A real-time quality control method, medium and system for offshore wind speed based on neural network

Abstract

The present invention provides a real-time quality control method, medium and system for offshore wind speed based on neural network, belonging to the technical field of offshore wind speed quality control. The real-time quality control method, medium and system for offshore wind speed based on neural network include performing range detection on offshore wind speed data, identifying abnormal peaks through spike detection, establishing a multi-element correlation model of wind speed and air pressure, performing adjacent station detection, establishing a quality control symbol statistical model, setting buoy stations as nodes in a spatio-temporal graph and constructing a nine-layer spatio-temporal graph convolutional neural network through natural connection edges and time connection edges, constructing a self-correlation time encoding layer including an adaptive delay conversion module and a self-correlation time convolution module for modeling the propagation delay and periodic characteristics of wind speed data, and outputting quality control warning information and sensor fault alarm information of wind speed data; the present invention can distinguish between weather changes and equipment failures, and improve the application universality of the system.

Description

Technical Field

[0001] The present invention belongs to the technical field of offshore wind speed quality control. Specifically, it relates to a real-time quality control method, medium and system for offshore wind speed based on neural network. Background Technique

[0002] Offshore wind speed monitoring is of great significance for marine weather forecasting, offshore wind energy development, offshore engineering construction and shipping safety. With the continuous expansion of the buoy observation network, data quality control has become a key link to ensure the reliability of observation data. At present, the quality control of offshore wind speed data mainly adopts statistical methods and empirical threshold judgment, including single-station extreme value test, adjacent point consistency test and time continuity test, etc. These traditional methods perform well in dealing with wind speed data under normal meteorological conditions, but have obvious deficiencies when facing the influence of complex weather systems.

[0003] In the prior art, the single-station extreme value test mainly relies on the threshold obtained by historical data statistics for judgment. This method judges the data too mechanically and cannot distinguish the sudden change of wind speed caused by weather processes and the abnormal data caused by equipment failures. For example, during a cold wave, the wind speed may increase significantly within a short period of time and exceed the statistical threshold, but this is a normal weather phenomenon and should not be judged as abnormal data. In the case of equipment failure, the measured wind speed value may happen to be within the threshold range but cannot reflect the true meteorological conditions.

[0004] The adjacent point consistency test method judges the data quality by comparing the observed values of adjacent stations. However, due to the significant non-uniformity of the spatial distribution of marine meteorological elements, the wind speed difference between adjacent stations may be due to the influence of local weather systems. The existing test methods are difficult to accurately distinguish this natural variation and observation anomalies, and are prone to misjudgment. Especially under the influence of weather systems such as typhoons and local convection, the reliability of the adjacent point consistency test is further reduced.

[0005] The time continuity test mainly focuses on the time variation characteristics of the wind speed at a single station. Traditional methods usually adopt a fixed change rate threshold, and this processing method is too simplistic and cannot adapt to the diversity of wind speed changes under different weather systems. For example, when a frontal system passes by, the wind speed may change rapidly, while under static and stable weather conditions, the wind speed changes relatively gently. The method of fixed threshold is difficult to take into account these two situations and often leads to over-rejection or omission of abnormal data.

[0006] In addition, the prior art generally has the problems of single criterion and lack of collaborative analysis of multiple factors. Wind speed changes are often closely related to other meteorological factors such as air pressure and temperature. Relying solely on wind speed data for quality control ignores the physical connections between meteorological factors and reduces the scientificity and accuracy of quality control. For example, during severe convective weather, a sudden increase in wind speed is usually accompanied by a sharp change in air pressure, and this collaborative relationship can be used as an important basis for judging data reliability. Summary of the Invention

[0007] In view of this, the present invention provides a real-time quality control method, medium and system for offshore wind speed based on a neural network, which can accurately distinguish between weather mutations and equipment failures, etc., and adapt to wind speed changes under different weather systems.

[0008] The present invention is implemented as follows:

[0009] In the first aspect of the present invention, a real-time quality control method for offshore wind speed based on a neural network is provided. A real-time quality control method for offshore wind speed based on a neural network includes the following steps: performing range detection on the input offshore wind speed data, identifying abnormal peaks through spike detection, establishing a multi-factor correlation model of wind speed and air pressure, performing adjacent station detection, establishing a quality control symbol statistical model, setting the buoy station as a node in the spatio-temporal graph and constructing a nine-layer spatio-temporal graph convolutional neural network through natural connection edges and time connection edges to extract the spatial correlation between stations and capture the temporal change characteristics, designing an adaptive graph convolution module, constructing a self-correlation time encoding layer including an adaptive delay conversion module and a self-correlation time convolution module for modeling the propagation delay and periodic characteristics of wind speed data, and outputting quality control warning information and sensor fault alarm information of the wind speed data.

[0010] Further, the range detection includes extreme value range detection, empirical range detection and equipment range detection. The quality level is determined by calculating the passing rate of wind speed data. The extreme value range detection threshold is 40 meters per second. The empirical range detection sets different thresholds according to seasons. The equipment range detection range is 0 to 60 meters per second. When the passing rate is greater than 70%, the quality control symbol is set to 1. When the passing rate is less than 70%, the quality control symbol is set to 3.

[0011] Further, the spike detection is performed by setting a sliding time window, calculating the average value of wind speed data at adjacent time points before and after as the peak reference value. When the difference between the wind speed value at the current moment and the peak reference value exceeds 3 times the standard deviation, the quality control symbol is set to 3, where the standard deviation is obtained by calculating the dispersion degree of wind speed data within the previous 12 hours.

[0012] Further, the wind speed and air pressure multi-factor correlation model calculates the synchronous observation values of wind speed data and air pressure data. When the change in wind speed exceeds 2.5 times the standard deviation of wind speed and the change in air pressure is less than 2.5 times the standard deviation of air pressure, the quality control flag is set to 3.

[0013] Further, for the adjacent station detection, buoy stations within a distance of 200 kilometers are selected as reference stations. When the change in wind speed at the adjacent moments of the station to be detected exceeds 2 times the first standard deviation while the change in wind speed at the adjacent moments of the reference station is less than 2 times the second standard deviation, the quality control flag is set to 3.

[0014] Further, the spatio-temporal graph convolutional neural network sets buoy stations as nodes in the spatio-temporal graph, represents the physical connection relationship between adjacent stations through natural connection edges, and represents the connection relationship of the same station at different time steps through time connection edges, constructing a 9-layer spatio-temporal graph convolutional structure.

