Fan operation data outlier screening method, system and storage medium

Abstract

The application discloses a fan operation data abnormal value screening method and system and a storage medium, and relates to the field of fan operation data processing.The method comprises the following steps: obtaining historical operation data and weather data of a fan group; obtaining jet index data by preprocessing the data; obtaining a fusion data set by processing all the data in time and space; constructing a boundary box model by taking fan power and jet core wind speed in the fusion data set as variables, and identifying data points outside the boundary box as suspected abnormal points; determining points with jet core wind speed values lower than a preset jet threshold as fault abnormal points; inputting geographical data and weather history data into a CFD model to calculate a theoretical fan power range for the remaining points; determining points with fan power values exceeding the theoretical fan power range as fault abnormal points, and taking the remaining points as extreme true value points; and removing the fault abnormal points to fill in data gaps and obtain a normal historical operation data set of the fan group.The application can screen out real data under extreme weather conditions and remove fault abnormal values.

Description

TECHNICAL FIELD

[0001] The present application relates to wind turbine power prediction, in particular to a wind turbine operation data outlier screening method, system and storage medium. BACKGROUND

[0002] Wind energy, as a clean and renewable energy, is increasingly occupying a larger proportion in the global energy system. However, the accuracy of wind power data directly affects the operation efficiency of wind farms and the reliability of power grid scheduling. In practical applications, wind power data collected during the operation of wind power equipment may contain outliers due to the influence of wind speed, wind direction, wind turbine internal factors and other factors. If these outlier data are not effectively processed, it may lead to an increase in wind power prediction error, affecting the health management of wind farms and the stability of power grids. Therefore, outlier detection and missing data filling and correction of wind power data have become key technical problems in wind power operation and maintenance.

[0003] In practical applications, there are mainly two types of outliers in wind turbine operation history data: one type is "fault outliers" caused by wind turbine faults, sensor errors or data transmission problems; the other type is "extreme true values" caused by extreme weather events such as gusts and storms, which are physically real but deviate from the norm.

[0004] In the prior art, conventional processing methods for handling these outliers often use simple statistical threshold methods or machine learning outlier detection algorithms (such as Isolation Forest, etc.). These methods are difficult to effectively distinguish between "fault outliers" and "extreme true values", and often exclude both, resulting in a "clean" data set, but losing valuable data representing extreme weather patterns, leading to a significant decline in prediction performance when encountering rare but critical extreme weather conditions, and failing to provide effective early warning for the power grid. SUMMARY

[0005] The purpose of the present application is to provide a wind turbine operation data outlier screening method, system and storage medium that can screen out real data under extreme weather conditions and exclude fault outliers.

[0006] Technical solution: The wind turbine operation data outlier screening method provided by the present application comprises the following steps:

[0007] Obtain the historical operation data of the target wind turbine group and the weather history data of the region to which it belongs; preprocess the weather history data to obtain gust index data;

[0008] Perform spatio-temporal alignment processing on the weather history data, gust index data and wind turbine historical operation data to obtain a fusion data set;

[0009] The wind turbine power in the historical operation data of the wind turbine set in the fusion data set and the wind core speed in the wind core index data are respectively taken as the horizontal and vertical coordinates as data points to construct a bounding box model, and the data points outside the bounding box are identified as suspected abnormal points;

[0010] It is judged whether the wind core speed value of the suspected abnormal point is lower than the preset wind core threshold, if yes, the point is determined as a fault abnormal point, otherwise, the static geographic data corresponding to the target wind turbine set and the weather history data corresponding to the data point are input into a CFD micro-scale model to calculate a theoretical wind turbine power range;

[0011] If the wind turbine power value corresponding to the data point exceeds the theoretical wind turbine power range, it is determined as a fault abnormal point, otherwise, it is determined as an extreme true value point;

[0012] The historical operation data of the wind turbine set corresponding to the fault abnormal point is removed, and the related data to be filled in the data gap is input into the trained prediction model to fill in the data gap to obtain a normal historical operation data set of the wind turbine set.

[0013] The present application extracts wind core index data from weather history data, and spatiotemporally aligns the two data with wind turbine set historical operation data to construct a comprehensive fusion data set, providing a reliable data basis for subsequent analysis; through three-level judgment, the fault abnormal value (which needs to be removed) and the extreme true value (which needs to be retained) are finely distinguished; the computational fluid dynamics (CFD) is used as the final arbitrator, so that the judgment result is no longer based on statistical probability, but on the first principle of physics, greatly improving the scientificity and reliability of the decision, solving the problem of low credibility of pure data-driven methods in extreme cases; the advantages of three-level screening: the first level is screened by the bounding box model: the calculation cost is small, the processing speed is fast, and most of the normal points are filtered out; the second level is further screened by the preset wind core threshold: the calculation cost is medium, and a part of the points obviously with faults are quickly diverted; the third level is finally screened by the theoretical wind turbine power range obtained by CFD simulation: the calculation cost is large, but only a few key suspected cases are started, this architecture ensures that the system can guarantee high efficiency when processing massive data, and can invest computing power in the most critical places to ensure the highest accuracy; through three-level screening, the extreme true value data is retained, providing valuable data for subsequent prediction, and the fault abnormal value is removed to process and screen the data to train the prediction model to fill in the data gap, so that the wind turbine historical operation data remains complete.

