Meteorological equipment abnormal data detection method and system based on graph neural network

Abstract

The application relates to a meteorological equipment abnormal data detection method and system based on a graph neural network and belongs to the meteorological data processing field. The method comprises the following steps: S1. data collection and preliminary processing; S2. data preprocessing; S3. network construction; S4. spatiotemporal graph convolution network modeling; S5. model training; and S6. abnormality determination. The application adopts a spatiotemporal graph convolution network to jointly model meteorological data, realizes deep-level representation learning of time sequence and spatial dependence, introduces a self-supervised reconstruction error strategy at an output end of the model, learns a spatiotemporal mode of normal data, can accurately diagnose potential abnormal points without a large number of labeled abnormal samples, and can effectively reduce the dependence on abnormal labeled data.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of meteorological data processing, and particularly relates to a meteorological equipment abnormal data detection method based on a graph neural network. BACKGROUND

[0002] Current meteorological equipment plays an indispensable role in monitoring meteorological elements (such as temperature, humidity, air pressure, wind speed, etc.) and environmental changes. However, with the continuous expansion of the monitoring range and the improvement of the automation level, meteorological equipment is extremely susceptible to sensor failure, external environmental interference, network transmission and other factors during long-term operation, thereby producing various types of abnormal data. In the traditional abnormal detection method, threshold rules, statistical analysis or machine learning technology are often used to diagnose meteorological data. However, threshold type methods usually rely on experience or prior knowledge and lack flexibility for potential diversification of anomalies; and traditional machine learning models often have difficulty in accurately capturing abnormal features when there is a lack of sufficient abnormal labeled samples. Especially when the spatio-temporal distribution of observation data is complex and there is a significant spatial dependence relationship between multiple sites, relying solely on time series or geographical location for abnormal detection cannot guarantee the detection effect. In recent years, with the development of deep learning technology, graph neural networks (GNN) have shown significant potential in processing data with complex topological structures, providing a new solution for anomaly detection based on spatio-temporal correlation. By regarding meteorological stations as nodes in the graph and constructing adjacency relationships based on sensor observation values and their spatial and meteorological correlations, the dependence structure between stations can be more comprehensively described. If combined with time series information, it is expected to mine potential abnormal patterns in the dynamic changes of multi-dimensional meteorological variables. However, there are still some deficiencies in the current application of graph neural networks in the meteorological field, such as the inability to consider the dynamic interaction between multiple sites, the reliance on a large number of labeled abnormal samples during training, etc., which makes it difficult to adapt to the changing environment and diverse abnormal distribution of meteorological application scenarios. Therefore, there is an urgent need for a solution that can capture complex spatio-temporal dependence relationships and effectively learn normal meteorological data patterns in the absence of sufficient abnormal labels. SUMMARY

[0003] The application provides a meteorological equipment abnormal data detection method based on a graph neural network.

[0004] In order to achieve the above purpose, the application realizes the following technical solutions:

[0005] The application provides a meteorological equipment abnormal data detection method based on a graph neural network, comprising the following steps:

[0006] S1. Data collection and preliminary processing: collecting multidimensional meteorological observation data of each meteorological station and geographical information, to obtain a multidimensional meteorological observation data time sequence of each meteorological station , geographical information features , and an anomaly label set ;

[0007] S2. Data preprocessing: preprocessing the multidimensional meteorological observation data time sequence of each meteorological station to obtain a preprocessed multidimensional meteorological observation data time sequence of each meteorological station ; ;

[0008] S3. Network construction: constructing a meteorological station network based on the multidimensional meteorological observation data time sequence of each meteorological station, the preprocessed multidimensional meteorological observation data time sequence of each meteorological station , and the geographical information features , to obtain node features and an adjacency matrix at time step t ;

[0009] S4. Spatio-temporal graph convolution network modeling: inputting the node features and the adjacency matrix into a spatio-temporal graph convolution network model to perform feature fusion through a time series convolution operation to obtain a high-dimensional embedding representation , decoding the high-dimensional embedding representation to obtain reconstructed meteorological features , and further obtaining a reconstruction error of the meteorological station at time step t ;

[0010] S5. Model training: iteratively updating model parameters using a reconstruction error loss function with a regularization term to obtain a trained model ;

[0011] S6. Abnormality determination: using the trained model to predict and reconstruct newly input data, calculating a reconstruction error, and determining that the data is an abnormal point if the reconstruction error exceeds a predetermined threshold.