[0015] Further, the adaptive graph convolution module represents the connection relationship between stations by constructing an initial adjacency matrix, introduces a learnable correlation matrix, uses the initial adjacency matrix for the first 10 iterations in the early stage of training, and then gradually introduces the influence of the correlation matrix.

[0016] Further, the autocorrelation time encoding layer includes an adaptive delay conversion module and an autocorrelation time convolution module, slices historical data through a 24-hour sliding window, uses the dynamic time warping algorithm for sequence alignment and clustering, and converts the signal to the frequency domain through Fourier transform to calculate the autocorrelation features at different time scales.

[0017] In the second aspect of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores program instructions, which are used to execute the above-mentioned real-time quality control method for offshore wind speed based on a neural network when running on a computer.

[0018] In the third aspect of the present invention, a real-time quality control system for offshore wind speed based on a neural network is provided, which includes the above-mentioned computer-readable storage medium. The system can be any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged inside the system, and a microprocessor for executing the program instructions stored in the computer-readable storage medium is arranged inside the system.

[0019] The present invention effectively improves the accuracy of quality control for offshore wind speed data by constructing a multi-level quality control system. First, an adaptive threshold system is adopted to dynamically adjust the judgment criteria according to seasonal characteristics and weather backgrounds, making the quality control more in line with meteorological laws. During strong weather processes such as cold snaps and typhoons, the system can automatically adjust the wind speed anomaly criterion according to the change of the pressure field, avoiding misjudging normal weather processes as data anomalies.

[0020] A multi-factor collaborative analysis mechanism is introduced to establish a correlation model between wind speed and pressure changes, providing a more reliable physical basis for the judgment of abnormal data. By analyzing the spatio-temporal consistency between wind speed changes and pressure changes, the system can accurately identify wind speed changes caused by weather systems and distinguish them from abnormal data caused by equipment failures. This judgment method based on physical processes significantly improves the scientific nature of quality control.

[0021] The spatio-temporal graph convolutional neural network model designed by the present invention can adaptively learn the spatial correlation between stations, overcoming the limitation of the fixed distance threshold in traditional adjacent point consistency tests. By capturing the dynamic correlation features between stations, the model can accurately identify the influence range of local weather systems and improve the accuracy of spatial consistency tests. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flowchart of the method of the present invention;

[0023] Figure 2 is the daily average wind speed curve graph in Embodiment 3;

[0024] Figure 3 is the scatter plot of the monthly average wind speed from 2021 to 2023 in Embodiment 3, where a is the effective value range corresponding to different extreme values in spring, b is the effective value range corresponding to different extreme values in summer, c is the effective value range corresponding to different extreme values in autumn, and d is the effective value range corresponding to different extreme values in winter;

[0025] Figure 4 is the monthly wind rose graph in Embodiment 3;

[0026] Figure 5 is the spatio-temporal graph of adjacent stations in Embodiment 4, where a is the spatio-temporal graph of the station and b is the schematic diagram of the station;

[0027] Figure 6 is the algorithm network structure diagram in Embodiment 4;

[0028] Figure 7 is the structure diagram of the dynamic delay conversion module in Embodiment 4;

[0029] Figure 8 is the structure diagram of the autocorrelation time convolution module in Embodiment 4

[0030] Figure 9 For the range check process in Embodiment 2;

[0031] Figure 10 For the real-time quality control workflow of offshore wind speed in Embodiment 2. Detailed implementation manners

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0033] As Figure 1 , Figure 10 shown, it is a flowchart of a real-time quality control method for offshore wind speed based on a neural network provided by the present invention. This method includes the following steps:

[0034] S10. Perform range detection on the input offshore wind speed data, calculate the passing rate of the wind speed data according to the extreme value range, empirical range, and equipment range, and determine the quality level of the wind speed data based on the passing rate;

[0035] S20. Perform spike detection on the wind speed data, and identify the single-value peak data in the wind speed data by calculating the standard deviation of the wind speed data at adjacent times;

[0036] S30. Establish a multi-element correlation model of wind speed and air pressure, calculate the absolute value of the change at adjacent times between the wind speed data and the air pressure data, and determine whether the wind speed data is affected by weather processes;

[0037] S40. Perform adjacent station detection, and identify wind speed anomalies caused by gusts or human factors by comparing the standard deviations of the data of the first wind speed sensor and the second wind speed sensor;

[0038] S50. Based on the range detection, the spike detection, the multi-element correlation model, and the adjacent station detection results, establish a quality control symbol statistical model and generate a quality control identifier for the wind speed data;

[0039] S60. Construct a spatio-temporal graph convolutional neural network, set the buoy station as a node in the spatio-temporal graph, and establish a multi-layer spatio-temporal graph convolutional structure based on natural connection edges and time connection edges;

[0040] S70. Design an adaptive graph convolutional module to obtain the long-distance spatial dependence relationship between stations by learning the adjacency matrix;

[0041] S80. Construct a self-correlation time encoding layer, including an adaptive delay conversion module and a self-correlation time convolutional module, for modeling the propagation delay and periodic characteristics of the wind speed data;

[0042] S90. Input the quality control identifier into the spatio-temporal graph convolutional neural network, and output the quality control warning information of the wind speed data and the sensor fault alarm information.

[0043] The following is a detailed description of the specific implementation of the above steps:

[0044] The specific implementation of step S10 is to first perform multi-level range detection on the received original offshore wind speed data, including three aspects: extreme value range detection, empirical range detection, and equipment range detection. In the extreme value range detection, based on meteorological knowledge and long-term observation experience, the wind speed data is compared with a preset threshold. When the wind speed value is greater than 40 meters per second or less than 0 meters per second, it is determined as abnormal data. In the empirical range detection, by analyzing the statistical characteristics of historical data, a seasonal empirical threshold model is established, and different wind speed thresholds are set for different seasons. For example, the threshold is set to 25 meters per second in spring, 30 meters per second in summer, 28 meters per second in autumn, and 35 meters per second in winter. In the equipment range detection, according to the factory technical specifications of the wind speed sensor, the range detection range is set to 0 to 60 meters per second. Subsequently, a weighted calculation is performed on the results of these three detections. The weight of the extreme value range detection is 0.4, the weight of the empirical range detection is 0.4, and the weight of the equipment range detection is 0.2. The passing rate of the wind speed data is calculated. When the passing rate is greater than 70%, the quality control symbol is set to 1, indicating that the data is normal; when the passing rate is less than 70%, the quality control symbol is set to 3, indicating that the data is suspicious. The main purpose of this step is to quickly identify significantly abnormal wind speed data through multi-dimensional range detection, providing a basis for subsequent refined quality control.