[0014] As a preferred, the CFD micro-scale model construction process comprises:

[0015] A height correction model based on a logarithmic wind profile is established to correct the height corresponding to the wind speed data in the weather history data to the height of the wind turbine hub;

[0016] The terrain acceleration factor lookup table method is used to give different correction coefficients to different terrains to correct the terrain effect.

[0017] The integrated physical constraint model includes air density correction, wake loss coefficient and turbulence intensity decay factor.

[0018] The standard power curve is established by piecewise linear interpolation, cubic spline interpolation is carried out between key points, and uncertainty is introduced to obtain the theoretical fan power range.

[0019] By integrating height correction, terrain effect correction, air density correction, wake loss and turbulence intensity decay, a high-fidelity engineering theoretical fan power range calculation model is constructed, so that the output power range is closer to the actual operation physical law of the fan, and a solid physical basis is provided for final arbitration.

[0020] The fan operation data outlier screening system comprises:

[0021] The data acquisition and preprocessing module is used to obtain the historical operation data of the target fan group and the weather historical data of the region to which it belongs; the weather historical data is preprocessed to obtain the jet index data;

[0022] The data fusion module is used to perform time and space alignment processing on the weather historical data, jet index data and fan historical operation data to obtain a fusion data set;

[0023] The fault value screening module is used to take the fan power in the fan group historical operation data and the jet core wind speed in the jet index data as the horizontal and vertical coordinates as data points to construct a boundary box model, and mark the data points outside the boundary box as suspected abnormal points;

[0024] Determine whether the jet core wind speed value of the suspected abnormal point is lower than the preset jet threshold value, if yes, the point is determined as a fault abnormal point; otherwise, the static geographic data corresponding to the target fan group and the weather historical data corresponding to the data point are input into the CFD microscale model to calculate the theoretical fan power range;

[0025] If the fan power value corresponding to the data point exceeds the theoretical fan power range, it is determined as a fault abnormal point, otherwise it is determined as an extreme true value point;

[0026] The data filling module is used to eliminate the fan group historical operation data corresponding to the fault abnormal point, and input the related data of the data gap to be filled into the trained prediction model to fill the data gap to obtain the normal historical operation data set of the fan group.

[0027] The computer-readable storage medium storing one or more programs of the present application includes one or more programs including instructions that, when executed by a computing device, cause the computing device to perform the above method.

[0028] Beneficial effects: The present application lays a foundation for abnormal data screening through multi-source data fusion, and takes into account data processing efficiency and screening accuracy through a three-level progressive screening method, ensuring that the training data is removed from fault noise while retaining valuable extreme weather data, thereby providing valuable data for prediction under extreme weather conditions, enabling timely and accurate prediction and early warning, and providing more reliable technical support for the safe and stable operation of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The figure is a schematic diagram of the overall process of the method of the present application;

[0030] Figure 2 The figure is a schematic diagram of the spatio-temporal alignment process for obtaining a fused data set;

[0031] Figure 3 The figure is a three-level judgment flowchart for screening extreme true values of the present application;

[0032] Figure 4 The figure is a schematic diagram of the wind turbine unit historical data set after removing fault outliers and filling data gaps. DETAILED DESCRIPTION

[0033] As shown in Figures 1-4 The wind turbine operation data outlier screening method of the present application comprises the following steps:

[0034] Obtain the target wind turbine unit historical operation data and the weather historical data of the region to which it belongs; preprocess the weather historical data to obtain the jet index data;

[0035] The target wind turbine unit historical operation data is read from the historical database of the monitoring and data acquisition system inside the wind farm, and the wind turbine unit historical operation data, i.e., SCADA data, is high-frequency sampled wind turbine operation data. SCADA data mainly represents high-frequency time series data, and its core elements include wind speed, wind direction, wind turbine power, generator speed, pitch angle, status code, etc., directly reflecting the true operating state of the wind farm.

[0036] Historical weather data can be obtained using GFS data. GFS data and corresponding historical reanalysis field data can be obtained through API interfaces. GFS data is usually represented as a four-dimensional data grid, such as longitude, latitude, pressure layer, and time. Its core elements include temperature, zonal wind U, meridional wind V, relative humidity, etc. GFS data can cover the globe, with a spatial resolution of 0.25°×0.25° or 0.5°×0.5° and a temporal resolution of 3 hours or 6 hours.

