Centralized wind farm power prediction method and system based on multi-source data fusion

Abstract

The application relates to the technical field of wind power prediction, and discloses a centralized wind farm power prediction method and system based on multi-source data fusion. The method comprises the following steps: cleaning and aligning multi-source meteorological data; fusing a composite credibility weight and a candidate abnormal source according to historical errors and source deviation degrees; constructing a wind farm group adjacency matrix, correcting the fused data according to a support rate correction weight; and inputting a graph convolution spatio-temporal network to obtain a centralized power prediction result. The application improves the identification capability of local meteorological anomalies and the accuracy of multi-wind farm joint prediction results in the meteorological data fusion link of regional centralized wind farm group power prediction.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, and in particular to a centralized wind farm power prediction method and system based on multi-source data fusion. Background Technology

[0002] Centralized wind farm power prediction methods based on multi-source data fusion typically integrate multiple meteorological data sources, such as numerical weather prediction, meteorological tower measured data, radar inversion data, and satellite inversion data, into a unified prediction system to jointly predict the power of multiple wind farms within a region. Existing technologies usually first acquire raw data from multiple meteorological data sources, then assign fixed weights or segmented weights based on historical forecast errors of each data source or wind speed intervals, and use the weighted meteorological data as input to a unified prediction system. This combined with a joint prediction model outputs the power prediction results for each wind farm within the region. This approach replaces the traditional "one wind farm, one measurement" model, where each wind farm has its own independent prediction system, and data aggregation between stations is not possible.

[0003] However, existing static or segmented weighting schemes based on historical errors rely solely on the historical statistical error between forecast values ​​and meteorological tower measurements—a single verification dimension. This verification dimension is essentially a post-hoc statistic. When a regional meteorological system undergoes a sudden change in a short period, even a meteorological data source with excellent historical performance may become inaccurate at the moment of system switching. The historical error weights cannot reflect this inaccuracy in a timely manner, resulting in the fused meteorological data still using the existing weights of the inaccurate meteorological sources at the moment of meteorological change. Furthermore, while existing regional centralized forecasting systems have achieved unified access to data from multiple wind farms at the data aggregation level, the weighting of each meteorological data source in the meteorological fusion process is still performed independently by each wind farm. This fails to establish a correlation between wind farms based on geographical location and historical power performance, and also fails to use this correlation to verify the weighting results of the meteorological data sources.

[0004] If other verification dimensions besides historical errors are introduced into the multi-source meteorological data fusion process, such as the degree of deviation between meteorological data sources at the same time, to compensate for the problem that the weight of historical errors lags behind meteorological changes, then this real-time verification dimension is usually based on the judgment logic that "if the output of most meteorological data sources is consistent, it is correct". At this time, if a small-scale real meteorological anomaly such as local gusts or local convective cells induced by terrain occurs at a certain location in the region, often only a few meteorological data sources can correctly capture the anomaly, while the majority of meteorological data sources fail to capture it due to spatial resolution or coverage limitations and still maintain false consistency with each other. According to the majority consistency logic, the minority meteorological data sources that truly reflect local anomalies will be judged as observation errors and their weights will be reduced, resulting in the effective local meteorological information being filtered out during the fusion process. In existing regional centralized forecasting systems, each wind farm independently determines the weight of meteorological data sources, lacking a cross-comparison mechanism for meteorological observation results from adjacent stations within the wind farm cluster. This means that misjudgments by a single station regarding a minority of meteorological signals cannot be verified or corrected by synchronous observations from other stations in the region that are closely related to it. Such misjudgments are then carried into subsequent joint forecasting models, affecting the accuracy of power forecasting results for multiple wind farms in the region. Summary of the Invention

[0005] This application provides a centralized wind farm power prediction method and system based on multi-source data fusion. It solves the problems that the judgment of the credibility of meteorological data sources lags behind meteorological changes during the multi-source meteorological data fusion process, and that the judgment based on majority consensus is prone to misjudging local real meteorological anomalies as observation noise and losing effective meteorological information. It improves the ability of meteorological data fusion in the power prediction of regional centralized wind farm groups to identify local meteorological anomalies and the accuracy of the joint prediction results of multiple wind farms.

[0006] In a first aspect, this application provides a centralized wind farm power prediction method based on multi-source data fusion, the centralized wind farm power prediction method based on multi-source data fusion comprising: Step S1: Clean and align the multi-source meteorological data with spatiotemporal references to obtain a spatiotemporally aligned meteorological dataset; Step S2: Based on the spatiotemporally aligned meteorological dataset, perform weighted fusion processing on the historical errors and inter-source deviation of each meteorological data source to obtain the composite confidence weight and fused meteorological input vector, and obtain candidate abnormal meteorological data sources by filtering the inter-source deviation. Step S3: Construct a weighted adjacency matrix for wind farm clusters based on the geographical distance attenuation coefficient and historical power correlation coefficient of wind farms. Filter wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold to obtain the associated wind farm set corresponding to the candidate abnormal meteorological data source. Statistically process the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate. When the support rate reaches a preset support rate threshold, amplify the composite credibility weight corresponding to the candidate abnormal meteorological data source, using the highest historical credibility score as the upper limit. When the support rate does not reach the preset support rate threshold, attenuate the composite credibility weight. Reweight the fused meteorological input vector with the corrected composite credibility weight to obtain the corrected fused meteorological input vector. Step S4: Process the corrected and fused meteorological input vector into the convolutional spatiotemporal prediction network to obtain the power prediction results of the centralized wind farm group.

[0007] Secondly, this application provides a centralized wind farm power prediction system based on multi-source data fusion, the centralized wind farm power prediction system based on multi-source data fusion comprising: The alignment module is used to clean and align multi-source meteorological data with spatiotemporal references to obtain a spatiotemporally aligned meteorological dataset. The weighting module is used to perform weighted fusion processing on the historical errors and inter-source deviation of each meteorological data source according to the spatiotemporally aligned meteorological dataset, to obtain the composite confidence weight and fused meteorological input vector, and to obtain candidate abnormal meteorological data sources by filtering the inter-source deviation. The attenuation module is used to construct a weighted adjacency matrix of wind farm groups based on the attenuation coefficient of wind farm geographical distance and the correlation coefficient of historical power. It selects the associated wind farm set corresponding to the candidate abnormal meteorological data source by filtering the wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold. It performs statistical processing on the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate. When the support rate reaches a preset support rate threshold, it amplifies the composite credibility weight corresponding to the candidate abnormal meteorological data source and uses the highest historical credibility score as the upper limit. When the support rate does not reach the preset support threshold, it attenuates the composite credibility weight. The fused meteorological input vector is re-weighted by the corrected composite credibility weight to obtain the corrected fused meteorological input vector. The prediction module is used to process the corrected and fused meteorological input vector into the graph convolutional spatiotemporal prediction network to obtain the power prediction results of the centralized wind farm group.

