non-meteorological cause labels to avoid mixing with normal meteorological response data.

[0088] For meteorological anomalies, which originate from real weather processes or complex meteorological conditions, such as gusts, wind shear, sudden changes in wind direction, enhanced turbulence, extreme low temperatures, icing, heavy precipitation, strong convection, or typhoons, this type of data, while potentially appearing as outliers in statistical distribution, is not necessarily erroneous. Instead, it may be an important sample reflecting extreme or complex weather processes. Therefore, in principle, it should not be directly deleted or simply smoothed out during the repair and fusion process. The system should perform cross-validation using multi-source meteorological data, nearby stations, wind towers, numerical weather prediction, radar, or reanalysis data. If multiple independent data sources can corroborate each other, the sample should be retained and labeled as a meteorological event. If only a single data source is abnormal and lacks support from other sources, its quality score should be lowered or it should be marked as pending confirmation. In the multi-source fusion stage, this type of data is weighted and fused based on cross-source consistency, source credibility, and quality score, preserving as much of the true meteorological disturbance characteristics as possible.

[0089] In step S530 of some embodiments, when it is confirmed that there are missing values ​​that need to be filled after anomaly classification (such as short-term missing values ​​in routine acquisition anomalies, or historical reference values ​​required for non-meteorological anomaly repair), candidate repair values ​​are searched according to preset priorities. The search scope is limited to cross-modal available data that has passed availability screening. The search is carried out strictly according to preset priorities, layer by layer. For example, the first priority is to search for cross-modal available data that is close to the target anchor point in time, the second priority is to search for cross-modal available data that is close to the target spatial location of the wind turbine, the third priority is to search for cross-modal available data that is close to the target area external observation station, and the fourth priority is to search for cross-modal available data that is close to the numerical forecast background field. After a candidate value that meets the requirements of comprehensive quality score is found, the subsequent search is stopped immediately, which ensures the timeliness of the repair value and avoids excessive search and consumption of computing power.

[0090] In step S540 of some embodiments, for the single-source retrieval results, the first candidate repair value that meets the quality requirements is determined as the target repair value, and the missing values ​​in the cross-modal aligned data are directly filled in. After filling, the repair source, repair time and repair confidence are simultaneously marked.

[0091] In some embodiments, step S500 may also include, but is not limited to, steps S550 to S590: Step S550: When multiple candidate repair values ​​are found within the same priority, or when the overall quality score of a low-priority candidate repair value is greater than the overall quality score of a high-priority candidate repair value, multi-source fusion processing is triggered. Step S560: In response to multi-source fusion processing, obtain the comprehensive weight of each candidate repair value based on the mapping weight, comprehensive quality score, temporal distance weight, high consistency weight, and source credibility weight of each candidate repair value. Step S570: When the candidate repair value is a continuous variable, the target repair value is obtained by weighting each candidate repair value according to the comprehensive weight. Step S580: When the candidate repair value belongs to the wind direction variable, the sine and cosine components of each candidate repair value are weighted and averaged, and the arctangent function is used to calculate the target repair value. Step S590: When the candidate repair value is a discrete variable, the target repair value is obtained by prioritizing high-confidence sources, majority voting, or determining by preset business rules.

[0092] In step S550 of some embodiments, two types of multi-source fusion triggering conditions are set. The first is that two or more candidate repair values ​​with satisfactory comprehensive quality scores are retrieved within the same priority (such as three qualified observation values ​​of neighboring wind turbines at the same time). The second is that the comprehensive quality score of the low-priority candidate repair value is significantly higher than that of the high-priority candidate (such as the score of the neighboring time data is 0.75, while the score of the neighboring wind turbine is 0.92). In this case, a single candidate repair value is no longer used, and the multi-source fusion process is automatically triggered to make full use of the complementarity of multi-source data, reduce the random error of single-source data, and improve the stability of the repair results.