[0012] Further, step S1 specifically comprises:

[0013] S11. For each meteorological station, collecting multidimensional meteorological observation data at discrete time steps , and uniformly representing the multidimensional meteorological observation data as a vector to obtain a multidimensional meteorological observation data time sequence of each meteorological station at time step , which is expressed as follows:

[0014] ,

[0015] wherein, Represents the dimensions of meteorological variables, including temperature, humidity, air pressure, and wind speed, Indicates the Meteorological variables at meteorological stations At time step The observation value of each meteorological station The multidimensional meteorological observation data of each time step is integrated to obtain the multidimensional meteorological observation data time series of each meteorological station , the formula is as follows:

[0016] ;

[0017] S12. Collect the geographic location information and meta-information of each meteorological station to obtain geographic information features , the formula is as follows:

[0018] ,

[0019] in, represent the longitude, latitude and altitude of the meteorological station respectively;

[0020] S13. Combine the sensor self-test log and manual test results to obtain a known abnormal data annotation set , the formula is as follows:

[0021] ,

[0022] in, Indicates a weather station with abnormal data, represents the time step corresponding to the abnormal data, Indicates a multidimensional time series of meteorological observation data with anomalies.

[0023] Furthermore, step S2 specifically includes:

[0024] S21. Standardization: Set The mean of the meteorological variables at all stations and all time steps is , the standard deviation is , for Weather stations at time step Multidimensional meteorological observation data time series Standardize and obtain the standardized multidimensional meteorological observation data time series , the formula is as follows:

[0025] ,

[0026] ,

[0027] S22. Data alignment: time alignment of the standardized multi-dimensional meteorological observation data time series The time alignment operation is performed, and if observation data is missing at a certain time step, the last observation value is used for replacement to obtain the aligned multi-dimensional meteorological observation data time series , which is expressed by the following formula:

[0028] ,

[0029] wherein represents the observation value of the aligned meteorological variable at the meteorological station at the time step , and finally, the preprocessed multi-dimensional meteorological observation data time series of each meteorological station is obtained, which is expressed by the following formula:

[0030] .

[0031] Further, step S3 specifically comprises:

[0032] S31. Node definition: each meteorological station is taken as a node, and the node features include the preprocessed multi-dimensional meteorological observation data time series of each meteorological station and geographical information features ;

[0033] S32. Edge and adjacency relationship construction: the adjacency matrix of the meteorological station network at each time step is defined as , and a dynamic adjacency construction method based on meteorological variable correlation is used to calculate the meteorological variable correlation coefficient between stations within a sliding window: for the time step , the Pearson correlation coefficient of the multi-dimensional meteorological observation data time series between the meteorological station and the meteorological station is calculated within the window size , which is expressed by the following formula:

[0034] ,

[0035] wherein represents the meteorological observation subsequence of the meteorological station at the time step to , and represents the meteorological observation subsequence of the meteorological station at the time step to , and ​​denotes the complete calculation process of Pearson correlation coefficient, and the result is a normalized correlation index, ranging from [-1, 1], when the adjacency matrix at each time step is and an edge is added and a value is assigned , otherwise 0, and the formula is as follows:

[0036] ,

[0037] wherein, denotes the weight value between and in the adjacency matrix at each time step , denotes the indicator function, which is 1 when the condition in the parentheses is met, otherwise 0, denotes the threshold value of the adjacency matrix weight setting;

[0038] S33. Constructing a dynamic graph structure: based on the adjacency matrix at each time step , combined with the aligned multi-dimensional meteorological observation data time series at the current time step, a series of dynamic graph representations are formed, and the formula is as follows:

[0039] ,

[0040] , ,

[0041] wherein, denotes the dynamic graph structure, denotes the node set of the meteorological station, used to represent all nodes in the graph, denotes the total number of time steps, denotes the node feature at time step t, denotes the aligned multi-dimensional meteorological observation data time series of the meteorological station at time step t, denotes the adjacency matrix at the time step.

[0042] Further, the threshold value in step S32 is taken as the 95th percentile of the distribution.