[0045] The specific implementation of step S20 is to implement spike detection through the sliding time window method to identify abnormal peaks in the wind speed data. First, set the sliding time window length to 3 time units. For the wind speed data at each time point, calculate the average value of the wind speed data at the two adjacent time points before and after it as the peak reference value. Then calculate the difference between the current wind speed value and the peak reference value, and compare it with 3 times the standard deviation. Among them, the standard deviation is obtained by calculating the dispersion degree of the wind speed data within the previous 12 hours. When the difference exceeds 3 times the standard deviation, the quality control symbol at this moment is set to 3, indicating that the data is suspicious. In practical applications, according to the wind speed change characteristics of different sea areas, the standard deviation multiple threshold can be dynamically adjusted, generally recommended to be between 2.5 and 3.5 times. The main function of this step is to identify short-term abnormal data caused by factors such as equipment failure, sudden weather, or human interference.

[0046] The specific implementation of step S30 is to construct a multi-factor correlation model of wind speed and air pressure. First, obtain the synchronous observation values of wind speed data and air pressure data, and calculate the change amounts at adjacent times. For wind speed data, calculate the absolute value of the difference in wind speed between the current time and the previous time; for air pressure data, also calculate the absolute value of the difference in air pressure at adjacent times. Calculate the standard deviations of the wind speed data series and the air pressure data series respectively, denoted as the wind speed standard deviation and the air pressure standard deviation. When the change amount of wind speed exceeds 2.5 times the wind speed standard deviation and the change amount of air pressure is less than 2.5 times the air pressure standard deviation, set the quality control flag at this time to 3, indicating that the data is suspicious. This judgment method is based on meteorological principles, that is, in normal weather processes, significant changes in wind speed are usually accompanied by corresponding changes in air pressure. The purpose of this step is to improve the recognition accuracy of wind speed changes caused by weather processes through multi-factor joint analysis.

[0047] The specific implementation of step S40 is to perform adjacent station detection. First, determine the adjacent reference stations for the station to be detected. Generally, select buoy stations within 200 kilometers as reference stations. Then obtain the wind speed observation data of the station to be detected and the reference stations in the same time period, and calculate the time length as 12 hours. Calculate the standard deviations of the wind speeds of the two stations respectively, denoted as the first standard deviation and the second standard deviation. When the change amount of the wind speed at adjacent times of the station to be detected exceeds 2 times the first standard deviation, while the change amount of the wind speed at adjacent times of the reference station is less than 2 times the second standard deviation, set the quality control flag at this time of the station to be detected to 3, indicating that the data is suspicious. The main purpose of this step is to use the observation data of adjacent stations for mutual verification to identify abnormal data caused by local gusts or human factors.

[0048] The specific implementation of step S50 is to establish a quality control flag statistical model. First, count the number of various quality control flags at each time point, including the quality control flags generated by range detection, spike detection, multi-factor detection, and adjacent station detection. Then, generate the final quality control label according to the combination of quality control flags. When all detections show normal, the label is 1; when only one detection shows suspicion, the label is 2; when two or more detections show suspicion, the label is 3; when any detection shows abnormality, the label is 4. For data with 5 or more abnormal labels accumulated within 3 consecutive hours, it is determined as equipment failure. The main purpose of this step is to form the final data quality assessment result through comprehensive analysis of various quality control results.

[0049] The specific implementation of step S60 is to construct a spatio-temporal graph convolutional neural network model. First, each buoy station is set as a node in the spatio-temporal graph, and the node features include information such as wind speed values and quality control identifiers. Then, based on the spatial relationships between stations, two types of edge connections are established: natural connection edges represent the physical connection relationships between adjacent stations, and time connection edges represent the connections of the same station at different time steps. On this basis, a 9-layer spatio-temporal graph convolutional structure is constructed, with each layer containing a graph convolution operation and a time convolution operation. The graph convolution operation is used to extract the spatial correlations between stations, and the time convolution operation is used to capture the temporal change features. The main purpose of this step is to establish a deep learning framework that can handle both spatial relationships and temporal evolution simultaneously.

[0050] The specific implementation of step S70 is to implement an adaptive graph convolution module. First, an initial adjacency matrix is constructed, and the matrix elements represent the connection relationships between stations. The value is 1 for connected stations and 0 otherwise. Then, a learnable correlation matrix is introduced, and the similarity between stations is calculated through a Gaussian kernel function. In the early stage of training, the initial adjacency matrix is used for the first 10 iterations, and then the influence of the correlation matrix is gradually introduced. The initial values of the elements of the correlation matrix are set to 1 and are continuously updated through the gradient descent method during the training process. The main purpose of this step is to achieve adaptive learning of the long-range spatial dependence relationships between stations.

[0051] The specific implementation of step S80 is to construct a self-correlation time encoding layer. In the adaptive delay conversion module, a sliding window with a length of 24 hours is used to slice the historical data, and the dynamic time warping algorithm is used for sequence alignment and clustering to extract typical wind speed change patterns. For the wind speed sequence at each moment, the similarity between it and the typical pattern is calculated through an attention mechanism and weighted fusion is performed. In the self-correlation time convolution module, first, a time convolutional network is used to extract time features, then the signal is transformed to the frequency domain through Fourier transform, and the self-correlation features at different time scales are calculated. The features are transformed back to the time domain through inverse Fourier transform, and the most significant time correlations are selected for feature fusion. The main purpose of this step is to effectively model the propagation delay effect and periodic characteristics of wind speed data.

[0052] The specific implementation of step S90 is to input the quality control identifier into the trained spatio-temporal graph convolutional neural network to obtain the quality control warning information and sensor fault alarm information of the wind speed data. For the warning information, according to the deviation degree between the predicted value and the measured value output by the model, the warning level is divided into three levels: when the prediction deviation is less than 1 meter per second, it is normal and no warning is required; when the prediction deviation is between 1 and 2 meters per second, it is a mild warning; when the prediction deviation is greater than 2 meters per second, it is a severe warning. For the fault alarm information, when there are 5 or more abnormal identifiers continuously accumulated within 3 hours at a certain station, the sensor fault alarm is triggered. The main purpose of this step is to realize the real-time monitoring and warning of the quality of wind speed data.

[0053] Through the implementation of the above steps, the method realizes the all-round quality control of the offshore wind speed data, can effectively identify and process various abnormal data, and ensure the reliability of the wind speed data.