[0093] In step S560 of some embodiments, a comprehensive weight is calculated for each candidate repair value. The mapping weight is calculated using a directional weighted mapping algorithm, which calculates the directional weighted mapping result of the source data points applied to the target anchor point, reflecting the physical correlation between the candidate repair value and the target spatial location and target height. The comprehensive quality score reflects the reliability of the candidate repair value itself. The time distance weight is calculated by the difference between the physical time of the candidate repair value and the time of the target anchor point; the smaller the difference, the higher the weight. The height consistency weight is calculated by the difference between the height of the candidate repair value and the target height; the smaller the difference, the higher the weight. The source credibility weight is determined by the calibration level and historical accuracy of the data source. The five weights are multiplied and then normalized to obtain the comprehensive weight of each candidate repair value. The weight value ranges from [0,1], with higher values ​​indicating greater reference value for the candidate value.

[0094] For example, for the first For each candidate repair value, the system comprehensively considers spatial mapping weight, overall quality score, temporal distance, high consistency, and data source credibility to calculate its comprehensive weight. The calculation formula used is as follows: ; in, Indicates the overall weight. Indicates the spatial mapping weights. This indicates the overall quality score. Indicates the time distance weight. Indicates a high degree of consistency weight. This represents the source credibility weight.

[0095] In step S570 of some embodiments, for continuous candidate repair values ​​such as wind speed, power, temperature, and air pressure, a weighted average method is used to fuse multi-source information. For example, for continuous variables, the calculation formula for the target repair value is: ; in, This represents the target repair value at the target anchor point. Indicates the first One candidate repair value, Indicates the first The fusion weights of the candidate repair values.

[0096] In step S580 of some embodiments, for the wind direction variable, since the wind direction has a periodicity with 0° and 360° consecutively connected, a simple weighted average cannot be directly performed. The system first converts the wind direction into sine and cosine components, then performs weighted fusion on the sine and cosine components respectively, and then calculates the corrected wind direction using the arctangent function. For example, for the wind direction variable, the calculation formula for the target corrected value includes: ; ; ; ; ; In the formula, Indicates the first The sinusoidal component of the candidate wind direction repair value Indicates the first The cosine component of each candidate wind direction repair value Indicates the first The candidate wind direction angles corresponding to the candidate wind direction repair values. Indicates the first The overall weight of each candidate wind direction correction value This indicates the total number of candidate repair values ​​participating in multi-source fusion. This represents the weighted average sinusoidal component after fusion. This represents the weighted average cosine component after fusion. This represents the optimal candidate wind direction angle obtained through fusion, i.e., the target repair value.

[0097] In step S590 of some embodiments, for discrete candidate repair values ​​such as status codes, operating statuses, and fault types, the error handling method of numerical weighted averaging is prohibited. According to the priority matching judgment logic, optionally, the high-confidence source priority rule is first adopted to select the candidate value with the highest comprehensive weight as the optimal candidate value; if the comprehensive weights of multiple candidate values ​​are similar and the values ​​are consistent, the majority voting rule is adopted to determine the result; if there is a conflict of candidate values ​​(such as SCADA reporting shutdown, but manual log marking as maintenance), the preset business rule is adopted to determine (such as the priority of SCADA real-time status codes is higher than that of manual log entries). The final determined optimal candidate value is attached with a judgment basis label to ensure that the repair result of discrete variables conforms to the business logic and avoids incorrect labeling of operating conditions.

[0098] In some embodiments, the uncertainty of the fusion result is also calculated based on the deviation between each candidate repair value and the target repair value obtained after fusion. The target data is then labeled with high, medium, or low confidence levels according to the magnitude of the uncertainty. During the downstream model training phase, the training weights of samples can be dynamically adjusted based on the uncertainty, so that samples with higher uncertainty contribute less to the model training. For example, after obtaining the target repair value, the business system simultaneously calculates the uncertainty of the fusion result to describe the confidence level of the repair value. For continuous variables, uncertainty can be expressed as a weighted standard deviation: ; in, This indicates the uncertainty of the fusion result. If the values ​​of multiple candidate repair values ​​are close, then... A smaller value indicates a stable and reliable repair result; if multiple candidate repair values ​​differ significantly, then... A large value indicates that the reliability of the repair result is low.