[0043] Further, step S4 specifically comprises:

[0044] the node feature at time step t ​and an adjacency matrix Input into the spatio-temporal graph convolution network model, the feature fusion is carried out between adjacent time steps through the time series convolution, and the propagation rule of the first layer is expressed as follows: The propagation rule of the first layer is expressed as follows:

[0045] ,

[0046] Wherein, represents the node representation of the first layer, represents the initial node feature, , , represents the unit matrix, is the degree matrix of the first layer, represents the trainable weight, represents the nonlinear activation function; after the time series convolution is completed, the high-dimensional embedding representation of the time step t is obtained, and the formula is expressed as follows: ,

[0047] ,

[0048] Wherein, represents the number of time series convolution layers, represents the final embedding dimension; the high-dimensional embedding representation is input into the decoder for reconstruction, and the reconstructed meteorological feature is obtained, , , , represents the reconstructed multi-dimensional meteorological observation data of the i meteorological station, and the decoder is a multilayer perceptron. Further, in step S4, the reconstruction error of the meteorological station

[0049] at the time step t is .

[0050] Further, step S5 specifically includes:

[0051] The reconstruction error loss function is expressed as follows:

[0052] ,

[0053] On the basis of the reconstruction error loss function, a regularization term is added to obtain the training target function, and the formula is expressed as follows:

[0054] ,

[0055] Wherein, represents a hyperparameter for adjusting the regularization strength, ​​​denotes an overall set containing all trainable weights, including trainable weights in the temporal convolution and the multi-layer perceptron internal parameters, the regularization term is adopted norm.

[0056] Further, the step S6 specifically comprises:

[0057] The predetermined threshold is set by using a standard deviation method, calculating the mean value of the reconstruction error in the training set and the standard deviation , and the threshold is set as:

[0058] ,

[0059] wherein, is a proportional coefficient, and the value is 3.

[0060] The application also provides a meteorological equipment abnormal data detection system based on a graph neural network, which executes the above-mentioned meteorological equipment abnormal data detection method based on a graph neural network, and comprises:

[0061] A data collection and preliminary processing module is used to collect multi-dimensional meteorological observation data and geographic information of meteorological stations, and obtain multi-dimensional meteorological observation data time series, geographic information features and an abnormal label set of each meteorological station.

[0062] A data preprocessing module is used to preprocess the multi-dimensional meteorological observation data time series of each meteorological station, and obtain the multi-dimensional meteorological observation data time series of each meteorological station after preprocessing.

[0063] A network construction module is used to construct a meteorological station network based on the meteorological stations, the multi-dimensional meteorological observation data time series of each meteorological station after preprocessing and the geographic information features, and obtain node features and an adjacency matrix.

[0064] A space-time graph convolution network modeling unit is used to input the node features and the adjacency matrix into a space-time graph convolution network model, perform feature fusion through a temporal convolution operation to obtain a high-dimensional embedding representation, input the high-dimensional embedding representation into a decoder to obtain reconstructed meteorological features, and further obtain a reconstruction error.

[0065] A model training unit is used to iteratively update model parameters by using a reconstruction error loss function with a regularization term, and obtain a trained model.

[0066] An abnormality determination unit is used to predict and reconstruct new input data by using the trained model, calculate a reconstruction error, and determine that the data is an abnormal point if the reconstruction error exceeds a predetermined threshold.

[0067] The application has the following advantages:

[0068] The meteorological equipment abnormal data detection method based on a graph neural network is provided, a dynamic adjacent relationship construction strategy is introduced, and the real meteorological correlation can be captured in the time and space dimensions. Compared with the abnormal detection scheme based on only time series or single site information, the meteorological equipment abnormal data detection method based on a graph neural network can more comprehensively identify the abnormality generated in the interaction of multiple sites. The meteorological data is jointly modeled by using a space-time graph convolution network, deep representation learning of time series and spatial dependence is realized, a self-supervised reconstruction error strategy is introduced at the output end of the model, the space-time mode of normal data is learned, and the accurate diagnosis of potential abnormal points can be completed without a large number of labeled abnormal samples, and the dependence on abnormal labeled data is effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0069] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application.

[0070] Figure 1 The step flowchart of the method of the application is shown in the figure.

[0071] Figure 2 The experimental results of the method of the application and the conventional method are shown in the figure. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0073] Embodiment 1

[0074] In this embodiment, as shown in the figure, a meteorological equipment abnormal data detection method based on a graph neural network is provided, and the specific steps include: Figure 1

[0075] S1. Data collection and preliminary processing: multi-dimensional meteorological observation data and geographic information of K meteorological sites are collected, and the time series of multi-dimensional meteorological observation data, geographic information features and abnormal annotation sets of each meteorological site are obtained.