[0054] In the present invention, the involved calculation process is described in detail as follows:

[0055] 1. The calculation equation of the passing rate of range detection is specifically expressed as follows:

[0056] ;

[0057] In the formula, is the passing rate of wind speed data; is the result of empirical range detection, taking values of 0 or 1; is the result of extreme value range detection, taking values of 0 or 1; is the result of equipment range detection, taking values of 0 or 1; is the weight coefficient and satisfies ;

[0058] 2. The calculation equation of spike detection is specifically expressed as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula, is the peak reference value; is the wind speed value at time t; is the standard deviation; is the time window length, taking the value of 12; is the current wind speed deviation value;

[0063] 3. The calculation equation of the multi-element correlation of wind speed and air pressure is specifically expressed as follows:

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] In the formula, is the wind speed change amount; is the air pressure change amount; is the standard deviation of wind speed change; is the standard deviation of air pressure change; is the time window length, with a value of 24;

[0069] 4. The calculation equation of the adaptive graph convolutional model is specifically expressed as follows:

[0070] ;

[0071] ;

[0072] In the formula, is the output feature; is the convolution kernel weight; is the input feature; is the adjacency matrix; is the learnable correlation matrix; is the distance between node i and node j; is the Gaussian kernel parameter;

[0073] 5. The calculation equation of the autocorrelation time encoding is specifically expressed as follows:

[0074] ;

[0075] ;

[0076] In the formula, is the spectral density function; is the time series is the Fourier transform of; represents the conjugate complex number; is the autocorrelation function; is the time delay term;

[0077] Description of the parameter acquisition method:

[0078] 1. The weight coefficients are obtained through statistical analysis of historical data, and the values are 0.4, 0.4, and 0.2 respectively;

[0079] 2. Obtained through calculation of the latitude and longitude coordinates of the buoy station;

[0080] 3. It is obtained by cross-validation method, and the value is generally taken as 1 / 3 of the minimum distance between stations;

[0081] 4. It is calculated by the dynamic time warping algorithm, and its calculation steps are:

[0082] Step 1: Calculate the distance matrix of two time series;

[0083] Step 2: Use dynamic programming algorithm to find the optimal alignment path;

[0084] Step 3: Calculate the time delay value according to the alignment path.

[0085] Explanation of the equation construction principle:

[0086] 1. The range detection pass rate adopts the weighted summation form, taking into account the importance of different detection methods;

[0087] 2. Peak detection uses the average value of adjacent points as a reference, which can effectively identify mutation points;

[0088] 3. The standard deviation comparison method is used for the correlation of multiple factors, which reflects the synergy between wind speed and air pressure changes;

[0089] 4. The adaptive graph convolution model uses a Gaussian kernel function to calculate spatial correlation, ensuring the physical property that correlation decreases with increasing distance;

[0090] 5. Autocorrelation time coding uses spectrum analysis method, which can effectively capture periodic characteristics.

[0091] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned real-time quality control method of offshore wind speed based on neural network.

[0092] A third aspect of the present invention provides a wind tunnel health monitoring data mining system, which includes the above-mentioned computer-readable storage medium, the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0093] Specifically, the principle of the present invention is: based on the basic principles of meteorology and deep learning technology, the present invention constructs a complete set of offshore wind speed data quality control methods. In principle, the effectiveness of this method is mainly reflected in the following aspects:

[0094] First, as a physical quantity describing the state of atmospheric motion, the change of wind speed must follow the basic laws of atmospheric motion. The present invention combines data quality control with atmospheric dynamics processes by establishing a multi-factor correlation model between wind speed and air pressure. When an abnormal wind speed occurs, the system will automatically check the corresponding characteristics of air pressure changes and judge the rationality of the data through the conservation relationship between physical quantities. This judgment method based on physical mechanisms essentially improves the scientific nature of quality control.

[0095] Second, the distribution of marine meteorological elements has obvious spatial correlation, and adjacent regions are often affected by the same weather system. The spatio-temporal graph convolutional structure designed in the present invention effectively describes the spatial relationship and temporal evolution characteristics between stations through natural connection edges and time connection edges. This network structure can automatically learn the spatial influence range of weather systems at different scales, making the quality control more in line with the actual meteorological process.

[0096] Finally, the autocorrelation time coding technology adopted in the present invention can effectively capture the periodic characteristics and propagation delay effects of wind speed data. Through Fourier transform and dynamic time warping algorithms, the system can identify typical patterns of wind speed changes and dynamically adjust the judgment criteria according to the characteristics of different weather systems. This adaptive mechanism makes the quality control method more universal.

[0097] A specific Embodiment 1 of the present invention is provided below. The specific implementation manners of each step in this Embodiment 1 are described in detail as follows:

[0098] The specific implementation manner of step S10 is to establish a multi-level range detection process to comprehensively evaluate the quality of marine wind speed data. First, an extreme value range detection threshold is established based on meteorological theory and long-term observation experience. When the wind speed measurement value is greater than 40 meters per second or less than 0 meters per second, it is determined as abnormal data. At the same time, combined with the statistical characteristics of historical data, a seasonal empirical threshold model is constructed. The threshold is set to 25 meters per second in spring, 30 meters per second in summer, 28 meters per second in autumn, and 35 meters per second in winter for differential detection according to the wind speed change characteristics in different seasons. According to the factory technical specifications of the wind speed sensor, the range detection range is set to 0 to 60 meters per second. The range detection pass rate is calculated using a weighted summation method, and its calculation formula is: , where is the wind speed data pass rate, is the empirical range detection result, is the extreme value range detection result, is the device range detection result, and the values of all three are 0 or 1. The weight coefficient Take the values 0.4, 0.4, and 0.2 respectively. When the calculated passing rate is greater than 70%, set the quality control flag to 1, indicating that the data is normal; when the passing rate is less than 70%, set the quality control flag to 3, indicating that the data is suspicious. This step quickly identifies significantly abnormal wind speed data through multi-dimensional range detection, providing a basis for subsequent refined quality control.

[0099] The specific implementation of step S20 is to use the sliding time window method for spike detection, mainly used to identify abnormal peaks in wind speed data. First, set the sliding time window length to 3 time units. For the wind speed data at each time point, calculate its peak reference value, and the calculation formula is: , where is the peak reference value, is the wind speed value at time t. Then calculate the standard deviation of the wind speed data within the previous 12 hours, and its calculation formula is: , where is the standard deviation, is the time window length. Calculate the wind speed deviation value at the current moment, and its calculation formula is: , where is the wind speed deviation value at the current moment. When exceeds , set the quality control flag at this moment to 3, indicating that the data is suspicious. This step can effectively identify short-term abnormal data caused by factors such as equipment failures, sudden weather, or human interference.