[0099] The calculated uncertainty can be used for subsequent data quality control, data fusion, and model training. A first uncertainty threshold is preset according to variable type and business scenario. and the second uncertainty threshold ;when When the repair value is marked as high confidence, it can participate normally in subsequent cleaning, fusion, and model training; when When the repair value is marked as medium confidence, it can participate in subsequent processes, but its weight should be appropriately reduced in fusion calculations or model training; when When this happens, the repaired value is marked as low confidence or pending review, and should not be used for key model training and business judgment if necessary.

[0100] During the model training phase, sample training weights can be set according to the uncertainty: ; in, This represents the training weights of the samples. The larger, The smaller the value, the less the model relies on highly uncertain samples during training.

[0101] In some embodiments, the generated target data is also output as a standard sample set with metadata identifiers, which include at least the data source, comprehensive quality score, repair identifier, operating condition label, and uncertainty. The downstream power prediction model is trained using the standard sample set, and the model prediction error is monitored. The prediction error is used as a feedback signal to adjust the dynamic lag parameter, spatial mapping weight function parameter, and criterion threshold in the cross-modal alignment processing, so as to achieve closed-loop iterative optimization oriented towards business results.

[0102] In some embodiments, such as Figure 2 As shown, the cross-modal data processing procedure for power meteorology in this invention includes four stages, constituting a complete process from data preparation to closed-loop optimization. (Reference) Figure 2 The various stages of cross-modal data processing include: Phase 1: Data Preparation and Standardization S1. Multi-source data access: Access at least two types of raw data, including but not limited to numerical weather forecasts, reanalysis data, ground meteorological stations, offshore wind measurement towers, and wind turbine SCADA data.

[0103] S2. Metadata Standardization: The time, spatial coordinates, variable names, physical units, and missing value markers of all accessed data are uniformly parsed and transformed to form a standard intermediate data model with unified description specifications.

[0104] S3. Construct spatiotemporal semantic anchor points: As the "bullseye" of the entire alignment and cleaning process, clearly define the target object to be processed (such as a wind turbine), target time, target spatial location, target height, target physical variables and their statistical windows.

[0105] Phase 2, Cross-modal core alignment: S4. Time Alignment: First, a dual-time system is constructed, defining two labels for each data point: physical time and available time. All subsequent alignment and model training are strictly based on the available time to prevent information leakage. Next, dynamic lag estimation is performed, analyzing time series from different data sources. Through sliding time window correlation analysis, the systematic offsets relative to the target anchor time are automatically estimated and corrected. Optionally, in addition to correlation analysis, dynamic lag estimation can also employ dynamic time warping algorithms to capture more complex and nonlinear time shifts.

[0106] S5. A directional weighted mapping algorithm is used to perform directional weighting of spatial mapping. For any source data point, the mapping weight applied to the target anchor point is calculated. Optionally, in the directional weighted mapping, the weight calculation function can be implemented using different kernel methods, such as Gaussian process regression or a nonlinear model based on random forest to learn the mapping from each factor to the weight, instead of using a manually preset explicit formula.

[0107] S6. Variable semantic alignment: Unifies fields through variable dictionary and unit conversion. Based on the embedded wind vector conversion module (U / V component to wind speed and direction), wind direction circumferential averaging module (handling cyclic statistics of wind direction crossing 0 degrees), and height conversion module based on power law / logarithmic wind profile, it realizes adaptive conversion of wind speed at different height levels.

[0108] Phase Three: Quality Assessment and Reliable Cleaning S7. Construct a quality profile: For each data point, generate a comprehensive quality score based on six dimensions.