[0037] Since the raw GFS data cannot be used directly, preprocessing is required to extract key information. This involves extracting specific pressure layers from the GFS data. Because the upper-level jet stream mainly occurs in the 200-400 hPa range, the zonal wind (U) and meridional wind (V) components of this range are extracted. Using these U and V components, the overall wind speed and wind direction are calculated.

[0038] Wind Speed ​​= sqrt(U² + V²)

[0039] Wind Direction = arctan2(U,V) (requires quadrant conversion)

[0040] The system scans the grid data in the target area (i.e., the wind farm monitoring area corresponding to the wind turbine) to find areas where the wind speed exceeds the preset wind speed threshold. These areas are candidate areas for the jet stream core. Then, by finding the local maximum wind speed point, the position of the jet stream axis is accurately determined. Since the maximum wind speed value on the jet stream axis represents the intensity of the jet stream, the difference between the latitude and longitude of the jet stream core point and the latitude and longitude of the center of the wind farm is calculated. This difference is converted into distance and direction to obtain jet stream index data. The spatial relative relationship between the jet stream and the wind farm is quantified based on the jet stream index data. The closer the distance, the greater the influence. The jet stream index data usually includes: the maximum wind speed of the jet stream core (i.e., the jet stream core wind speed), the relative distance between the jet stream core and the wind farm, and the azimuth angle of the jet stream core relative to the wind farm.

[0041] Traditional wind power forecasting only uses surface or low-altitude wind speeds, neglecting the influence of upper-level jet streams on surface wind speed variability. In particular, extreme winds significantly impact low-altitude wind speeds. Therefore, models must consider the impact of wind speeds under extreme wind conditions during forecasting. High maximum wind speeds in the jet stream core near the wind farm indicate a high probability of abnormal power fluctuations due to extreme winds during that period. By calculating jet stream index data, the impact of extreme weather is incorporated into data processing, thus improving the model's physical foundation from the outset. Furthermore, jet stream index data provides extreme meteorological evidence for subsequent outlier screening. Preprocessing GFS data to obtain jet stream index data reduces subsequent data complexity while retaining core extreme meteorological information affecting wind farm power output, thereby improving the efficiency of subsequent data processing and calculations.

[0042] A fused dataset is obtained by performing spatiotemporal alignment processing on historical weather data, jet stream index data, and historical wind turbine operation data.

[0043] Spatiotemporal alignment processing includes: using the timestamps of historical wind turbine operation data as a benchmark, performing cubic spline interpolation on historical weather data and jet stream index data to achieve time alignment of the three types of data.

[0044] Specifically, due to the asynchronous timestamps of the three types of data, SCADA data is high-frequency measured data (e.g., one point every 15 minutes), while GFS data and its derived jet stream index data are low-frequency forecast data (e.g., one point every 3 hours). Using the timestamps of the SCADA data as a benchmark, interpolation is performed on the low-frequency GFS and jet stream index data. The timestamp T of a SCADA data point is read, and the two closest time points T1 and T2 before and after T in the GFS data are found. Cubic spline interpolation is used, and based on the GFS forecast values ​​at T1 and T2, the interpolated estimates of the GFS meteorological elements and jet stream index at time T are calculated, thus aligning the three data sets in time. This ensures that each SCADA data point's timestamp has corresponding and continuous historical weather data (GFS data) and jet stream index data.

[0045] Spatial alignment of the three types of data is achieved by mapping historical weather data and jet stream index data to the wind turbine geographical location in historical wind turbine operation data through bilinear interpolation.

[0046] Specifically, since GFS data is regular grid data while SCADA data is point data at specific locations within the wind farm, and the two are not spatially aligned, the GFS grid data is interpolated to the specific geographical location of the wind farm. Based on the precise latitude and longitude coordinates of the wind farm center, the GFS grid cell containing it is located. Using bilinear interpolation, the meteorological element values ​​for the wind farm's location are calculated based on the values ​​of the surrounding four grid points, thus ensuring that the GFS data matches the actual environment of the wind farm. The jet stream index data is derived from GFS, so the spatial alignment process is the same as for GFS, and will not be elaborated further. Therefore, the GFS data, jet stream index data, and SCADA data are fused together in both time and space to obtain a unified dataset.

[0047] First-level judgment: Use the wind turbine power in the historical operation data of the wind turbine units in the fused dataset and the core wind speed of the jet flow index data as the horizontal and vertical coordinates to construct a bounding box model, and mark the data points outside the bounding box as suspected outliers.