[0100] The specific implementation of step S30 is to construct a multi-factor correlation model of wind speed and air pressure. First, calculate the change amounts of adjacent moments of wind speed and air pressure, and their calculation formulas are: and , where is the wind speed change amount, is the air pressure change amount. Then calculate the standard deviations of the wind speed and air pressure changes, and their calculation formulas are: and , where is the standard deviation of wind speed change, is the standard deviation of air pressure change. When the wind speed change amount exceeds and the air pressure change amount is less than , set the quality control flag at this moment to 3, indicating that the data is suspicious. This step improves the recognition accuracy of wind speed changes caused by weather processes through multi-factor joint analysis.

[0101] The specific implementation of step S40 is to perform adjacent station detection. First, determine the adjacent reference stations of the station to be detected, and select buoy stations within a distance of 200 kilometers as reference stations. Calculate the standard deviation of the wind speed at the station to be detected and the reference stations within the same time period, with a time length of 12 hours. When the wind speed change amount at adjacent moments of the station to be detected exceeds 2 times its standard deviation, while the wind speed change amount at adjacent moments of the reference station is less than 2 times its standard deviation, set the quality control flag of the station to be detected at that moment to 3, indicating that the data is suspicious. This step uses the observed data of adjacent stations for mutual verification to identify abnormal data caused by local gusts or human factors.

[0102] The specific implementation of step S50 is to establish a quality control flag statistical model. First, count the number of quality control flags of each type at each time point, including the quality control flags generated by range detection, spike detection, multi-element detection, and adjacent station detection. When all detections show normal, it is marked as 1; when only one detection shows suspicion, it is marked as 2; when two or more detections show suspicion, it is marked as 3; when any detection shows abnormality, it is marked as 4. For data with 5 or more abnormal marks accumulated within 3 consecutive hours, it is determined as equipment failure. This step forms the final data quality assessment result through comprehensive analysis of various quality control results.

[0103] The specific implementation of step S60 is to construct a spatio-temporal graph convolutional neural network model. Set each buoy station as a node in the spatio-temporal graph, and the node features include information such as wind speed value and quality control flag. The calculation formula of the adaptive graph convolution model is: , where is the output feature, is the convolution kernel weight, is the input feature, is the adjacency matrix, is the learnable correlation matrix. The calculation formula of the spatial correlation between stations is: , where is the distance between node i and node j, is the Gaussian kernel parameter. This step establishes a deep learning framework that can handle spatial relationships and temporal evolution simultaneously.

[0104] The specific implementation of step S70 is to implement the adaptive graph convolution module. First, construct an initial adjacency matrix, where the matrix elements represent the connection relationship between stations. The value is 1 for connected stations and 0 otherwise. Introduce a learnable correlation matrix and calculate the similarity between stations through the Gaussian kernel function. Use the initial adjacency matrix for the first 10 iterations in the early stage of training, and then gradually introduce the influence of the correlation matrix. This step realizes the adaptive learning of long-distance spatial dependence relationships between stations.

[0105] The specific implementation of step S80 is to construct an autocorrelation time encoding layer. The autocorrelation time encoding calculation formula is: and , where is the spectral density function, is the time series is the Fourier transform of, represents the conjugate complex number, is the autocorrelation function, is the time delay term. A sliding window with a length of 24 hours is used to slice the historical data, and the dynamic time warping algorithm is used for sequence alignment and clustering to extract typical wind speed change patterns. This step effectively models the propagation delay effect and periodic characteristics of wind speed data.

[0106] The specific implementation of step S90 is to input the quality control identifier into the trained spatio-temporal graph convolutional neural network to obtain the quality control warning information and sensor fault alarm information of the wind speed data. The warning levels are divided into three levels: a prediction deviation less than 1 meter per second is normal and no warning is required; a prediction deviation between 1 and 2 meters per second is a mild warning; a prediction deviation greater than 2 meters per second is a severe warning. When there are 5 or more abnormal identifiers continuously for 3 hours at a certain station, a sensor fault alarm is triggered. This step realizes the real-time monitoring and warning of the quality of wind speed data.

[0107] To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: The present invention has been practically applied to 15 offshore buoy stations arranged in the Yellow Sea and Bohai Sea areas, and the research area involves important sea areas such as the Bohai Bay, Laizhou Bay, and Liaodong Bay. First, the classification standard of the quality control symbols used in the quality control of wind speed data is given in Table 1 as follows:

[0108] Table 1 Classification standard of quality control symbols for wind speed data quality control

[0109]

[0110] The specific distribution information of the 15 buoy stations is as shown in the following table:

[0111] Table 2 Specific distribution information of buoy stations

[0112]

[0113] The range detection determination method is as shown in the following table:

[0114] Table 3 Range detection determination method

[0115]

[0116] Among them, is the wind speed at the nth moment, and USER_MAX is the historical threshold obtained by combining the buoy historical experience with the cumulative probability and wind direction. is the set passing rate, as Figure 9 shown, satisfies the following conditions:

[0117] ;

[0118] Among them, JZ is the extreme value range, Wdir*JY is the empirical wind direction, PJZ is the average value, and LC is the range.

[0119] The peak detection judgment conditions are shown in the following table:

[0120] Table 4 Peak Detection Judgment Conditions

[0121]

[0122] Among them, SPK_REF is the average value of adjacent data points n-2 and n as the peak reference, SD is the standard deviation, and satisfies the following conditions:

[0123] ;

[0124] In the above formula, TIM_DEV is the period for calculating SD, is the wind speed at any moment, is the average wind speed within the period.

[0125] The multi-factor detection judgment conditions are shown in the following table:

[0126] Table 5 Multi-factor Detection Judgment Conditions

[0127]

[0128] is the absolute value of the change in adjacent moments of the average wind speed, is the absolute value of the change in air pressure at the same moment, SD_WS is the wind speed standard deviation, and SD_BP is the air pressure standard deviation.

[0129] Adjacent station detection can be used for short-term large fluctuations in wind speed caused by gusts, accidents, or human factors during non-weather processes. Compare the wind data of Station 1 (WS1) with the wind data of Station 2 (Ws2), and calculate the standard deviations (SD1, SD2) of each device during this period (TIM_DEV).