[0109] S8. Anomaly Classification: For routine identification, this includes detecting missing, duplicate, and data exceeding physical limits. For multivariate joint identification of non-meteorological anomalies, this involves correlation analysis of wind speed, power, pitch angle, status codes, etc., to identify typical patterns such as power curtailment (high wind speed, low power, large pitch angle) and shutdown (with wind speed, zero power, specific status code).

[0110] S9. Repair and Fusion: Four processing strategies are adopted for outlier data, including direct removal, retention but labeling, repair, and labeling only. When anchor data is missing and needs repair, candidate repair values ​​are searched according to the priority order of the data sources. Optionally, the priority order is: nearby time > nearby wind turbines > external observation stations > forecast background field. Once a candidate repair value is found, the search stops and the missing value is filled. If multiple high-quality sources exist, multi-source fusion and uncertainty estimation are performed. The target repair value is obtained by weighted fusion using spatial mapping weights and comprehensive quality scores, and the uncertainty of the fusion result is calculated simultaneously. Optionally, when obtaining repair values ​​through multi-source fusion, in addition to weighted averaging, methods based on Bayesian inference or conditional generative adversarial networks can be used to generate the most likely repair value from the perspective of probability distribution and directly output the uncertainty.

[0111] Phase 4, Output and Closed-Loop Optimization: S10. Output Standard Sample Set: Each data sample produced is output in a structured format and includes complete data source identification and a comprehensive quality score. Metadata such as repair labels, operating condition labels, and uncertainty are included to ensure full traceability.

[0112] S11. Task Feedback Optimization: The dataset produced by the method of this invention is used to train the downstream power prediction model, and the model prediction error is monitored. The prediction error is used as a feedback signal to optimize the dynamic lag parameter, the spatial mapping weight function parameter, and the criterion threshold in anomaly classification, forming a closed-loop iterative optimization oriented towards business results.

[0113] This invention also provides a cross-modal data processing device for power meteorology, which can implement the above-mentioned cross-modal data processing method for power meteorology. The device includes: The data standardization processing module is used to respond to data access requests, standardize the accessed cross-modal raw power meteorological data to obtain cross-modal standard data, and construct a standard intermediate data model based on the cross-modal standard data. The anchor point construction module is used to construct spatiotemporal semantic anchor points based on the target object to be processed, the target anchor point time, the target spatial location, the target height, the target physical variables, and the statistical window. The data alignment module is used to perform cross-modal alignment processing on cross-modal standard data in the standard intermediate data model based on spatiotemporal semantic anchors, so as to obtain cross-modal aligned data. The quality assessment module is used to assess the quality of cross-modal aligned data and obtain a comprehensive quality score. The anomaly handling module is used to classify anomalies and perform anomaly differentiation processing on cross-modal aligned data based on the comprehensive quality score, and generate target data.

[0114] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0115] This invention also provides an electronic device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including a tablet computer, an in-vehicle computer, or similar device.

[0116] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0117] refer to Figure 3 , Figure 3 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 601 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention. The memory 602 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 602 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 602 and is called and executed by the processor 601. The input / output interface 603 is used to implement information input and output; The communication interface 604 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 605 transmits information between various components of the device (e.g., processor 601, memory 602, input / output interface 603, and communication interface 604); The processor 601, memory 602, input / output interface 603, and communication interface 604 are connected to each other within the device via bus 605.

[0118] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0119] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0120] This invention also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions to cause the computer device to perform the aforementioned method.

[0121] In summary, the cross-modal data processing method and related equipment for power meteorology according to embodiments of the present invention have the following advantages: 1. The embodiments of the present invention comprehensively consider factors such as wind direction and land-sea distribution through directional weighted mapping, which is more in line with the laws of atmospheric physics. Compared with the single nearest neighbor method, it can significantly reduce the spatial representativeness error in complex environments, thereby improving the input accuracy of downstream models.

[0122] 2. The dual time-label system of this invention strictly distinguishes between phenomenon time and business availability time from the data labeling level, ensuring the authenticity and reliability of model evaluation and training, objectively reflecting the true predictive ability of the business system, and avoiding inflated evaluation results.