[0076] Specifically, S11. For each meteorological site, multi-dimensional meteorological observation data is collected at discrete time steps, and the multi-dimensional meteorological observation data is uniformly represented as a vector, so that the multi-dimensional meteorological observation data of each meteorological site at time step t is obtained. ​​​​​​Multidimensional meteorological observation data time series , the formula is as follows:

[0077] ,

[0078] in, Represents the dimensions of meteorological variables, including temperature, humidity, air pressure, and wind speed, Indicates the Meteorological variables at meteorological stations At time step The observation value of each meteorological station The multidimensional meteorological observation data of each time step is integrated to obtain the multidimensional meteorological observation data time series of each meteorological station , the formula is as follows:

[0079] ;

[0080] S12. Collect the geographic location information and meta-information of each meteorological station to obtain geographic information features , the formula is as follows:

[0081] ,

[0082] in, represent the longitude, latitude and altitude of the meteorological station respectively;

[0083] S13. Combine the sensor self-test log and manual test results to obtain a known abnormal data annotation set , the formula is as follows:

[0084] ,

[0085] in, Indicates a weather station with abnormal data, represents the time step corresponding to the abnormal data, Indicates a multidimensional time series of meteorological observation data with anomalies.

[0086] S2. Data preprocessing: Multidimensional meteorological observation data time series of each meteorological station Perform preprocessing to obtain the multi-dimensional meteorological observation data time series after preprocessing at each meteorological station ;

[0087] Specifically, S21. Standardization processing: Assume The mean of the meteorological variables at all stations and all time steps is , the standard deviation is , for Weather stations at time step multi-dimensional meteorological observation data time series standardization, obtaining the standardized multi-dimensional meteorological observation data time series , which is expressed by the following formula:

[0088] ,

[0089] ,

[0090] S22. Data alignment: performing time alignment operation on the standardized multi-dimensional meteorological observation data time series , if there is missing observation data at a certain time step, the last observation value is used for replacement, obtaining the aligned multi-dimensional meteorological observation data time series , which is expressed by the following formula:

[0091] ,

[0092] wherein, represents the observation value of the aligned meteorological variable at the meteorological station at time step ; finally obtaining the preprocessed multi-dimensional meteorological observation data time series of each meteorological station, which is expressed by the following formula:

[0093] .

[0094] S3. Network construction: based on meteorological stations, the preprocessed multi-dimensional meteorological observation data time series of each meteorological station, and geographical information features , constructing a meteorological station network, obtaining node features and adjacency matrix at time step t;

[0095] Specifically, S31. Node definition: taking each meteorological station as a node, the node features include the preprocessed multi-dimensional meteorological observation data time series of each meteorological station and geographical information features ;

[0096] S32. Edge and adjacency relationship construction: defining the adjacency matrix of the meteorological station network at each time step as , using a dynamic adjacency construction method based on meteorological variable correlation to calculate the meteorological variable correlation coefficient between stations within a sliding window: for time step , the window size is Computing the Pearson correlation coefficient of multidimensional weather observation data time series between weather stations within a range and weather stations , The formula is as follows:

[0097] ,

[0098] wherein, denotes the weather observation sub-sequence of weather station at time step to , denotes the weather observation sub-sequence of weather station at time step to , denotes the complete calculation process of the Pearson correlation coefficient, and the result is a standardized correlation index ranging from [-1, 1], when , the adjacency matrix at each time step is , and add edges and assign values

[0099] , otherwise 0, the formula is as follows:

[0100] , wherein, denotes the weight between and in the adjacency matrix at each time step , denotes the indicator function, which is 1 when the condition in the parentheses is met, otherwise 0,

[0101] denotes the threshold value of the adjacency matrix weight setting; S33. Constructing a dynamic graph structure: based on the adjacency matrix at each time step , combined with the aligned multidimensional weather observation data time series

[0102] at the current time step, a series of dynamic graph representations are formed, the formula is as follows:

[0103] , ,

[0104] wherein, denotes the dynamic graph structure, denotes the node set of weather stations, used to represent all nodes in the graph, denotes the total number of time steps, denotes the node features at time step t, denotes the aligned multi-dimensional weather observation data time series of the th weather station at time step t, denotes the aligned multi-dimensional weather observation data time series of the th weather station at time step t, denotes the adjacency matrix at the