[0048] The bounding box model is constructed as follows:

[0049] Based on SCADA data and jet stream index data, a scatter plot is constructed with the core jet stream wind speed as the x-axis and the turbine power as the y-axis. DBSCAN clustering is used to perform density analysis on the data points in the scatter plot, and an irregular α-Shape closed boundary is automatically fitted to surround the high-density normal data area. This closed boundary is the bounding box model.

[0050] The bounding box model defines the decision boundary for the distribution range of "normal operating data points" on a "wind speed-power" scatter plot; points outside the boundary are considered suspected anomalies. The density-based DBSCAN algorithm is used for data distribution analysis. The algorithm controls the clustering effect through two key parameters: the neighborhood radius ε and the minimum number of samples min_samples. Innovatively, a k-distance curve method is used to automatically optimize parameter selection: the distance from each point to its k-th nearest neighbor is calculated, sorted, and plotted as a curve. The distance value at the inflection point of the curve is selected as the optimal ε parameter, and the min_samples parameter is adaptively set to the square root of the total number of data points based on the dataset size. During algorithm execution, unvisited points are first randomly selected, and all points within their ε-neighborhood are found. If the number of points within the neighborhood reaches the min_samples threshold, a new cluster is created, and this cluster is recursively expanded to include all density-reachable points. Points with insufficient points within their neighborhood are marked as noise points (anomalies). This process continues until all points have been visited.

[0051] Traditional methods, which manually define a rectangular boundary or a boundary based on a simple sigma rule, fail to effectively capture the nonlinearity and local density variations of wind turbine power characteristic curves. This leads to the erroneous classification of many normal data points at the extremes of the power curve as anomalies, while simultaneously missing some hidden anomalies within normal data density. By using the core jet velocity as the horizontal axis and wind turbine power as the vertical axis, all data points are plotted on a two-dimensional plane, forming a scatter plot. Under normal circumstances, these points are closely distributed around an "S"-shaped curve, reflecting the standard power characteristics of the wind turbine. After identifying all anomalies, the algorithm uses the α-Shape algorithm to generate an irregular closed curve that encloses the core, high-density cluster of normal data points, ultimately forming a "boundary box model." This model maximizes the retention of reasonable data points while accurately excluding statistically abnormal points. The boundary box model can automatically learn and generate from historical data of any wind farm, eliminating the need for manual threshold setting based on expert experience. This achieves adaptive and automated anomaly identification, and the fitted irregular boundary significantly reduces false alarm and false negative rates.

[0052] Anomaly screening is performed using a bounding box model to quickly filter out obviously abnormal points. By using the core jet speed value from the jet index data and the wind speed power value from the SCADA data for each data point in the fused dataset as a coordinate point, it is matched with the pre-built bounding box model to check whether the data point is located outside the closed boundary of the bounding box. If the data point falls inside the box, it is marked as "normal" and the process ends. If the data point falls outside the box, it is marked as "suspected anomaly" and a second-level judgment is triggered.

[0053] Second-level judgment: Determine whether the core wind speed of the jet stream at the suspected anomaly point is lower than the preset jet stream threshold. If so, the point is determined to be a fault anomaly point; otherwise, input the static geographic data corresponding to the target wind turbine and the historical weather data corresponding to the data point into the CFD microscale model to calculate the theoretical wind turbine power range.

[0054] The preset rapid flow threshold includes a short-term threshold and a long-term threshold, which are determined as follows:

[0055] Short-term jet threshold: Select the core wind speed of the jet within the first preset time range before the time of the suspected anomaly point to be judged to form the first jet core wind speed set. Select the lower quantile of the set or the jet core wind speed in front of it as the short-term jet threshold. Here, the 10th quantile is taken as the short-term threshold. The first preset time range can be 24 hours, or the range can be adjusted according to actual needs.

[0056] Long-term jet threshold: Select the core wind speed of the jet within the second preset time range before the time of the suspected anomaly point to be determined to form the second jet core wind speed set. Select the lower quantile of this set or the jet core wind speed in front of it as the long-term jet threshold. Here, the 5th quantile is taken as the short-term threshold. The second preset time range can be 30 days, or the range can be adjusted according to actual needs.

[0057] The first preset time range should be smaller than the second preset time range.

[0058] If the core jet velocity value of a suspected fault location is simultaneously less than two jet thresholds, it is considered to be below the preset jet threshold and thus identified as a fault anomaly. Otherwise, the static geographic data corresponding to the target wind turbine and the historical weather data corresponding to that data point are input into the CFD microscale model to calculate the theoretical wind turbine power range. The process then proceeds to the third level of judgment.

[0059] The method for constructing the CFD microscale model of this invention specifically includes:

[0060] First, a height correction model based on logarithmic wind profiles is established to correct the height corresponding to the wind speed data in historical weather data to the height of the wind turbine hub.