[0130] The adjacent station detection judgment conditions are shown in the following table:

[0131] Table 6 Adjacent Station Detection Judgment Conditions

[0132]

[0133] Among them, is the wind speed at the nth moment of Device 1, is the wind speed at the (n - 1)th moment of Device 1, is the wind speed at the nth moment of Device 2, is the wind speed at the (n - 1)th moment of Device 2; SD1 is the standard deviation of the wind speed of Device 1 during TIM_DEV, and SD2 is the standard deviation of the wind speed of Device 2 during TIM_DEV.

[0134] During the period from January 15th to 18th, 2024, the Yellow Sea and Bohai Sea areas were affected by strong cold air, resulting in a strong wind weather process. The present invention conducts quality control on the wind speed data during this period, and the specific implementation process is as follows:

[0135] First, range detection is carried out. Considering the characteristics of the winter wind speed threshold, empirical range thresholds for different sea areas are set: the Bohai Bay stations (BH1 - BH5) are 35 meters per second, the Laizhou Bay stations (LZ1 - LZ5) are 32 meters per second, and the Liaodong Bay stations (LD1 - LD5) are 38 meters per second; the extreme value range threshold is uniformly set to 40 meters per second, and the device range is from 0 to 60 meters per second. The specific implementation process takes the data of Station BH1 as an example:

[0136] The continuous observation data from 10:00 to 13:00 on January 15th are shown in the following table:

[0137] Table 7 Continuous Observation Data of Station BH1 from 10:00 to 13:00 on January 15th

[0138]

[0139] Among them, the passing rate calculation uses a weight system: the empirical range detection weight is 0.4, the extreme value range detection weight is 0.4, and the device range detection weight is 0.2. For example, the passing rate calculation at 10:00 is: , since the passing rate is less than 70%, the quality control symbol is set to 3.

[0140] Subsequently, spike detection is carried out. Taking the data of Station BH2 from 14:00 to 17:00 on January 15th as an example:

[0141] Table 8 Continuous Observation Data of Station BH2 from 14:00 to 17:00 on January 15th

[0142]

[0143] Among them, the analysis process at 16:00 is: calculate the peak reference value meters per second, calculate the standard deviation of the wind speed in the previous 12 hours meters per second, the wind speed deviation value at the current moment meters per second, less than m / s, so the quality control flag remains 1.

[0144] When performing multi-factor correlation detection, analyze the synergy between wind speed changes and air pressure changes. Taking the observation data of the BH3 station from 18:00 to 21:00 on January 15 as an example:

[0145] Table 9 Continuous Observation Data of the BH3 Station from 18:00 to 21:00 on January 15

[0146]

[0147] Calculate the standard deviation of wind speed changes within 24 hours m / s, and the standard deviation of air pressure changes hPa. Taking the moment of 18:00 as an example, the wind speed change is 6.7 m / s, exceeding m / s, and the air pressure change is 3.7 hPa, less than hPa, so the quality control flag is set to 3.

[0148] In the adjacent station detection, select the BH4 station and the BH3 station for analysis. The observation data from 20:00 to 23:00 on January 15 are as follows:

[0149] Table 10 Continuous Observation Data of the BH4 Station from 20:00 to 23:00 on January 15

[0150]

[0151] Taking the moment of 20:00 as an example, the wind speed change of the BH4 station m / s exceeds m / s, while the wind speed change of the BH3 station m / s is less than m / s, so the quality control flag of the BH4 station at this moment is set to 3.

[0152] In terms of quality control flag statistics and equipment failure judgment, analyze the data of the BH5 station from 2:00 to 5:00 on January 16:

[0153] Table 11 Continuous Observation Data of the BH5 Station from 2:00 to 5:00 on January 16

[0154]

[0155] Since the quality control flag is 3 for 10 consecutive times, it is determined that there is a fault in the equipment of the BH5 station, and the system automatically issues an equipment failure warning message.

[0156] The statistical results of the application of the method of the present invention at 15 buoy stations are as follows:

[0157] Table 12 Statistical Data of the Application Results of 15 Buoy Stations

[0158]

[0159] Through comparative analysis, the advantages of the present invention over traditional quality control methods are shown in the following table:

[0160] Table 13 Advantages of the Present Invention over Traditional Quality Control Methods

[0161]

[0162] The practical application of the present invention in the Yellow Sea and Bohai Sea areas shows that:

[0163] 1. Through a multi-level quality control process, abnormal data can be effectively identified, and the quality control accuracy rate reaches 96%;

[0164] 2. By adopting a multi-factor collaborative analysis method, the wind speed changes caused by weather processes and the abnormal data caused by equipment failures can be successfully distinguished;

[0165] 3. By introducing spatial correlation analysis, the ability to identify local abnormal weather is improved.

[0166] The following provides a specific Embodiment 3 of step S10 of the present invention:

[0167] The data standardization processing module continuously reads the database, automatically obtains the buoy wind data and its related data, and then standardizes the data to ensure that the data from different manufacturers have a unified format, which is convenient for subsequent operations. During the data quality control stage, the data will not be modified, and quality control symbols are marked for recording. The alarm processing module triggers different alarm prompts according to the quality control symbol records, including data anomaly alarms and equipment failure alarms.

[0168] In the visualization stage, different colors are set for different types of data, and data graphs are drawn in combination with the quality control symbol record table, visually presenting the abnormal data, which helps the auditor quickly grasp the overall situation and trend of the data. The auditor comprehensively analyzes various situations to complete the data review work.

[0169] Data standardization processing module: Obtain real-time data streams from multiple different data sources such as sensors from different manufacturers through advanced data acquisition tools, and map them into a unified data structure framework. Remove noise, redundant information, and incomplete data through data cleaning technology to ensure the consistency and comparability of the data, providing a standardized basis for subsequent data analysis.

[0170] In this step, by combining time series analysis methods and geographic information system technology, the timestamps and spatial location information in the data stream are identified and processed, and the Gregorian calendar and lunar calendar dates are recorded. Combining with subsequent detection methods, the combination method of the data stream is dynamically adjusted to generate a high-quality target real-time data stream, improving the quality control efficiency.

[0171] Threshold determination method based on cumulative frequency method: The selection of the threshold is crucial, which directly affects the accuracy of range detection. The wind threshold selection method proposed in this patent considers various influencing factors, including seasonal changes and extreme weather events, etc. The threshold obtained by this method can represent the normal wind speed in most cases. While excluding the influence of extreme weather, the main wind direction and occurrence probability of extreme wind are also given, making the range detection more accurate.