[0008] Thirdly, a centralized wind farm power prediction device based on multi-source data fusion is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the centralized wind farm power prediction device based on multi-source data fusion to execute the aforementioned centralized wind farm power prediction method based on multi-source data fusion.

[0009] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described centralized wind farm power prediction method based on multi-source data fusion.

[0010] The technical solution provided in this application obtains a spatiotemporally aligned meteorological dataset by cleaning and aligning multi-source meteorological data with spatiotemporal references. This unifies the grid data of numerical weather prediction, the surface data retrieved by radar and satellite, and the point-time series data of wind towers and SCADA systems to the same geographic coordinate reference and time resolution. This provides a unified data structure that can be directly computed for time-by-time and location-by-location comparisons and weighted fusions between different meteorological data sources in subsequent steps, avoiding fusion errors caused by heterogeneous data sources. Furthermore, by weighting and fusing the historical errors and inter-source deviations of each meteorological data source, a composite reliability weight is obtained. This ensures that the reliability judgment of the meteorological data source simultaneously covers both historical statistics and real-time horizontal comparisons. The independent verification paths, historical errors reflecting the long-term stable performance of meteorological data sources under similar operating conditions, and inter-source deviation reflecting the instantaneous consistency between meteorological data sources at the current moment, combine the two paths to overcome the limitation of relying solely on historical statistical weights in responding to sudden changes in the meteorological system. This allows the composite credibility weight to reflect the actual credibility of the meteorological data source at the current moment more promptly, thus making the fused meteorological input vector obtained based on this weight more closely resemble the actual meteorological state of the wind farm at the current moment. At the same time, the inter-source deviation is used to screen candidate abnormal meteorological data sources, providing a clear judgment object for further identification of the credibility of the meteorological data source, avoiding the uniform processing of all meteorological data sources indiscriminately.

[0011] Based on this, a weighted adjacency matrix for wind farm clusters is constructed using the geographical distance attenuation coefficient and historical power correlation coefficient. A set of associated wind farms is obtained by filtering wind farm nodes whose edge weights in the adjacency matrix reach a preset threshold. This establishes a data association structure for cross-validation among closely meteorologically correlated stations within the wind farm cluster, changing the existing centralized forecasting system's approach where each wind farm independently judges the credibility of meteorological data sources, and observation results between stations are isolated. By statistically analyzing the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction, a support rate is obtained. Based on this support rate, the composite credibility weights corresponding to candidate abnormal meteorological data sources are amplified or attenuated. This ensures that the correction result of the composite credibility weights no longer relies solely on the judgment of a single station, but introduces spatial verification from observation results of adjacent stations within the region. This reduces the likelihood of effective meteorological information being filtered out due to misjudging local real meteorological anomalies as observation noise. Yes, the corrected fused meteorological input vector is obtained by re-weighting the corrected composite credibility weights. This vector retains effective information about local meteorological anomalies while also taking into account the spatial consistency of regional meteorological observations. Finally, the corrected fused meteorological input vector is input into a graph convolutional spatiotemporal prediction network for processing to obtain the power prediction results of centralized wind farm clusters. The graph convolution and temporal feature extraction processing applied by this network further extracts the spatial correlation features and temporal evolution features of wind farm clusters based on the fused meteorological data that has already undergone credibility identification and spatial correlation correction in the aforementioned steps. This allows the algorithm to extract spatial features from the combined features of the fused meteorological data and unit operation data after local anomaly identification and correction, rather than the unidentified raw multi-source data, in the specific application scenario of regional centralized wind farm cluster power prediction. This reflects the inherent correlation between the algorithm structure and the data fusion correction logic in this specific application scenario, rather than a simple application of a general graph convolutional spatiotemporal network. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an embodiment of the centralized wind farm power prediction method based on multi-source data fusion in this application. Figure 2 This is a schematic diagram illustrating the historical reliability scores of various meteorological data sources under different operating condition labels in the embodiments of this application; Figure 3This is a schematic diagram comparing the power prediction results of a centralized wind farm group in the embodiments of this application. Detailed Implementation

[0014] This application provides a centralized wind farm power prediction method and system based on multi-source data fusion. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the centralized wind farm power prediction method based on multi-source data fusion in this application includes: Step S1: Clean and align the multi-source meteorological data with spatiotemporal references to obtain a spatiotemporally aligned meteorological dataset; Specifically, the spatiotemporal aligned meteorological dataset addresses the inconsistencies in temporal resolution, spatial resolution, and observation coordinate system among multi-source meteorological data. Numerical weather forecast data exists in grid form, while wind tower and generator data exist in point-time sequence form. Without unified coordinate mapping and time resampling, subsequent steps cannot compare or weight multiple data sources at the same time and location. The "cleaning" step refers to removing abnormal records that exceed the physical range or deviate too far from the historical distribution. The "spatiotemporal benchmark alignment" step maps grid data to station coordinate points and unifies them into sequences with the same time interval. This processing result constitutes the original input for subsequent steps involving weighted fusion.

[0016] Step S2: Based on the spatiotemporally aligned meteorological dataset, perform weighted fusion processing on the historical errors and inter-source deviation of each meteorological data source to obtain the composite confidence weight and fused meteorological input vector, and obtain candidate abnormal meteorological data sources by filtering the inter-source deviation. Specifically, the composite credibility weight and the static historical error weight corresponding to the fused meteorological input vector lag behind meteorological system mutations and the problem that a single verification dimension cannot identify meteorological source inaccuracies in a timely manner. Historical error reflects the long-term performance of meteorological sources under similar operating conditions in the past, while inter-source deviation reflects the horizontal differences between meteorological sources at the current moment. The two respectively cover two non-overlapping verification paths: post-event statistics and real-time horizontal comparison. The composite credibility weight obtained after combination has stronger timeliness than the weight that only relies on a single path. Candidate abnormal meteorological data sources refer to meteorological data sources that have obvious deviations from other meteorological sources in the inter-source deviation matrix and are marked as needing further determination of authenticity. Their screening results are directly used as the judgment objects for subsequent steps.