[0061] The expression for the height-corrected model is:

[0062]

[0063] in, and These are the wind speed values ​​from historical weather data and the corrected wind speeds, respectively. and These represent the altitude corresponding to the wind speed value in historical weather data and the altitude of the wind turbine hub, respectively. This is the roughness length; in this embodiment... and Take 100 meters and 80 meters respectively. The depth is 0.1-0.3 meters depending on the landform type.

[0064] Secondly, a terrain effect correction module was constructed. Using the terrain acceleration factor lookup table method, correction coefficients of 1.00, 1.08, 1.15, 0.92 and 1.05 were assigned to plains, hills, mountains, valleys and coastal terrain, respectively.

[0065] Then, the physical constraint model is integrated, including the air density correction based on the ideal gas law, expressed as follows:

[0066]

[0067] in, P is the corrected air density, R is the atmospheric pressure, R is the dry air gas constant with a value of 287.05 J / (kg·K), and T is the absolute temperature.

[0068] The fixed wake loss factor of 0.97 indicates that the model predicts the power of a single unit and roughly considers the wake effect from the upwind turbine (3% loss); and the turbulence intensity linear attenuation factor (1-0.1×I), where high turbulence intensity I reduces the wind energy conversion efficiency, and the linear attenuation factor (1-0.1×I) attempts to quantify this effect.

[0069] Finally, a standard power curve was established by piecewise linear interpolation. Cubic spline interpolation was performed between key points of cut-in wind speed of 3 m / s, rated wind speed of 12.5 m / s, and cut-out wind speed of 25 m / s (these are determined by the characteristics and operating strategies of the wind turbine, and are usually determined by the technical parameters of the wind turbine manufacturer). 15% basic uncertainty was introduced and superimposed with 5-10% wind speed-related uncertainty to obtain the theoretical wind turbine power range, thus completing the construction of the engineering CFD microscale model.

[0070] The third level of judgment: If the wind turbine power value corresponding to the data point exceeds the theoretical wind turbine power range, it is judged as a fault point; otherwise, it is judged as an extreme true value point.

[0071] The historical operating data of the wind turbine corresponding to the fault point is removed, and the prediction model is trained using this data. The relevant data to be filled in the data gaps are then input into the trained prediction model to fill the data gaps, thus obtaining the normal historical operating dataset of the wind turbine.

[0072] Data points marked as "outliers" are removed from the original SCADA data sequence. Directly removing outliers results in gaps in the data sequence, causing discontinuity in the time series. Since most prediction models require continuous input data, data gap filling is necessary. This invention uses a random forest model for this purpose. The specific process is as follows:

[0073] Training the populated model: Using the high-quality data retained after removing outliers, including "normal values" and "extreme true values", a random forest regression model is trained.

[0074] Model features: The features are timestamps (reflecting periodicity and trends) and GFS data at the corresponding time (reflecting meteorological conditions).

[0075] Model label: The actual power value is used as the label.

[0076] Perform filling: For each empty space to be filled, input its corresponding timestamp and GFS data into the trained random forest model. The model will output a physically reasonable power prediction value to fill the empty space.

[0077] This method is superior to traditional interpolation because it takes into account the influence of meteorological factors, and the filled values ​​more accurately reflect the output level that the wind turbine should have under the meteorological conditions at that time, rather than just smoothing over time. After elimination and filling, we obtained a continuous, complete, and high-quality wind power time series data, which contains neither interfering fault data nor any valuable extreme weather data.

[0078] The wind turbine operation data outlier filtering system of the present invention includes:

[0079] Data acquisition and preprocessing module: used to acquire historical operating data of the target wind turbine and historical weather data of its region; preprocess the historical weather data to obtain jet stream index data;

[0080] Data fusion module: used to perform spatiotemporal alignment processing on historical weather data, jet stream index data, and historical wind turbine operation data to obtain a fused dataset;

[0081] Fault value filtering module: Used to construct a bounding box model by taking the wind turbine power from the historical operation data of the wind turbine units in the fused dataset and the core wind speed of the jet flow index data as the horizontal and vertical coordinates respectively, and marking the data points outside the bounding box as suspected anomalies;

[0082] Determine whether the core wind speed of the jet stream at the suspected anomaly point is lower than the preset jet stream threshold. If so, the point is identified as a fault anomaly point. Otherwise, input the static geographic data corresponding to the target wind turbine and the historical weather data corresponding to the data point into the CFD microscale model to calculate the theoretical wind turbine power range.

[0083] If the wind turbine power value corresponding to the data point exceeds the theoretical wind turbine power range, it is determined to be a fault point; otherwise, it is determined to be an extreme true value point.