[0172] Wind characteristic analysis: By analyzing wind data, the characteristic distribution of wind in this area is found. In order to better understand the seasonal changes of the ocean, the four seasons of the ocean are adopted in the present invention. Taking Station A as an example, the daily average of wind speed data is analyzed to analyze the characteristic changes of wind speed at this station. As Figure 2 shown, the daily average wind speed is concentrated at 2 - 10 m / s. The annual average wind speed from 2021 to 2023 is 6.40 m / s, and the wind speed distribution trend is relatively consistent. The wind speed gradually decreases at the end of winter and the beginning of spring, and drops to the lowest in the whole year in summer. The wind speed gradually increases in autumn and reaches the maximum in the whole year at the turn of autumn and winter. Therefore, the change of wind speed at Station A has obvious seasonal characteristics, and a threshold can be set for each season.

[0173] Selection of threshold reference value: First, the maximum wind data of each month is statistically analyzed. Combining the results of scatter plot and seasonal analysis, a threshold that is neither too high nor too low is selected as the threshold reference value for each season.

[0174] The threshold reference value is obtained by using the method of theoretical analysis. Next, the cumulative frequency method can be used to verify the rationality of the reference value. First, set a screening rate P for range detection, representing that the probability that the data cannot pass the range detection is 1 - P. If the cumulative frequency corresponding to the reference value is basically consistent with 1 - P, then this reference value is reasonable. If the difference between the two is large, it is necessary to reselect the reference value and verify it again to obtain the most suitable threshold. The cumulative frequency can be set according to its own needs.

[0175] ;

[0176] Among them, is the wind speed threshold for season X, is the cumulative probability in season x over the years.

[0177] Next, set the minimum value of the seasonal threshold as Hseaso-min, use the wind rose diagram to count the wind directions when H > Hseaso-min, determine the main wind directions detected beyond the empirical range, and set the probability of the main wind directions as the weight W for the empirical range detection. dir Take Station A as an example to demonstrate the process of threshold selection.

[0178] (1) Selection of reference values. As shown in Table 1 and Figure 3 shown, set the spring threshold as 17 m / s, summer 14 m / s, autumn 20 m / s, and winter 16 m / s.

[0179] Table 14 Monthly maximum average wind speed statistical table from 2021 to 2023

[0180]

[0181] (2) Calculate the cumulative probability. Set P = 97%, then 1 - P = 3%, that is, it is more appropriate for Px to be around 3%. The winter threshold is appropriate, and the others need to be adjusted. Repeat this process until the appropriate threshold is obtained.

[0182] Table 15 Statistical table of seasonal threshold distribution probabilities in the past three years

[0183]

[0184] Assume Hseaso-min = 16, as Figure 4 shown, use the wind rose diagram to count the wind directions when H > 16, the main wind direction is northeast, and the occurrence probability = 97%,

[0185] Example 4: A specific Example 4 of steps S70 and S80 in the present invention is provided below. To address the above problems, the present invention constructs a new short- and medium-term wind speed prediction model, which can automatically capture the patterns embedded in the spatial structure of the observation stations and their temporal dynamics, and predict the wind speed values at multiple future time steps. As Figure 5 shown, when constructing the spatio-temporal graph, the buoy stations are regarded as the nodes of the graph, and the edges are divided into two types. One is the natural edge, that is, the natural connection between the nodes (orange); the other is the time edge (blue), which connects the same nodes across consecutive time steps, as Figure 5 shown. On this basis, a multi-layer spatio-temporal graph convolution is constructed to achieve the integration of information along the graph dimension and the time dimension.

[0186] A North Sea buoy dataset was constructed, on which the model was trained and tested. The mean absolute error and mean squared error were used to evaluate the model, and a comparison curve graph of predicted values and true values was provided. In addition, since the learnable adjacency matrix is the only factor determining the relationship between cities, the content learned by the network was visualized and some explanations were provided visually. The contributions of the present invention are as follows:

[0187] (1) A wind speed prediction model based on adaptive graph convolution and autocorrelation time encoding was proposed to predict the accurate wind speed in the North Sea area, solving the problems of regionality, periodicity, long distance, and time delay of complex wind data proposed above;

[0188] (2) An adaptive graph convolution module was proposed, which can obtain the long-distance spatial dependence relationship between stations; a self-correlation time encoding layer was designed, including two core components: the adaptive delay conversion module can dynamically model the time delay in the data propagation process, and the self-correlation time convolution module is used to learn the inherent periodicity in the wind data sequence;

[0189] (3) The method proposed in the present invention was experimented on the collected real dataset. The experimental results showed that the mean absolute error MAE for predicting the next 1 hour was 0.79, the mean squared error MSE was 1.08, and the Pearson coefficient between the predicted value and the actual value was 96.5%.

[0190] As Figure 6 shown, the network is composed of 9 AST-Blocks. Each AST-Block includes an adaptive graph convolution module and a self-correlation time encoding layer. The adaptive graph convolution is executed before the self-correlation time encoding layer, and a residual connection is added to the entire block. Each block contains its own, separate, learnable adjacency matrix in the spatial convolution.

[0191] The consistency of adjacent stations plays an important role in analyzing elements with poor internal consistency such as precipitation and wind. Researchers often use graph convolution to obtain this correlation. However, in reality, the distance between any two stations is hundreds of nautical miles, which means that the spatial dependence is long-distance. In addition, the GCN-based model has the problem of over-smoothing and is difficult to capture the long-distance spatial dependence. The present invention proposes an adaptive graph convolution module to obtain the long-distance spatial dependence relationship between stations.

[0192] ;

[0193] ;

[0194] Among them, is the correlation matrix, indicating whether there is a connection between two stations. is a learnable matrix that learns a unique graph for each sample and uses the classical Gaussian function to capture the similarity between sites. In Equation 2, represents the fixed connections between sites, represents the first 10 iterations ; for the subsequent iterations, A = P, and the elements of P are initialized to 1. The correlation between nodes is calculated using the K-shape function and is continuously updated during the training process. In this way, prior knowledge makes the training easier in the early stage, and the adaptive graph structure brings more flexibility.

[0195] The autocorrelation time encoding layer models the dynamic temporal correlation based on the self-attention mechanism and includes two core components: the adaptive delay transformation module can dynamically model the time delay in the data propagation process, and the autocorrelation time convolution module is used to learn the inherent periodicity in the wind data sequence.

[0196] The propagation of wind speed is affected by distance. In the offshore sites, the settings are relatively sparse and the distance between sites is far. The wind speed change at a certain site takes several minutes to affect the adjacent sites. The present invention proposes a dynamic delay feature transformation module, the goal of which is to optimize the perception ability of the network by reducing the influence of delay.