[0123] 3. The embodiments of the present invention can accurately identify and retain real extreme weather data (such as typhoons and strong wind shears) by using multivariate joint identification of non-meteorological anomalies, while effectively cleaning up false anomalies caused by power outages, shutdowns, etc., and protecting valuable high wind speed samples.

[0124] 4. The priority-based repair and multi-source fusion in this embodiment of the invention preserve and utilize high-quality information to the greatest extent; while the output with uncertainty estimation provides prior knowledge of the credibility of the data for downstream AI prediction models, enabling the models to better handle uncertainty and make more robust decisions.

[0125] 5. Through the task feedback optimization steps in this embodiment of the invention, the business system is no longer a one-way preprocessing pipeline, but can be linked with downstream tasks (such as power prediction) to automatically adjust its internal parameters with the end in mind, thereby achieving a dual improvement in data quality and business performance.

[0126] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0127] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0128] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0129] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0130] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0131] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0132] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0133] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A cross-modal data processing method for power meteorology, characterized in that, Includes the following steps: In response to a data access request, the cross-modal raw data of the accessed power meteorology is standardized to obtain cross-modal standard data, and a standard intermediate data model is constructed based on the cross-modal standard data. Based on the target object to be processed, the target anchor time, the target spatial location, the target height, the target physical variables, and the statistical window, construct spatiotemporal semantic anchors; Based on the spatiotemporal semantic anchor points, cross-modal alignment processing is performed on the cross-modal standard data in the standard intermediate data model to obtain cross-modal aligned data; The quality of the cross-modal alignment data is evaluated to obtain a comprehensive quality score; Based on the comprehensive quality score, the cross-modal aligned data is classified as abnormal and processed for abnormal differentiation to generate target data.

2. The method according to claim 1, characterized in that, The process of performing cross-modal alignment processing on the cross-modal standard data in the standard intermediate data model based on the spatiotemporal semantic anchor points to obtain cross-modal aligned data includes the following steps: Based on the spatiotemporal semantic anchors and preset dual time tags, the cross-modal standard data is subjected to availability filtering and physical time series matching to obtain time series aligned data. Based on the spatiotemporal semantic anchors and atmospheric transport direction characteristics, the mapping weights from the temporal alignment data to the target spatial location and the target altitude are obtained, and the temporal alignment data is directionally weighted according to the mapping weights to obtain spatial mapping data. Based on the spatiotemporal semantic anchors and variable semantic dictionary, variable semantic alignment processing is performed on the spatial mapping data to obtain the cross-modal aligned data.

3. The method according to claim 2, characterized in that, The process of performing availability filtering and physical time-series matching on the cross-modal standard data based on the spatiotemporal semantic anchor points and preset dual time tags to obtain time-series aligned data includes the following steps: Based on the dual time tags, each piece of cross-modal standard data is labeled with physical time and available time; the physical time represents the observation time or forecast time described by the cross-modal standard data; the available time represents the earliest time when the cross-modal standard data is allowed to be called in the business system. Based on the spatiotemporal semantic anchor points, the target anchor point time is determined, and a candidate data window is constructed based on the target anchor point time. The cross-modal standard data whose physical time is within the candidate data window and whose available time is before the preset business processing time are selected to obtain cross-modal available data. Based on the physical time of the cross-modal available data, a physical correspondence between the cross-modal available data and the target anchor time is established through one or more of the following methods: nearest neighbor matching, fixed time interval matching, statistical window aggregation matching, or prediction step size matching. By using sliding time window correlation analysis, the dynamic lag of the cross-modal available data relative to the target anchor time is estimated, and the physical correspondence is corrected based on the dynamic lag. The timing alignment data is generated based on the available cross-modal data and the corrected physical correspondence.