[0105] th time step. S4. Spatio-temporal graph convolution network modeling: input the node features and the adjacency matrix into the spatio-temporal graph convolution network model to obtain high-dimensional embedding representation through the time series convolution operation, and the high-dimensional embedding representation is input into the decoder to obtain reconstructed weather features , and further obtain the reconstruction error of the weather station

[0106] at time step t; Specifically, the node features and the adjacency matrix at time step t are input into the spatio-temporal graph convolution network model, and the features are fused between adjacent time steps through time series convolution. The propagation rule formula of the

[0107] th layer is as follows:

[0108] wherein, denotes the node representation of the th layer, denotes the initial node feature, , denotes the unit matrix, is the degree matrix of , denotes the trainable weight, denotes the nonlinear activation function; after the time series convolution is completed, the high-dimensional embedding representation at time step t is finally obtained, and the formula is as follows:

[0109] ,

[0110] wherein, denotes the number of time series convolution layers, denotes the final embedding dimension; the high-dimensional embedding representation is input into the decoder for reconstruction to obtain reconstructed weather features , , denotes the reconstructed multidimensional meteorological observation data of a meteorological station, the decoder is a multilayer perceptron;

[0111] meteorological station The reconstruction error at time step t is .

[0112] S5. Model training: iteratively update the model parameters using the reconstruction error loss function with a regularization term to obtain a trained model;

[0113] Specifically, the reconstruction error loss function is represented as follows:

[0114] ,

[0115] On the basis of the reconstruction error loss function, a regularization term is added to obtain the training target function, which is represented as follows:

[0116] ,

[0117] wherein, is a hyperparameter for adjusting the regularization strength, represents the overall set containing all trainable weights, including the trainable weights in the time series convolution and the internal parameters of the multilayer perceptron, and the regularization term uses the norm.

[0118] S6. Abnormality determination: use the trained model to predict and reconstruct new input data, calculate the reconstruction error, and if the reconstruction error exceeds a predetermined threshold, it is determined to be an abnormal point.

[0119] Specifically, the predetermined threshold is set using the standard deviation method, the mean and the standard deviation of the reconstruction error in the training set are calculated, and the threshold is set as:

[0120] ,

[0121] wherein, is a proportionality coefficient.

[0122] Embodiment 2

[0123] This embodiment is to verify the effectiveness of the meteorological equipment abnormal data detection method based on the graph neural network of the present application. A comparative experiment is designed to simulate the time series data of a real meteorological station. The data contains multidimensional meteorological observation variables (such as temperature, humidity, air pressure, etc.) and labeled abnormal points. For example, Figure 2The experiment compares the method of the application with the following algorithms: a baseline model (Z-Score detection based on a sliding window), a machine learning model (LSTM and GRU time series model), and a graph neural network model (static GCN); the machine learning model is a method based only on single-site time series information, and the graph neural network model is a method of multi-site static information. The comparison results under the four indicators of accuracy (Accuracy), precision (Precision), recall (Recall), and F1 score (F1-Score) are as shown in the figure. The experimental results show that compared with the abnormal detection scheme based only on time series or single-site information, the evaluation indicators of the method of the application are higher than those of the other two methods, proving that the application can more comprehensively identify the abnormalities generated in the multi-site interaction; the application combines a dynamic adjacency matrix and a spatio-temporal graph convolution network, and can simultaneously capture the temporal variation and spatial correlation of the weather station, which is significantly better than the comparison algorithm.

[0124] Embodiment 3

[0125] The embodiment provides a meteorological equipment abnormal data detection system based on a graph neural network, which executes the meteorological equipment abnormal data detection method based on the graph neural network described in Embodiment 1, and comprises the following steps of:

[0126] A data collection and preliminary processing module is configured to collect multi-dimensional meteorological observation data and geographical information of the weather stations, and obtain a multi-dimensional meteorological observation data time series, geographical information features, and an abnormal label set of each weather station.

[0127] A data preprocessing module is configured to preprocess the multi-dimensional meteorological observation data time series of each weather station, and obtain a preprocessed multi-dimensional meteorological observation data time series of each weather station.

[0128] A network construction module is configured to construct a weather station network based on the weather stations, the preprocessed multi-dimensional meteorological observation data time series of each weather station, and the geographical information features, and obtain node features and an adjacency matrix.