[0017] Step S3: Construct a weighted adjacency matrix for wind farm clusters based on the geographical distance attenuation coefficient and historical power correlation coefficient of wind farms. Select wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold to obtain the associated wind farm set corresponding to the candidate abnormal meteorological data source. Statistically process the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate. When the support rate reaches a preset support rate threshold, amplify the composite credibility weight corresponding to the candidate abnormal meteorological data source and use the highest historical credibility score as the upper limit. When the support rate does not reach the preset support rate threshold, attenuate the composite credibility weight. Re-weight the fused meteorological input vector with the corrected composite credibility weight to obtain the corrected fused meteorological input vector. Specifically, the weighted adjacency matrix of wind farm clusters and the support rate correspond to the majority consensus criterion, which may misjudge real local meteorological anomalies as observation noise, thus reducing the weight of effective signals. The adjacency matrix describes the degree of meteorological correlation between wind farms in the region based on geographical location and historical power performance. The support rate is the number of adjacent stations that have also observed anomalies in the same direction within the correlation range. It is used to distinguish whether a candidate anomaly meteorological data source is a local real meteorological event or an isolated observation error. The judgment result determines whether to amplify or attenuate the composite confidence weight, thereby obtaining a corrected fused meteorological input vector. This vector corrects the signal false rejection that may be caused by the simple majority consensus criterion.

[0018] Step S4: Process the corrected and fused meteorological input vector into the convolutional spatiotemporal prediction network to obtain the power prediction results of the centralized wind farm group.

[0019] Specifically, the corrected fusion of meteorological input vectors and centralized wind farm power prediction results addresses the problem of "data from adjacent wind farms not being able to converge under the one-farm-one-measurement model, and the model being unable to utilize the spatiotemporal dependencies between wind farms." The graph convolutional spatiotemporal prediction network's input includes corrected meteorological data from each wind farm and a weighted adjacency matrix between wind farms. The graph convolutional layer processes the spatial relationship structure between wind farms, while the time series processing layer processes the feature evolution of the same wind farm at different times. After the two work together, the output layer maps the power sequence of multiple wind farms in the region, achieving unified joint prediction of all wind farms in the region rather than independent prediction for each wind farm.

[0020] In one specific embodiment, step S1 includes: The raw meteorological data is processed by physical range verification and outlier removal based on the three-standard-deviation criterion to obtain cleaned meteorological data. Linear interpolation is performed on data segments in the cleaned meteorological data whose sampling interval is less than the preset missing duration threshold, and data segments in the cleaned meteorological data whose sampling interval is greater than the preset missing duration threshold are removed to obtain complete meteorological data. Based on the geographical coordinates of the wind farm, bilinear interpolation is performed on the grid data in the complete meteorological data to obtain the meteorological data of the wind farm location. Meteorological data of wind farm locations are resampled according to a unified time resolution to obtain a spatiotemporally aligned meteorological dataset.

[0021] Specifically, the wind speed determination range in the raw meteorological data is set to 0 m / s to 50 m / s. Records exceeding this range are judged as sensor malfunctions or transmission errors and are discarded. The wind direction determination range is set to 0 degrees to 360 degrees. Records exceeding this range are also discarded. The temperature determination range is set to -40 degrees Celsius to 50 degrees Celsius. The air pressure determination range is set to 850 hPa to 1050 hPa. The humidity determination range is set to 0 to 100. All of the above ranges are determined according to the range calibration range of anemometers, wind vanes, and temperature, humidity and pressure sensors in the meteorological observation industry. Values ​​exceeding the range are physically impossible and are directly treated as invalid records and not included in subsequent calculations. The three-standard-deviation criterion uses 96 samples from the past 24 hours, with one sampling point every 15 minutes, as a statistical window. It calculates the mean and standard deviation of the same observation from the same meteorological data source within the window. The absolute value of the difference between the current sampling point value and the mean is compared with three times the standard deviation. Sampling points with a difference exceeding three times the standard deviation are identified as outliers and removed. The statistical window is set at 24 hours to cover a complete day-night meteorological change cycle, avoiding the misjudgment of normal diurnal temperature differences and daily wind speed variations as outliers due to an excessively short window.

[0022] For missing data segments appearing in the cleaned data, linear interpolation is used for missing segments with a duration of less than 30 minutes. The linear interpolation takes the values ​​of the two valid sampling points before and after the missing period as endpoints. The estimated value of the missing point is obtained by linearly interpolating the values ​​of the two endpoints according to the proportion of the time interval between each sampling point and the previous valid point within the missing period to the total time interval between the two valid points. Data segments with a missing duration of 30 minutes or more are not interpolated and are directly marked as invalid and removed from the calculation in subsequent steps. The 30-minute threshold is set based on the fact that meteorological quantities such as wind speed and wind direction usually do not undergo drastic changes within 30 minutes, and the deviation between the linear interpolation result and the true value is within an acceptable range. After 30 minutes, the meteorological system may have changed, and linear interpolation no longer has reliability. Bilinear interpolation is applied to the grid data of numerical weather prediction. Using the geographical coordinates of the wind farm as the target location, four nearest grid points are found around this coordinate point. The distance ratio of the target coordinate point relative to these four grid points in both the longitude and latitude directions is calculated. First, the values ​​of adjacent sets of grid points are weighted and summed along the longitude direction to obtain two intermediate interpolation results. Then, these two intermediate interpolation results are weighted and summed along the latitude direction to obtain the meteorological value at the target coordinate point. The weights are proportional to the distance ratio, with grid points closer to the target point having a larger weight. All data sources are uniformly resampled to a 15-minute time resolution. During resampling, the arithmetic mean of the existing sampled points within the 15-minute window is used to obtain the representative value of the window. The selection of the 15-minute resolution is consistent with the regular upload cycle of the wind farm unit SCADA data, ensuring that the meteorological data and the unit operation data are aligned at the same timestamp. This provides a unified time reference for subsequent steps of calculating the deviation between different data sources and weighted fusion.