4. The method according to claim 2, characterized in that, The step of obtaining the mapping weights from the temporal alignment data to the target spatial location and the target altitude based on the spatiotemporal semantic anchor points and atmospheric transport direction characteristics includes the following steps: Obtain the geographic location corresponding to the time-series aligned data, and obtain the geometric distance between the geographic location and the target spatial location; Obtain the wind direction information corresponding to the time-series aligned data, and obtain the wind direction sector factor based on the geographical location, the wind direction information, and the target spatial location; Obtain the height difference between the height of the time-aligned data and the target height; Based on the first land-sea attribute corresponding to the time-series aligned data and the second land-sea attribute of the target spatial location, obtain the land-sea attribute consistency factor; Obtain the historical correlation coefficient between the temporal alignment data and the spatiotemporal semantic anchor point; Obtain the wake influence factor at the target spatial location; The geometric distance, wind direction sector factor, height difference, land-sea property consistency factor, historical correlation coefficient, and wake influence factor are input into a preset weighted mapping function to calculate the mapping weight.

5. The method according to claim 1, characterized in that, The process of evaluating the quality of the cross-modal aligned data to obtain a comprehensive quality score includes the following steps: Obtain the quantization metrics of the cross-modal alignment data in each quantization dimension; the quantization dimensions include integrity dimension, continuity dimension, physical consistency dimension, cross-modal consistency dimension, operational validity dimension, and source credibility dimension; The comprehensive quality score is obtained by weighting and summing the quantitative indicators according to the preset adjustable weight coefficients.

6. The method according to claim 1, characterized in that, The step of classifying and processing anomalies in the cross-modal aligned data based on the comprehensive quality score to generate target data includes the following steps: Obtain abnormal data from the cross-modal alignment data, classify the abnormal data, and obtain the abnormal type of the abnormal data; Based on the anomaly type, a differentiated processing strategy is adopted to perform anomaly differentiation processing on the anomaly data; the differentiated processing strategy includes removing the anomaly data, retaining and marking the anomaly data, repairing the anomaly data, or only marking the anomaly data; When the cross-modal alignment data contains missing values, and the missing values ​​are data to be repaired, candidate repair values ​​are searched according to a preset priority. Obtain the target repair value from the candidate repair values, and fill in the missing values ​​according to the target repair value to obtain the target data.

7. The method according to claim 6, characterized in that, The step of classifying and processing anomalies in the cross-modal aligned data based on the comprehensive quality score to generate target data further includes the following steps: When multiple candidate repair values ​​are found within the same priority range, or when the overall quality score of a lower priority candidate repair value is greater than the overall quality score of a higher priority candidate repair value, multi-source fusion processing is triggered. In response to multi-source fusion processing, the comprehensive weight of each candidate repair value is obtained based on the mapping weight of each candidate repair value, the comprehensive quality score, the time distance weight, the high consistency weight, and the source credibility weight. When the candidate repair value is a continuous variable, the target repair value is obtained by weighting each candidate repair value according to the comprehensive weight. When the candidate repair value belongs to the wind direction variable, the sine and cosine components of each candidate repair value are weighted and averaged, and the arctangent function is used to calculate the target repair value. When the candidate repair value is a discrete variable, the target repair value is obtained by prioritizing high-confidence sources, majority voting, or determining by preset business rules.

8. A cross-modal data processing device for power meteorology, characterized in that, include: The data standardization processing module is used to respond to data access requests, standardize the accessed cross-modal raw power meteorological data to obtain cross-modal standard data, and construct a standard intermediate data model based on the cross-modal standard data. The anchor point construction module is used to construct spatiotemporal semantic anchor points based on the target object to be processed, the target anchor point time, the target spatial location, the target height, the target physical variables, and the statistical window. The data alignment module is used to perform cross-modal alignment processing on the cross-modal standard data in the standard intermediate data model based on the spatiotemporal semantic anchor points to obtain cross-modal aligned data. The quality assessment module is used to assess the quality of the cross-modal alignment data and obtain a comprehensive quality score. The anomaly handling module is used to classify anomalies and perform anomaly differentiation processing on the cross-modal aligned data based on the comprehensive quality score, and generate target data.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images