[0084] Data imputation module: This module removes the historical operating data of the wind turbine corresponding to the fault or anomaly point, uses this data to train the prediction model, and inputs the relevant data to be filled into the trained prediction model to fill the data gaps and obtain the normal historical operating dataset of the wind turbine.

[0085] The computer-readable storage medium for storing one or more programs according to the present invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform the method described above.

[0086] To better illustrate the method of the present invention, a specific example is provided below for further explanation:

[0087] A wind field (33 wind turbines, coordinates: 117°42′57″-117°53′58″ E, 34°26′50″-34°32′51″ N, turbine height approximately 80m) was selected in Pizhou City, Jiangsu Province. Historical wind turbine operating data (SCADA data), including wind speed (m / s) and turbine power (kW), was used, from July 2024 to January 2025, with a time resolution of 15 minutes per data point. Historical weather data: Historical reanalysis data from the Global Forecast System (GFS) of the National Center for Environmental Prediction (NCEP) was selected, including temperature (K), zonal wind U (m / s), meridional wind V (m / s), and atmospheric pressure (Pa). The spatial resolution was 0.25° × 0.25°, and the temporal resolution was 6 hours. Data from the 300 hPa pressure layer was particularly focused on to capture upper-level jet streams.

[0088] Jet index extraction: The U and V components of the 300 hPa pressure layer are extracted from GFS data to calculate the wind speed and wind direction.

[0089] Wind Speed ​​= sqrt(U² + V²)

[0090] Wind Direction = arctan2(U,V) (requires quadrant conversion)

[0091] Scan the target area (wind farm and its surrounding area) to identify grid areas where wind speeds consistently exceed 30 m / s for more than one hour, defining these as "jet core candidate areas." Within each candidate area, locate the point with the highest local wind speed, precisely identifying it as the "jet core point." Calculate jet index data, including: Jet Core Speed: the maximum wind speed value at the core point; Distance: the great circle distance between the jet core point and the center point of the wind farm; and Bearing: the direction of the jet core relative to the wind farm.

[0092] Time alignment: Using the SCADA data timestamps (every 15 minutes) as a baseline, low-frequency GFS data (every 6 hours) and jet stream index data are interpolated to each SCADA timestamp using cubic spline interpolation to generate continuous meteorological and jet stream time series. Spatial alignment: Using the latitude and longitude of the wind farm center point as a baseline, the GFS grid data (latitude and longitude grid) and jet stream index data are interpolated to the wind farm center location using bilinear interpolation to obtain meteorological element values ​​that match the actual location of the wind farm.

[0093] Three-level progressive outlier screening:

[0094] The first-level screening: a rapid initial screening based on a bounding box model. A two-dimensional scatter plot is formed, with the core jet stream wind speed of each data point in the fused dataset as the x-axis and the turbine power as the y-axis. Density clustering is performed using the DBSCAN algorithm (parameters adaptively determined via a k-distance curve), and the α-Shape algorithm is used to generate closed boundaries, i.e., "bounding boxes," that enclose the core normal data points. Each data point is compared to its bounding box. Points within the boundary are marked as "normal points," and the process terminates. Points outside the boundary are marked as "suspected outliers" and proceed to the second-level screening. This step identified a total of 5405 suspected outliers.

[0095] The second level of screening involves physical diversion based on preset jet stream thresholds. The short-term jet stream threshold is the 10th quartile of the core jet stream wind speed data within the 24 hours preceding the suspected anomaly point; the long-term jet stream threshold is the 5th quartile of the core jet stream wind speed data within the 30 days preceding the suspected anomaly point. The core jet stream wind speed value at each suspected anomaly point is checked. If the value is lower than both the short-term and long-term thresholds, it indicates that there is no significant extreme weather activity at that time, and the anomaly is highly likely caused by wind turbine failure or sensor error; therefore, it is directly identified as a "fault anomaly point." Otherwise, it indicates that the point may be under extreme weather conditions, requiring further physical verification, and proceeding to the third level of screening.

[0096] Contribution of the jet flow index to abnormal shunting

[0097] Indicator Explanation: This quantifies how many "suspected outliers" are correctly diverted through the rapid flow index.

[0098] Evaluation method: Statistically count the number of data points that were directly judged as "fault abnormal value" in the second-level judgment because the core wind speed of the jet stream was lower than the dynamic threshold (low), and the number of data points that entered the third-level CFD judgment because the core wind speed of the jet stream was higher than the dynamic threshold (high).

[0099] formula:

[0100]

[0101] Reflecting the importance of the rapid flow index, it undertook the task of diverting the flow, effectively reducing the computational burden on subsequent CFD models.