[0023] In one specific embodiment, step S2 includes: Based on the wind speed range, wind direction standard deviation, and the ratio of wind speed standard deviation to mean in the spatiotemporally aligned meteorological dataset, operating condition labels are constructed. Historical meteorological data are grouped and statistically processed according to operating condition labels to obtain the historical reliability score of each meteorological data source under each operating condition label. The meteorological data from each meteorological data source at the same time are processed by pairwise difference calculation to obtain the inter-source deviation matrix. The meteorological data source is normalized according to the inter-source deviation matrix to obtain the real-time consistency reliability score. Candidate abnormal meteorological data sources are obtained by screening the meteorological data sources whose deviation reaches the preset deviation threshold in the inter-source deviation matrix. Based on the working condition label corresponding to the current moment, the historical credibility weight is retrieved from the historical credibility score. The historical credibility weight is then weighted and summed with the real-time consistency credibility score to obtain the composite credibility weight. Based on the composite credibility weight, the meteorological data from each meteorological data source in the spatiotemporally aligned meteorological dataset are weighted and fused to obtain the fused meteorological input vector.

[0024] Specifically, the operating condition label consists of three indicators. Wind speed is divided into four intervals: 0 to 3 m / s, 3 to 8 m / s, 8 to 12 m / s, and 12 to 25 m / s. The boundaries of these four intervals correspond to the differences in operating characteristics of the wind turbine at the cut-in wind speed, the climb phase, near the rated wind speed, and before the cut-out wind speed. The impact mechanism of wind speed fluctuations on power output differs in different intervals, and the forecast error of meteorological sources also shows systematic differences in different intervals. The standard deviation of wind direction is calculated based on the wind direction observation values ​​at a resolution of 15 minutes in the past hour. A standard deviation less than 15 degrees is classified as stable, and greater than or equal to 15 degrees is classified as unstable. The 15-degree value is based on the fact that in conventional meteorological observations, wind direction swings exceeding this range can be considered as a significant deflection of the dominant wind direction. Turbulence intensity is calculated as the ratio of the standard deviation of wind speed to the mean in the past 10 minutes. A ratio less than 0.1 is classified as low turbulence, and greater than or equal to 0.1 is classified as high turbulence. The 0.1 value refers to the conventional boundary for medium and low turbulence sites in the wind power industry's classification of turbulence intensity levels. The three indicators are combined to form 16 operating condition categories. For the past 30 natural days, a total of 2880 historical samples at 15-minute resolution, statistics are collected for each meteorological data source according to its operating condition category. The percentage of the absolute error between the forecast wind speed and the measured wind speed under the meteorological source in the operating condition category is calculated. The average absolute percentage error is obtained by summing the percentages of all samples under the operating condition category and dividing by the number of samples. The reciprocal of the error is then divided by the sum of the reciprocals of the errors of all meteorological data sources in the same operating condition category for normalization. The historical reliability score of the meteorological source under the operating condition category is obtained. The smaller the error, the higher the score of the meteorological source. The sum of the scores of all meteorological sources in the same operating condition category is 1.

[0025] The calculation of the inter-source deviation matrix involves taking the wind speed values ​​output by each of the M meteorological data sources at the current moment, calculating the absolute value of the difference between any two wind speed values, forming an M-by-M dimensional matrix. The element in the j-th row and k-th column of the matrix represents the absolute value of the wind speed difference between the j-th and k-th meteorological sources at the current moment, with the diagonal elements being 0. For the j-th meteorological source, the arithmetic mean of the absolute values ​​of its deviation from the remaining M-1 meteorological sources is calculated to obtain the average deviation of that meteorological source at the current moment. The smaller the average deviation, the closer the output of that meteorological source is to the output of the majority of other meteorological sources. The average deviation is then negative and transformed by an exponential function. After transformation, meteorological sources with smaller deviations receive larger values. The transformed values ​​of the M meteorological sources are then normalized by dividing each by its sum to obtain the real-time consistency reliability score for each meteorological source at the current moment. When the average deviation of a meteorological source exceeds a preset threshold, it is determined to be a candidate abnormal meteorological data source. This threshold is set based on the deviation level outside the normal fluctuation range of the meteorological source in historical data and is used to distinguish between normal observation fluctuations and significant deviations. The historical reliability weight is derived from the historical reliability score pre-calculated under the corresponding operating condition category at the current moment, while the real-time consistency reliability score is the result calculated in real time for the current moment. The two respectively cover two independent dimensions: long-term statistical regularity and instantaneous horizontal comparison. The historical reliability weight is multiplied by 0.6 and the real-time consistency reliability score is multiplied by 0.4 and then added to obtain the composite reliability weight. The coefficients of 0.6 and 0.4 are set based on the fact that historical reliability reflects the long-term stable performance of meteorological sources and has a higher statistical confidence, and should be used as the main basis for weight calculation. Real-time consistency reliability is used to capture anomalies at the current moment that are not reflected in time by historical statistics, and serves as an auxiliary correction basis. The two scores are added together and then the results of M meteorological sources are normalized to obtain the final composite reliability weight. The wind speed, wind direction, temperature, air pressure, and humidity values ​​of each meteorological data source at the current moment are weighted and summed. Among them, the wind direction data is first decomposed into cosine and sine components, then weighted and summed respectively, and finally restored to the angle value through the arctangent function to obtain the fused meteorological input vector of the wind farm at the current moment.

[0026] Figure 2 This diagram illustrates the historical reliability scores of various meteorological data sources under different operating condition labels in this application embodiment. The horizontal axis represents the four operating condition labels divided by wind speed range: low wind speed, climbing section, near rated wind speed, and high wind speed. The vertical axis represents the historical reliability score. Different fill patterns are used in the diagram to distinguish the historical reliability scores of the four types of meteorological data sources—numerical weather forecast model 1, numerical weather forecast model 2, radar inversion, and satellite inversion—obtained through grouping and statistical analysis under each operating condition label. This is used to reflect the differences in historical reliability scores of the same meteorological data source under different operating conditions, as well as the distribution of relatively high and low historical reliability scores of different meteorological data sources under the same operating condition.