[0129] A spatio-temporal graph convolution network modeling unit is configured to input the node features and the adjacency matrix into a spatio-temporal graph convolution network model, perform feature fusion through a time series convolution operation to obtain a high-dimensional embedding representation, input the high-dimensional embedding representation into a decoder to obtain reconstructed meteorological features, and then obtain a reconstruction error.

[0130] A model training unit is configured to iteratively update model parameters by using a reconstruction error loss function with a regularization term, and obtain a trained model.

[0131] Anomaly determination unit: used to predict and reconstruct the new input data using the trained model, calculate the reconstruction error, and determine it as an anomaly point if the reconstruction error exceeds a predetermined threshold.

[0132] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although in the foregoing embodiments of the application has been described in detail, for those skilled in the art, it still can be modified, or the equivalent replacement of part of the technical features described in the foregoing embodiments. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for detecting abnormal data of meteorological equipment based on graph neural network, characterized in that: The following steps are involved: S1. Data collection and preliminary processing: Collection Multidimensional meteorological observation data and geographic information of each meteorological station are used to obtain the multidimensional meteorological observation data time series of each meteorological station. , geographic information features and anomaly annotation collection ; S2. Data preprocessing: Multidimensional meteorological observation data time series of each meteorological station Perform preprocessing to obtain the multi-dimensional meteorological observation data time series after preprocessing at each meteorological station ; S3. Network construction: based on Meteorological stations, and time series of multidimensional meteorological observation data after preprocessing at each meteorological station and geographic information features Construct a weather station network and obtain node characteristics at time step t and the adjacency matrix ; The specific steps are: S31.Node definition: Each meteorological station As a node, the node features include the time series of multidimensional meteorological observation data preprocessed by each meteorological station and geographic information features ; S32. Edge and adjacency relationship construction: define the weather station network at each time step The adjacency matrix of , a dynamic neighbor construction method based on meteorological variable correlation is used to calculate the meteorological variable correlation coefficients between stations within the sliding window: , in the window size Calculate weather stations within range and weather stations Pearson correlation coefficient of multidimensional meteorological observation data time series , the formula is as follows: ， in, Indicates a weather station At time step to The meteorological observation subsequence of Indicates a weather station At time step to The meteorological observation subsequence of Represents the complete calculation process of the Pearson correlation coefficient, the result of which is a standardized correlation index in the range of [-1,1]. At each time step The adjacency matrix Zhongwei and Add edges and assign weights , otherwise it is 0. The formula is as follows: ， in, Represents each time step The adjacency matrix middle and The weight between Represents an indicator function, which is 1 when the condition in the brackets is met, otherwise it is 0. Indicates the threshold for setting the adjacency matrix weights; S33.Build dynamic graph structure: Based on each time step The adjacency matrix , combined with the aligned multidimensional meteorological observation data time series of the current time step Constitute a series of dynamic graph representations, the formula is as follows: ， ， ， in, Represents a dynamic graph structure, A node set representing a weather station, used to represent all nodes in the graph, represents the total number of time steps, represents the node features at time step t, Indicates the The aligned multidimensional meteorological observation data time series of meteorological stations at time step t, Indicates the The adjacency matrix of time steps; S4. Spatiotemporal graph convolutional network modeling: node features and the adjacency matrix Input into the spatiotemporal graph convolutional network model and perform feature fusion through temporal convolution operation to obtain a high-dimensional embedding representation , the high-dimensional embedding representation The reconstructed meteorological features are obtained through the decoder , and then get the weather station The reconstruction error at time step t; S5. Model training: The model parameters are iteratively updated using a reconstruction error loss function with a regularization term to obtain a trained model. S6. Anomaly determination: Use the trained model to predict and reconstruct the newly input data, calculate the reconstruction error, and if the reconstruction error exceeds the predetermined threshold, it is identified as an anomaly point.

2. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 1 is characterized in that: Step S1 specifically includes: S11. For each meteorological station, at discrete time steps Collect multidimensional meteorological observation data and represent them uniformly as vectors to obtain the time step of each meteorological station. Multidimensional meteorological observation data time series , the formula is as follows: ， in, Represents the dimensions of meteorological variables, including temperature, humidity, air pressure, and wind speed, Indicates the Meteorological variables at meteorological stations At time step The observation value of each meteorological station The multidimensional meteorological observation data of each time step is integrated to obtain the multidimensional meteorological observation data time series of each meteorological station , the formula is as follows: ； S12. Collect the geographic location information and meta-information of each meteorological station to obtain geographic information features , the formula is as follows: ， in, represent the longitude, latitude and altitude of the meteorological station respectively; S13. Combine the sensor self-test log and manual test results to obtain a known abnormal data annotation set , the formula is as follows: ， in, Indicates a weather station with abnormal data, represents the time step corresponding to the abnormal data, Indicates a multidimensional time series of meteorological observation data with anomalies.

3. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 2 is characterized in that: Step S2 specifically includes: S21. Standardization: Set The mean of the meteorological variables at all stations and all time steps is , the standard deviation is , for Weather stations at time step Multidimensional meteorological observation data time series Standardize and obtain the standardized multidimensional meteorological observation data time series , the formula is as follows: ， ， S22. Data alignment: Standardized multidimensional meteorological observation data time series Perform time alignment. If observation data is missing at a certain time step, the most recent observation value is used to replace it, and an aligned multidimensional meteorological observation data time series is obtained. , the formula is as follows: ， in, Indicates the alignment Meteorological variables at meteorological stations At time step Finally, we get the multidimensional meteorological observation data time series after preprocessing at each meteorological station. , the formula is as follows: 。 4. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 3 is characterized in that: Step S4 specifically includes: The node features at time step t and the adjacency matrix Input into the spatiotemporal graph convolutional network model, and perform feature fusion between adjacent time steps through temporal convolution. The propagation rule formula of the layer is expressed as follows: ， in, Indicates the Node representation of the layer, , represents the initial node features, , represents the identity matrix, for The degree matrix of represents the trainable weights, Represents a nonlinear activation function; after the temporal convolution is completed, the time step is finally obtained High-dimensional embedding representation of , the formula is as follows: ， in, represents the number of temporal convolution layers, Represents the final embedding dimension; high-dimensional embedding representation Input to the decoder for reconstruction to obtain reconstructed meteorological features , , Reconstructed The decoder is a multi-layer perceptron.

5. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 4 is characterized in that: In step S4, the weather station The reconstruction error at time step t is .

6. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 5, characterized in that: Step S5 specifically includes: The reconstruction error loss function formula is as follows: ， Add a regularization term based on the reconstruction error loss function , we get the training objective function, which is expressed as follows: ， in, represents a hyperparameter that adjusts the regularization strength, Represents the overall set of all trainable weights, including the trainable weights in the temporal convolution and the internal parameters of the multilayer perceptron, the regularization term use norm.

7. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 6, characterized in that: The threshold value in step S32 Pick The 95% quantile of the distribution.

8. The method for detecting abnormal data of meteorological equipment based on graph neural network according to claim 7 is characterized in that: Step S6 specifically includes: The predetermined threshold is set using the standard deviation method, and the mean of the reconstruction error in the training set is calculated. and standard deviation , the threshold is set to: ， in, is the proportional coefficient, and its value is 3.

9. A meteorological equipment abnormal data detection system based on graph neural network, executing the meteorological equipment abnormal data detection method based on graph neural network according to claim 1, characterized in that: include: Data collection and preliminary processing module: used to collect multi-dimensional meteorological observation data and geographic information of meteorological stations, and obtain the multi-dimensional meteorological observation data time series, geographic information features and anomaly annotation set of each meteorological station; Data preprocessing module: used to preprocess the multidimensional meteorological observation data time series of each meteorological station to obtain the preprocessed multidimensional meteorological observation data time series of each meteorological station; Network construction module: used to construct a meteorological station network based on the meteorological stations, the pre-processed multi-dimensional meteorological observation data time series of each meteorological station, and the geographic information features, and obtain node features and adjacency matrix; The spatiotemporal graph convolutional network modeling unit is used to input node features and adjacency matrices into the spatiotemporal graph convolutional network model, perform feature fusion through temporal convolution operations, and obtain a high-dimensional embedding representation. The high-dimensional embedding representation is then input into the decoder to obtain reconstructed meteorological features, and then the reconstruction error is obtained. Model training unit: used to iteratively update model parameters using a reconstruction error loss function with a regularization term added to obtain a trained model; Anomaly determination unit: used to use the trained model to predict and reconstruct the newly input data, calculate the reconstruction error, and if the reconstruction error exceeds the predetermined threshold, it is identified as an anomaly point.

Images