[0102] Table 1. Statistics on the number of data points of different types

[0103]

[0104]

[0105] Results of this step: Of the 5405 suspected anomalies, 1782 were identified as faulty anomalies and removed, resulting in a diversion ratio of 32.969%.

[0106] The third level of screening involves physical arbitration based on a CFD microscale model. For points entering this level, their corresponding static geographic data and historical weather data are input into a pre-constructed CFD microscale model. First, a height correction model based on logarithmic wind profiles is established to correct the height corresponding to wind speed data in historical weather data to the height of the wind turbine hub. The expression for the height correction model is as follows:

[0107]

[0108] in, and These are the wind speed values ​​from historical weather data and the corrected wind speeds, respectively. and These represent the altitude corresponding to the wind speed value in historical weather data and the altitude of the wind turbine hub, respectively. This is the roughness length; in this embodiment... and Take 100 meters and 80 meters respectively. The depth is 0.1-0.3 meters depending on the landform type.

[0109] A terrain effect correction module was constructed, and the terrain acceleration factor lookup table method was used to assign correction coefficients of 1.00, 1.08, 1.15, 0.92 and 1.05 to plains, hills, mountains, valleys and coastal terrains respectively.

[0110] Then, the physical constraint model is integrated, including the air density correction based on the ideal gas law, expressed as follows:

[0111]

[0112] in, P represents the corrected air density, P represents atmospheric pressure, R is the dry air gas constant with a value of 287.05 J / (kg·K), and T represents absolute temperature. A fixed wake loss factor of 0.97 indicates that the model predicts the power of a single unit and roughly considers the wake effect from the upwind turbine (a loss of 3%). A turbulence intensity linear attenuation factor (1-0.1×I) is also included, where high turbulence intensity I reduces wind energy conversion efficiency, and the linear attenuation factor (1-0.1×I) attempts to quantify this effect.

[0113] Finally, a standard power curve was established through piecewise linear interpolation. Cubic spline interpolation was performed between key points: an inrush wind speed of 3 m / s, a rated wind speed of 12.5 m / s, and an outrush wind speed of 25 m / s (these are determined by the characteristics and operating strategies of the wind turbine, typically through the technical parameters of the wind turbine manufacturer). A 15% basic uncertainty was introduced, superimposed with a 5-10% wind speed-related uncertainty to obtain the theoretical wind turbine power range, thus completing the construction of the engineering-scale CFD model. The actual wind turbine power at the data points was compared with the theoretical power range calculated by the CFD model. If the actual power exceeded the theoretical range, it was identified as a "fault anomaly point." If the actual power was within the theoretical range, it was considered a physically real "extreme truth point" and retained.

[0114] All data marked as "fault anomalies" were removed from the original SCADA sequence, resulting in a high-quality training dataset containing both "normal points" and "extreme ground truth points." Using this training set, a random forest regression model was trained, using timestamps and corresponding GFS meteorological data as features and wind turbine power as a label. The timestamps and meteorological data to be filled were then input into the trained model, predicting reasonable power values ​​for filling. This process ultimately generates a continuous, complete, and high-quality dataset of normal historical wind turbine operation. Specific results are shown below. Figure 4 As shown.

Claims

1. A method for filtering outliers in wind turbine operating data, characterized in that, Includes the following steps: Acquire historical operating data of the target wind turbine and historical weather data of its region; preprocess the historical weather data to obtain jet stream index data; A fused dataset is obtained by spatiotemporally aligning historical weather data, jet stream index data, and historical wind turbine operation data. The wind turbine power from the historical operation data of the wind turbine units in the fused dataset and the wind speed from the jet flow index data are used as the horizontal and vertical axes to construct a bounding box model. Data points outside the bounding box are marked as suspected outliers. Determine whether the core wind speed of the jet stream at a suspected abnormal point is lower than the preset jet stream threshold. If so, the point is identified as a fault abnormal point. Otherwise, input the static geographic data corresponding to the target wind turbine and the historical weather data corresponding to the data point into the CFD microscale model to calculate the theoretical wind turbine power range; If the wind turbine power value corresponding to the data point exceeds the theoretical wind turbine power range, it is determined to be a fault point; otherwise, it is determined to be an extreme true value point. The CFD microscale model construction process includes: Establish a height correction model based on logarithmic wind profiles to correct the height corresponding to wind speed data in historical weather data to the height of the wind turbine hub. The terrain acceleration factor lookup table method is used to assign different correction coefficients to different terrains in order to correct the terrain effect; An integrated physical constraint model is used, including air density correction, wake loss coefficient, and turbulence intensity attenuation factor. A standard power curve is established by piecewise linear interpolation, cubic spline interpolation is performed between key points, and uncertainty is introduced to obtain the theoretical power range of the wind turbine. The historical operating data of the wind turbine corresponding to the fault point is removed, and the prediction model is trained using this data. The relevant data to be filled in the data gaps are then input into the trained prediction model to fill the data gaps, thus obtaining the normal historical operating dataset of the wind turbine.