[0027] In one specific embodiment, in step S3, a weighted adjacency matrix of wind farm groups is constructed based on the wind farm geographical distance attenuation coefficient and historical power correlation coefficient. The associated wind farm set corresponding to candidate abnormal meteorological data sources is obtained by filtering wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold. This includes: Calculate the geographical distance between any two wind farms, substitute the geographical distance into the negative exponential decay function, and obtain the geographical distance decay coefficient. Based on the historical actual power sequence of each wind farm, the Pearson correlation coefficient is calculated for any two wind farms to obtain the historical power correlation coefficient. The geographical distance attenuation coefficient and the historical power correlation coefficient are multiplied to obtain the edge weights between wind farm nodes, and the weighted adjacency matrix of the wind farm group is constructed from the edge weights. The set of associated wind farms corresponding to candidate abnormal meteorological data sources is obtained by filtering wind farm nodes whose edge weights reach a preset edge weight threshold in the weighted adjacency matrix of wind farm clusters.

[0028] Specifically, the geographical distance between any two wind farms is calculated using the spherical distance formula based on the coordinates of the geographical centers of the two wind farms. The calculation result is substituted into a negative exponential decay function, which takes the geographical distance as the independent variable and sets the decay constant to 20 kilometers. That is, the farther the distance, the closer the function output value is to 0, and the closer the distance, the closer the function output value is to 1, thus obtaining the geographical distance decay coefficient. The decay constant of 20 kilometers is set based on the typical influence radius of the regional meteorological system. Wind farms within 20 kilometers of the same weather system are usually affected by similar meteorological processes. Beyond this range, the meteorological correlation weakens significantly. The historical power correlation coefficient is calculated based on the historical actual power series of 2880 sampling points at a 15-minute resolution over the past 30 natural days of the two wind farms. The calculation process of the Pearson correlation coefficient is as follows: calculate the mean of each of the two power series, multiply the difference between the power value of each sampling point and the mean of the series point by point between the two stations and sum them to obtain the covariance term, then sum the squares of the sums of the differences between the two series and the mean and multiply them to obtain the standard deviation product term. The covariance term divided by the standard deviation product term is the historical power correlation coefficient between the two wind farms. The value of this coefficient ranges from -1 to +1. The closer the value is to +1, the more consistent the historical power change trend of the two stations is.

[0029] The edge weights between wind farm nodes are obtained by multiplying the geographical distance attenuation coefficient with the historical power correlation coefficient. The edge weights reflect both the spatial proximity and the consistency of historical power performance. The edge weights of wind farm pairs that are only geographically close but have inconsistent historical power change trends, or wind farm pairs that are only related in historical power but are geographically far apart, will be reduced to avoid miscorrelation caused by single-dimensional judgment. The edge weights between each pair of N wind farms are filled into an N×N matrix, and the diagonal elements of the matrix are set to 0, thus forming the weighted adjacency matrix of the wind farm group. The preset edge weight threshold is set to 0.3. No connection edge is established for station pairs with edge weights less than 0.3. This threshold is set based on the historical distribution of the strength of meteorological correlation between stations in the region. Station pairs with weights below this value have a weak correlation in terms of geographical or power performance and do not have reference value for subsequent support rate determination. For a candidate abnormal meteorological data source in a certain wind farm, all wind farm nodes in the corresponding row of the wind farm with edge weights reaching or exceeding the preset threshold are retrieved in the weighted adjacency matrix. The retrieved set of wind farm nodes is used as the set of associated wind farms corresponding to the candidate abnormal meteorological data source. The stations in this set are used in subsequent determination to cross-validate whether the candidate abnormal meteorological data source reflects the real local meteorological anomaly.

[0030] In one specific embodiment, step S3 involves statistically processing the deviation compliance status of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate, including: For each wind farm in the associated wind farm set, retrieve whether the inter-source deviation of the corresponding wind farm at the current time and the candidate abnormal meteorological data source reaches the preset deviation threshold to obtain the wind farms that meet the deviation standard. Determine whether the deviation direction of the meteorological data source is consistent between the wind farm that deviates from the standard and the wind farm that generates candidate abnormal meteorological data source, and obtain the wind farm with consistent deviation direction. The number of wind farms that deviate from the same direction is statistically processed to obtain the number of support farms; The support rate is obtained by calculating the ratio of the number of supported wind farms to the total number of wind farms in the associated wind farm set.

[0031] Specifically, each wind farm in the associated wind farm set calculates its own source deviation matrix at the current time according to the method described in step S2. For the average deviation value corresponding to the candidate abnormal meteorological data source in the wind farm that generated the candidate label, the average deviation value corresponding to the meteorological data source is retrieved in each station of the associated wind farm set. If the value also reaches the preset deviation threshold of 0.15, the wind farm is marked as a wind farm with acceptable deviation. This threshold is consistent with the threshold used in step S2 to screen candidate abnormal meteorological data sources, and is used to determine whether the same meteorological data source also shows the characteristic of deviating from most meteorological sources in the associated stations. The deviation direction is determined by comparing the current wind speed value of the meteorological data source with the average wind speed of other meteorological data sources. If the wind speed value of the meteorological data source is higher than the average wind speed of other meteorological data sources, it is judged as a positive deviation; if it is lower, it is judged as a negative deviation. The deviation direction is determined separately between the wind farm that generated the candidate abnormal meteorological data source and the wind farm that deviated to the standard. If the two deviation directions are the same, the wind farm that deviated to the standard is marked as a wind farm with the same deviation direction. Wind farms that only meet the deviation standard but have opposite deviation directions are not included, because opposite deviation directions indicate that the abnormal performance of the meteorological data sources of the two stations does not originate from the same meteorological system and cannot be used as a basis for cross-validation.

[0032] The number of wind farms in the associated wind farm set that deviate in the same direction is counted to obtain the support count. This count reflects how many stations within the range of stations closely associated with the station that generated the candidate anomalous meteorological data source have observations that synchronously support the anomaly judgment. The support count is divided by the total number of wind farms in the associated wind farm set to obtain the support rate. This total number is the number of stations included in the associated wind farm set obtained in step S3 based on the edge weight threshold. The support rate ranges from 0 to 1. A higher value indicates more sufficient consistency verification of the candidate anomalous meteorological data source within the spatial association range, while a lower value indicates that the anomaly only occurs in isolation at a single station. The calculated support rate serves as the direct basis for subsequent amplification or attenuation processing of the composite confidence weight.