[0197] As Figure 7 shown, first, the historical data of the target station is sliced using a sliding window of size S to obtain N wind speed sequences denoted as . Next, the dynamic time warping (DTW) function is used to perform clustering operations on each to obtain the short-term wind speed patterns composed of the clustering centers . Then, the historical data sequence of each site is compared with the extracted P, and the similar pattern information is fused into the sequence representation of each node. Specifically, for the sequence of site B at time t, first, a 1X1Conv is used to obtain the high-dimensional representation of , and then the weight matrix is used to convert each data sequence in the short-term wind speed pattern P into a memory vector ; next, the softmax function is used to obtain the similarity vector , and then the weighted sum is used to obtain the final representation of the historical sequence at this moment. The weight matrices , are randomly initialized first and then continuously updated during the training process.

[0198] The wind speed has daily periodicity and seasonal characteristics, which provide regular information that can be utilized for wind speed prediction. Researchers have proposed an autocorrelation time convolution module to capture such characteristics.

[0199] As Figure 8 shown, for the input vector, first, it passes through the TCN to obtain information in the time dimension, and then through a linear transformation to get Q, K, and V. Q and K perform the Fourier transform (FFT) to obtain the corresponding frequency-domain signals Q FFT , K FFT . Take the conjugate of K FFT and multiply it element-wise with Q FFT to get QK, aiming to combine the frequency-domain information of Q and K. Perform the inverse Fourier transform on QK, and this step generates a time-domain result that fuses the frequency-domain information of Q and K. Next, select the TopK important moments from the result of the inverse Fourier transform to retain important information. The aggregated result passes through a linear layer and is added to the original V vector to supplement the underlying detailed information. In the time domain, perform the SoftMax operation on the results with different time delays to calculate the weights, and then use these weights for weighted summation, and finally fuse to obtain the final result . The overall process realizes the correlation calculation of Q and K in the frequency domain through the operations of the Fourier transform and the inverse Fourier transform, and then through time-delay aggregation and linear transformation, combines the information of V, and finally obtains the fused output result. This autocorrelation mechanism can more effectively capture the periodic characteristics in time-series data.

[0200] ;

[0201] ;

[0202] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 16 below.

[0203] Table 16 Variable Explanation Table

[0204]

[0205] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A real-time quality control method for offshore wind speed based on neural network, characterized in that, Including the following steps: Perform range detection on the input offshore wind speed data, identify abnormal peaks through spike detection, establish a multi-factor correlation model of wind speed and air pressure, perform adjacent station detection, establish a quality control symbol statistical model, set the buoy station as a node in the spatio-temporal graph, and construct a nine-layer spatio-temporal graph convolutional neural network through natural connection edges and time connection edges to extract the spatial correlation between stations and capture the temporal change characteristics, design an adaptive graph convolutional module, construct a self-correlation time encoding layer including an adaptive delay conversion module and a self-correlation time convolutional module to model the propagation delay and periodic characteristics of wind speed data, and output the quality control warning information and sensor fault alarm information of wind speed data; Among them, setting the buoy station as a node in the spatio-temporal graph and constructing a nine-layer spatio-temporal graph convolutional neural network through natural connection edges and time connection edges to extract the spatial correlation between stations and capture the temporal change characteristics includes the following steps: First, set each buoy station as a node in the spatio-temporal graph, and the node features include wind speed value and quality control identification information; then, based on the spatial relationship between stations, establish two types of edge connections: the natural connection edge represents the physical connection relationship between adjacent stations, and the time connection edge represents the connection relationship of the same station at different time steps; on this basis, construct a 9-layer spatio-temporal graph convolutional structure AST-Block, and each spatio-temporal graph convolutional structure AST-Block includes an adaptive graph convolutional module and a self-correlation time encoding layer. The adaptive graph convolution is executed before the self-correlation time encoding layer, and a residual connection is added to the entire block. Each block contains its own, separate, learnable adjacency matrix in the spatial convolution; Among them, the self-correlation time encoding layer includes an adaptive delay conversion module and a self-correlation time convolutional module. In the adaptive delay conversion module, slice the historical data through a 24-hour sliding window, and use the dynamic time warping algorithm for sequence alignment and clustering. In the self-correlation time convolutional module, convert the signal to the frequency domain through Fourier transform to calculate the self-correlation features at different time scales.

2. The real-time quality control method for offshore wind speed according to claim 1, characterized in that: The range detection includes extreme value range detection, empirical range detection, and equipment range detection. Determine the quality grade by calculating the passing rate of wind speed data. The extreme value range detection threshold is 40 meters per second. The empirical range detection sets different thresholds according to seasons. The equipment range detection range is 0 to 60 meters per second. When the passing rate is greater than 70%, the quality control symbol is set to 1. When the passing rate is less than 70%, the quality control symbol is set to 3.

3. The real-time quality control method for offshore wind speed according to claim 2, characterized in that: The spike detection is performed by setting a sliding time window, calculating the average value of wind speed data at adjacent time points before and after as the peak reference value. When the difference between the wind speed value at the current moment and the peak reference value exceeds 3 times the standard deviation, the quality control symbol is set to 3, where the standard deviation is obtained by calculating the dispersion degree of wind speed data within the previous 12 hours.

4. The real-time quality control method for offshore wind speed according to claim 3, wherein: The multi-factor correlation model of wind speed and air pressure calculates the synchronous observation values of wind speed data and air pressure data. When the wind speed change amount exceeds 2.5 times the wind speed standard deviation and the air pressure change amount is less than 2.5 times the air pressure standard deviation, the quality control symbol is set to 3.

5. The real-time quality control method for offshore wind speed according to claim 4, wherein: The adjacent station detection selects buoy stations within a distance of 200 kilometers as reference stations. When the wind speed change amount at the adjacent moments of the station to be detected exceeds 2 times the first standard deviation while the wind speed change amount at the adjacent moments of the reference station is less than 2 times the second standard deviation, the quality control flag is set to 3.

6. The real-time quality control method for offshore wind speed according to claim 5, characterized in that: The adaptive graph convolution module represents the connection relationship between stations by constructing an initial adjacency matrix, introduces a learnable correlation matrix, uses the initial adjacency matrix for the first 10 iterations in the early stage of training, and then gradually introduces the influence of the correlation matrix.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which are used to execute the method for real-time quality control of offshore wind speed based on neural network according to any one of claims 1-6 when running on a computer.

8. A real-time quality control system for offshore wind speed based on neural network, characterized in that, It includes the computer-readable storage medium according to claim 7. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is set inside the system, and a microprocessor for executing the program instructions stored in the computer-readable storage medium is set inside the system.

Images