2. The method according to claim 1, characterized in that, The preprocessing of historical weather data includes the following steps: extracting the zonal and meridional wind components of the preset altitude pressure layer from the historical weather data; Calculate the combined wind speed of zonal and meridional winds, and identify the area with a combined wind speed greater than a preset wind speed threshold as the core of the jet stream; Calculate the distance and direction vector between the jet stream core and the center of the target wind turbine, and use the obtained distance and direction vectors and the maximum wind speed at the jet stream core as the jet stream index data.

3. The method according to claim 1, characterized in that, The spatiotemporal alignment process includes the following steps: Based on the timestamps of historical wind turbine operation data, cubic spline interpolation is performed on historical weather data and jet stream index data to achieve time alignment of the three types of data. Spatial alignment of the three types of data is achieved by mapping historical weather data and jet stream index data to the wind turbine geographical location in historical wind turbine operation data through bilinear interpolation.

4. The method according to claim 1, characterized in that: The bounding box model is constructed using the DBSCAN algorithm.

5. The method according to claim 1, characterized in that, The preset swift flow threshold includes a short-term swift flow threshold and a long-term swift flow threshold, which are determined as follows: Short-term jet threshold: Select the core wind speed of the jet within the first preset time range before the time of the suspected anomaly point to be judged to form the first set of core wind speeds of the jet, and select the lower quantile of the set or the core wind speed of the jet ahead of it as the short-term jet threshold. Long-term jet threshold: Select the core wind speed of the jet within the second preset time range before the time of the suspected anomaly point to be determined to form a second set of core jet wind speeds, and select the lower quantile of the set or a jet core wind speed ahead of it as the long-term jet threshold. The first preset time range is less than the second preset time range; when the core wind speed value of the jet stream at the suspected fault point is less than both jet stream thresholds at the same time, it is determined to be less than the preset jet stream threshold.

6. The method according to claim 1, characterized in that, The expression for the height correction model is: ， in, and These are the wind speed values ​​from historical weather data and the corrected wind speeds, respectively. and These represent the altitude corresponding to the wind speed value in historical weather data and the altitude of the wind turbine hub, respectively. Roughness length; The expression for the air density correction is: ， in, P is the corrected air density, R is the atmospheric pressure, and T is the dry air gas constant.

7. The method according to claim 1, characterized in that: The prediction model adopts a random forest model; the prediction model is trained using historical operating data of wind turbines that provide data corresponding to fault anomalies.

8. A system for filtering outomas in wind turbine operating data, characterized in that, include: Data acquisition and preprocessing module: used to acquire historical operating data of the target wind turbine and historical weather data of its region; Preprocessing historical weather data yields jet stream index data; Data fusion module: used to perform spatiotemporal alignment processing on historical weather data, jet stream index data, and historical wind turbine operation data to obtain a fused dataset; Fault value filtering module: Used to construct a bounding box model by taking the wind turbine power from the historical operation data of the wind turbine units in the fused dataset and the core wind speed of the jet flow index data as the horizontal and vertical coordinates respectively, and marking the data points outside the bounding box as suspected anomalies; Determine whether the core wind speed of the jet stream at a suspected abnormal point is lower than the preset jet stream threshold. If so, the point is identified as a fault abnormal point. Otherwise, input the static geographic data corresponding to the target wind turbine and the historical weather data corresponding to the data point into the CFD microscale model to calculate the theoretical wind turbine power range; If the wind turbine power value corresponding to the data point exceeds the theoretical wind turbine power range, it is determined to be a fault point; otherwise, it is determined to be an extreme true value point. The CFD microscale model construction process includes: Establish a height correction model based on logarithmic wind profiles to correct the height corresponding to wind speed data in historical weather data to the height of the wind turbine hub. The terrain acceleration factor lookup table method is used to assign different correction coefficients to different terrains in order to correct the terrain effect; An integrated physical constraint model is used, including air density correction, wake loss coefficient, and turbulence intensity attenuation factor. A standard power curve is established by piecewise linear interpolation, cubic spline interpolation is performed between key points, and uncertainty is introduced to obtain the theoretical power range of the wind turbine. Data imputation module: This module removes the historical operating data of the wind turbine corresponding to the fault or anomaly point, uses this data to train the prediction model, and inputs the relevant data to be filled into the trained prediction model to fill the data gaps and obtain the normal historical operating dataset of the wind turbine.

9. A computer-readable storage medium for storing one or more programs, characterized in that: The program includes one or more instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1 to 7.

Images