[0033] In one specific embodiment, in step S3, when the support rate reaches a preset support rate threshold, the composite confidence weight corresponding to the candidate abnormal meteorological data source is amplified and the highest historical confidence score is used as the upper limit. When the support rate does not reach the preset support rate threshold, the composite confidence weight is attenuated. The fused meteorological input vector is then reweighted using the corrected composite confidence weight to obtain the corrected fused meteorological input vector, including: When the support rate reaches the preset support rate threshold, the composite credibility weight corresponding to the candidate abnormal meteorological data source is multiplied with the preset amplification coefficient to obtain the amplification weight. The amplification weight is compared with the highest historical credibility score, and the smaller value is determined as the corrected composite credibility weight. When the support rate does not reach the preset support rate threshold, the composite confidence weight corresponding to the candidate abnormal meteorological data source is multiplied with the preset attenuation coefficient to obtain the corrected composite confidence weight. The corrected composite confidence weight is obtained by replacing the corresponding meteorological data source weight in the composite confidence weight with the corrected composite confidence weight. Based on the corrected composite confidence weights, the meteorological data from the corresponding meteorological data sources in the fused meteorological input vector are re-weighted and summed to obtain the corrected fused meteorological input vector.

[0034] Specifically, the preset support rate threshold is set to 0.6. When the support rate reaches or exceeds 0.6, it indicates that more than half of the stations in the associated wind farm set have synchronously observed the same meteorological data source deviating in the same direction. It can be determined that the candidate abnormal meteorological data source reflects the local real meteorological anomalies of multiple stations in the coverage area rather than isolated observation noise. The value of 0.6, which is higher than half, is to avoid the situation where only a few stations coincidentally show the same deviation and are mistakenly judged as real anomalies. When the support rate reaches 0.6, the composite confidence weight of the candidate abnormal meteorological data source in the wind farm that generated the label is multiplied by the preset amplification factor of 1.5 to obtain the amplification weight. The value of 1.5 increases the weight of the meteorological data source by 50% on the original basis, which is used to reflect that its observation results should obtain a higher degree of confidence after obtaining spatial cross-validation. At the same time, the largest value is taken from the historical confidence scores calculated by the meteorological data source under the 16 operating condition labels in step S2, which is taken as the historical highest confidence score. The amplification weight is compared with this value, and the smaller value is taken as the correction composite confidence weight. This upper limit setting is to avoid the amplification process causing the weight to exceed the best performance level that the meteorological data source has reached in history, and to prevent the correction result from deviating from the historical capability range of the meteorological data source itself.

[0035] When the support rate is below 0.6, the composite confidence weight corresponding to the candidate abnormal meteorological data source is multiplied by the preset attenuation coefficient of 0.5 to obtain the corrected composite confidence weight, that is, the original weight is reduced by half. The value of 0.5 is used to reflect that when the abnormal performance has not been supported by cross-validation of related stations and is more likely to be judged as observation noise or data quality problems, its role in the fusion process should be reduced. However, the weight is not directly set to zero because it is considered that a single failure to obtain support cannot completely rule out the possibility that the meteorological data source still has some reference value.

[0036] The corrected composite confidence weights calculated above are located in the original composite confidence weight vector according to the meteorological data source number. The original value at that position is replaced with the corrected composite confidence weight. The weights of the remaining meteorological data sources that are not judged as candidate anomalies remain unchanged from the original calculation results in step S2. After the replacement, the weights of the M meteorological data sources are normalized again so that the sum of the weights of all meteorological data sources is restored to 1, thus obtaining the corrected composite confidence weights. The corrected composite confidence weights are then recalculated with the original values ​​of wind speed, wind direction, temperature, air pressure, and humidity of the M meteorological data sources at the current time of the wind farm according to the weighted summation method described in step S2. The wind direction data is also first decomposed into cosine and sine components, then weighted and summed separately, and then restored by the arctangent function. This yields the corrected fused meteorological input vector of the wind farm at the current time after spatial correlation correction, which replaces the uncorrected fused meteorological input vector as the input data for the graph convolutional spatiotemporal prediction network in the subsequent steps.

[0037] In one specific embodiment, step S4 includes: The corrected and fused meteorological input vector is spliced ​​with the corresponding wind farm unit operation data to obtain the wind farm node feature vector. The feature vectors of wind farm nodes are arranged according to a preset number of time steps to obtain the temporal feature matrix of wind farm nodes; The temporal feature matrix of wind farm nodes and the weighted adjacency matrix of wind farm group are input into the graph convolutional layer of the graph convolutional spatiotemporal prediction network for spatial feature extraction to obtain the spatial features of wind farm. The spatial features of wind farm are then input into the gated recurrent unit layer of the graph convolutional spatiotemporal prediction network for temporal feature extraction to obtain the spatiotemporal features of wind farm. By mapping the fully connected output layer of the convolutional spatiotemporal prediction network that inputs spatiotemporal features of wind farms, the power prediction results of centralized wind farm clusters are obtained.

[0038] Specifically, the wind farm unit operation data includes four fields: active power, pitch angle, nacelle wind direction, and nacelle wind speed at the current moment. These four values ​​are then combined with the five values ​​of wind speed, wind direction, temperature, air pressure, and humidity from the corrected meteorological input vector produced in step S3, and concatenated sequentially to obtain a wind farm node feature vector with a dimension of 9. This concatenation method allows the fused meteorological data and the actual operating status of the unit to jointly represent the current state of the wind farm in the same feature vector. The preset number of time steps is 12, corresponding to the past 3 hours, with one time step every 15 minutes. The node feature vectors of each wind farm in the 12 consecutive time steps are arranged in chronological order to form a wind farm node time series feature matrix with a dimension of 12 by 9. The N wind farms correspond to their respective matrices. The value of 12 time steps is based on the fact that the change trend of wind power in the 3-hour scale is sufficient to reflect the impact of short-term meteorological system evolution on the power output of the station. If the time step is too short, it will be difficult to capture trend changes, and if it is too long, it will introduce historical information with low correlation to the current prediction time.

[0039] The node time-series feature matrices of each of the N wind farms, along with the weighted adjacency matrix of the wind farm group constructed in step S3, are input into the graph convolutional layer. The graph convolutional layer is calculated by symmetrically normalizing the weighted adjacency matrix to obtain a normalized adjacency matrix. This normalized adjacency matrix is ​​then multiplied by the node feature matrices of the N wind farms at the same time step, realizing the information transfer of adjacent wind farm features in the graph structure. The product result is then multiplied by the weight matrix of the graph convolutional layer and processed by an activation function. The output dimension is set to 64 spatial features. The 64-dimensional spatial features are calculated separately at each time step, so the output result is the 12 x 64-dimensional spatial features corresponding to each of the N wind farms. The spatial feature sequence integrates the station's own state with the state information of the associated stations in the weighted adjacency matrix. The 12×64-dimensional spatial feature sequence is sequentially input into the gated cyclic unit layer according to the time step order. The hidden layer dimension of the gated cyclic unit layer is set to 128. The hidden state is updated step by step by calculating the update gate and reset gate. The hidden state of the previous time step is combined with the spatial feature of the current time step to obtain the new hidden state of the time step. After processing 12 time steps in sequence, the 128-dimensional hidden state corresponding to the last time step is obtained, which serves as the spatiotemporal feature of the wind farm of the station. This feature contains both spatial association information and temporal evolution information.

[0040] The 128-dimensional spatiotemporal features of the wind farm are input into the fully connected output layer. The fully connected output layer maps the 128-dimensional features into 16 values ​​through the weight matrix and bias term, which correspond to the power prediction sequence for the next 4 hours at 15-minute intervals. The values ​​of the 16 output time steps, together with the aforementioned 12 input time steps, constitute the time span setting for predicting the power of the next 4 hours based on the state of the past 3 hours. The mapping result is output according to the wind farm number. N wind farms correspond to N sets of 16-dimensional power prediction sequences, which is the final power prediction result of the centralized wind farm group obtained by this method.

[0041] Figure 3 This is a schematic diagram comparing the power prediction results of a centralized wind farm group in an embodiment of this application. The horizontal axis represents the prediction time step, and the vertical axis represents the power value. The actual power curve is marked with a solid line, the predicted power curve of the existing static weight fusion method is marked with a dashed line, and the predicted power curve of the present invention is marked with a dotted line. This is used to show the deviation between the predicted power curve of the present invention and the predicted power curve of the existing static weight fusion method and the actual power curve.

[0042] The above describes the centralized wind farm power prediction method based on multi-source data fusion in the embodiments of this application. The following describes the centralized wind farm power prediction system based on multi-source data fusion in the embodiments of this application. One embodiment of the centralized wind farm power prediction system based on multi-source data fusion in the embodiments of this application includes: The alignment module is used to clean and align multi-source meteorological data with spatiotemporal references to obtain a spatiotemporally aligned meteorological dataset. The weighting module is used to perform weighted fusion processing on the historical errors and inter-source deviation of each meteorological data source according to the spatiotemporally aligned meteorological dataset, to obtain the composite confidence weight and fused meteorological input vector, and to obtain candidate abnormal meteorological data sources by filtering the inter-source deviation. The attenuation module is used to construct a weighted adjacency matrix of wind farm groups based on the attenuation coefficient of wind farm geographical distance and the correlation coefficient of historical power. It selects the associated wind farm set corresponding to the candidate abnormal meteorological data source by filtering the wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold. It performs statistical processing on the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate. When the support rate reaches a preset support rate threshold, it amplifies the composite credibility weight corresponding to the candidate abnormal meteorological data source and uses the highest historical credibility score as the upper limit. When the support rate does not reach the preset support threshold, it attenuates the composite credibility weight. The fused meteorological input vector is re-weighted by the corrected composite credibility weight to obtain the corrected fused meteorological input vector. The prediction module is used to process the corrected and fused meteorological input vector into the graph convolutional spatiotemporal prediction network to obtain the power prediction results of the centralized wind farm group.

[0043] This invention also provides a centralized wind farm power prediction device based on multi-source data fusion, which can be a server. The centralized wind farm power prediction device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the centralized wind farm power prediction device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the centralized wind farm power prediction device is used to store the data corresponding to this embodiment. The network interface of the centralized wind farm power prediction device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0044] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the centralized wind farm power prediction method based on multi-source data fusion.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0046] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part 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 centralized wind farm power prediction device based on multi-source data fusion (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 the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A centralized wind farm power prediction method based on multi-source data fusion, characterized in that, The method includes: Step S1: Clean and align the multi-source meteorological data with spatiotemporal references to obtain a spatiotemporally aligned meteorological dataset; Step S2: Based on the spatiotemporally aligned meteorological dataset, perform weighted fusion processing on the historical errors and inter-source deviation of each meteorological data source to obtain the composite confidence weight and fused meteorological input vector, and obtain candidate abnormal meteorological data sources by filtering the inter-source deviation. Step S3: Construct a weighted adjacency matrix for wind farm clusters based on the geographical distance attenuation coefficient and historical power correlation coefficient of wind farms. Filter wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold to obtain the associated wind farm set corresponding to the candidate abnormal meteorological data source. Statistically process the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate. When the support rate reaches a preset support rate threshold, amplify the composite credibility weight corresponding to the candidate abnormal meteorological data source, using the highest historical credibility score as the upper limit. When the support rate does not reach the preset support rate threshold, attenuate the composite credibility weight. Reweight the fused meteorological input vector with the corrected composite credibility weight to obtain the corrected fused meteorological input vector. Step S4: Process the corrected and fused meteorological input vector into the convolutional spatiotemporal prediction network to obtain the power prediction results of the centralized wind farm group.

2. The centralized wind farm power prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S1 includes: The raw meteorological data is processed by physical range verification and outlier removal based on the three-standard-deviation criterion to obtain cleaned meteorological data. Linear interpolation is performed on data segments in the cleaned meteorological data whose sampling interval missing duration is less than a preset missing duration threshold, and data segments in the cleaned meteorological data whose sampling interval missing duration reaches the preset missing duration threshold are removed to obtain complete meteorological data; Based on the geographical coordinates of the wind farm, bilinear interpolation is performed on the grid data in the complete meteorological data to obtain the meteorological data of the wind farm location; The meteorological data of the wind farm location is resampled according to a uniform time resolution to obtain the spatiotemporally aligned meteorological dataset.

3. The centralized wind farm power prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S2 includes: Based on the wind speed range, wind direction standard deviation, and the ratio of wind speed standard deviation to mean in the spatiotemporally aligned meteorological dataset, operating condition labels are constructed. Historical meteorological data are grouped and statistically processed according to the operating condition labels to obtain the historical reliability score of each meteorological data source under each operating condition label. The meteorological data from each meteorological data source at the same time are processed by pairwise difference calculation to obtain the inter-source deviation matrix. The meteorological data source is normalized according to the inter-source deviation matrix to obtain the real-time consistency reliability score. The candidate abnormal meteorological data source is obtained by screening the meteorological data source whose deviation reaches the preset deviation threshold in the inter-source deviation matrix. Based on the working condition label corresponding to the current moment, the historical credibility weight is retrieved from the historical credibility score, and the historical credibility weight is weighted and summed with the real-time consistency credibility score to obtain the composite credibility weight. The meteorological data from each meteorological data source in the spatiotemporally aligned meteorological dataset are weighted and fused based on the composite credibility weight to obtain the fused meteorological input vector.

4. The centralized wind farm power prediction method based on multi-source data fusion according to claim 1, characterized in that, In step S3, a weighted adjacency matrix of wind farm groups is constructed based on the geographical distance attenuation coefficient and historical power correlation coefficient of wind farms. The associated wind farm set corresponding to the candidate abnormal meteorological data source is obtained by filtering wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold. This set includes: Calculate the geographical distance between any two wind farms, substitute the geographical distance into the negative exponential decay function, and obtain the geographical distance decay coefficient. Based on the historical actual power sequence of each wind farm, the Pearson correlation coefficient is calculated for any two wind farms to obtain the historical power correlation coefficient. The geographical distance attenuation coefficient is multiplied by the historical power correlation coefficient to obtain the edge weights between wind farm nodes, and the weighted adjacency matrix of the wind farm group is constructed from the edge weights. The associated wind farm set corresponding to the candidate abnormal meteorological data source is obtained by filtering wind farm nodes whose edge weights reach a preset edge weight threshold in the weighted adjacency matrix of the wind farm group.

5. The centralized wind farm power prediction method based on multi-source data fusion according to claim 4, characterized in that, In step S3, the deviation compliance status of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction is statistically processed to obtain the support rate, including: For each wind farm in the associated wind farm set, retrieve whether the inter-source deviation of the corresponding wind farm at the current time and the candidate abnormal meteorological data source reaches a preset deviation threshold to obtain the wind farms that meet the deviation standard. Determine whether the deviation direction of the meteorological data source is consistent between the wind farm that deviated from the standard and the wind farm that generated the candidate abnormal meteorological data source, and obtain the wind farm with consistent deviation direction. The number of wind farms with the same deviation direction is statistically processed to obtain the number of support farms; The support rate is obtained by calculating the ratio of the number of supports to the total number of wind farms in the associated wind farm set.

6. The centralized wind farm power prediction method based on multi-source data fusion according to claim 5, characterized in that, In step S3, when the support rate reaches a preset support rate threshold, the composite confidence weight corresponding to the candidate abnormal meteorological data source is amplified and capped at the highest historical confidence score. When the support rate does not reach the preset support rate threshold, the composite confidence weight is attenuated. The fused meteorological input vector is then reweighted using the corrected composite confidence weight to obtain the corrected fused meteorological input vector, including: When the support rate reaches the preset support rate threshold, the composite credibility weight corresponding to the candidate abnormal meteorological data source is multiplied with the preset amplification coefficient to obtain the amplification weight. The amplification weight is compared with the highest historical credibility score, and the smaller value is determined as the corrected composite credibility weight. When the support rate does not reach the preset support rate threshold, the composite confidence weight corresponding to the candidate abnormal meteorological data source is multiplied with the preset attenuation coefficient to obtain the corrected composite confidence weight. The corrected composite confidence weight is obtained by replacing the weight value of the corresponding meteorological data source in the composite confidence weight with the corrected composite confidence weight. Based on the corrected composite confidence weights, the meteorological data from the corresponding meteorological data source in the fused meteorological input vector are re-weighted and summed to obtain the corrected fused meteorological input vector.

7. The centralized wind farm power prediction method based on multi-source data fusion according to claim 1, characterized in that, Step S4 includes: The corrected and fused meteorological input vector is spliced ​​with the corresponding wind farm unit operation data to obtain the wind farm node feature vector. The feature vectors of the wind farm nodes are arranged according to a preset number of time steps to obtain the temporal feature matrix of the wind farm nodes; The temporal feature matrix of the wind farm node and the weighted adjacency matrix of the wind farm group are input into the graph convolutional layer of the graph convolutional spatiotemporal prediction network for spatial feature extraction to obtain the wind farm spatial features. The wind farm spatial features are then input into the gated recurrent unit layer of the graph convolutional spatiotemporal prediction network for temporal feature extraction to obtain the wind farm spatiotemporal features. The spatiotemporal features of the wind farm are input into the fully connected output layer of the graph convolutional spatiotemporal prediction network for mapping processing to obtain the power prediction results of the centralized wind farm group.

8. A centralized wind farm power prediction system based on multi-source data fusion, characterized in that, For implementing the centralized wind farm power prediction method based on multi-source data fusion as described in any one of claims 1-7, the centralized wind farm power prediction system based on multi-source data fusion comprises: The alignment module is used to clean and align multi-source meteorological data with spatiotemporal references to obtain a spatiotemporally aligned meteorological dataset. The weighting module is used to perform weighted fusion processing on the historical errors and inter-source deviation of each meteorological data source according to the spatiotemporally aligned meteorological dataset, to obtain the composite confidence weight and fused meteorological input vector, and to obtain candidate abnormal meteorological data sources by filtering the inter-source deviation. The attenuation module is used to construct a weighted adjacency matrix of wind farm groups based on the attenuation coefficient of wind farm geographical distance and the correlation coefficient of historical power. It selects the associated wind farm set corresponding to the candidate abnormal meteorological data source by filtering the wind farm nodes in the weighted adjacency matrix whose edge weights reach a preset edge weight threshold. It performs statistical processing on the deviation compliance of each wind farm in the associated wind farm set under the same meteorological data source and the same deviation direction to obtain the support rate. When the support rate reaches a preset support rate threshold, it amplifies the composite credibility weight corresponding to the candidate abnormal meteorological data source and uses the highest historical credibility score as the upper limit. When the support rate does not reach the preset support threshold, it attenuates the composite credibility weight. The fused meteorological input vector is re-weighted by the corrected composite credibility weight to obtain the corrected fused meteorological input vector. The prediction module is used to process the corrected and fused meteorological input vector into the graph convolutional spatiotemporal prediction network to obtain the power prediction results of the centralized wind farm group.

9. A centralized wind farm power prediction device based on multi-source data fusion, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the centralized wind farm power prediction method based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the centralized wind farm power prediction method based on multi-source data fusion as described in any one of claims 1 to 7.

Images