Simultaneous inversion method of earthquake source location and magnitude based on graph neural network

Abstract

The present invention provides a method for simultaneously inverting the location and magnitude of induced earthquakes based on a graph neural network. The method comprises the following steps: Step 1: Data input and preprocessing; Step 2: Generating graph neural network training label data; Step 3: Establishing a graph neural network model; Step 4: Training the graph neural network model; and Step 5: Using the graph neural network to predict the location and magnitude of earthquakes. The method creates training and test datasets based on a previously accumulated catalog of gas storage-induced earthquake events. The three-component original induced earthquake waveforms and station coordinates of each station are used as the node attributes of the graph neural network. The model is trained and predicted using an adaptive moment estimation method to obtain the location and magnitude of induced earthquakes in the gas storage monitoring area. This method provides technical support for efficient, high-precision, intelligent source location of gas storage-induced earthquake events, lays a theoretical foundation for evaluating the dynamic sealing of gas storage reservoirs, guiding the optimization and adjustment of injection and production parameters, and ensuring safe production operations.

Description

Technical Field

[0001] The present invention belongs to the field of long-term monitoring of gas storage-induced earthquakes; in particular, it relates to a method for simultaneous inversion of the source location and magnitude of induced earthquakes based on a graph neural network. Background Art

[0002] During the dynamic monitoring of gas storage facilities, induced seismic monitoring, with its fast feedback speed and high sensitivity, can identify seismic activity that may be caused by production activities and, in turn, image the geometric occurrence of induced fractures. By inverting the temporal and spatial distribution and intensity of the earthquake source through induced seismicity in real time, it is possible to estimate the geometric occurrence of induced fractures, changes in ground stress, and the connectivity of the fracture grid. This allows for real-time monitoring of rock fracture activity in faults, caprocks, and substrata, providing a scientific basis for gas storage safety warnings and optimization of injection and production parameters.

[0003] Unlike hydraulic fracturing monitoring of tight oil and gas reservoirs, dynamic monitoring of gas storage facilities requires long-term monitoring of the target area, resulting in large amounts of data and significant interference from human activities and high levels of noise. This noise intersects with the distribution range of effective signals in the time and frequency domains, making it difficult to accurately pick up the first arrival of the induced seismic event, resulting in low accuracy for source location methods based on first arrival information. Furthermore, the accuracy of conventional source location methods strongly depends on the accuracy of the velocity model of the monitored area, often requiring experienced processing personnel to repeatedly adjust the velocity model and corresponding processing parameters to obtain reasonable positioning results. This method is inefficient and cannot meet the requirements for real-time monitoring of gas storage-induced earthquakes.

[0004] With the rise of deep learning in recent years, some scholars have used machine learning to locate earthquake sources. Perol et al. (2018) used a convolutional neural network to roughly zonal earthquake locations based on a single seismic record input. However, this method can only predict the approximate area where the earthquake occurred and cannot obtain the exact source location. Zhu et al. (2022) combined machine learning with cloud computing for large-capacity seismic records. They used the deep learning model PhaseNet to select the first arrivals of P and S waves and used machine learning methods to correlate the picked P and S wave first arrivals with the source location. This demonstrated the feasibility of using machine learning to predict the source location, but this method still does not directly predict the exact spatial location of the source. Cui Yi et al. (2023) published a microseismic event location method that combines spectral analysis and convolutional neural networks. This method uses wavelet transform to convert single-component high-energy frequency band microseismic waveforms into time-frequency spectrum diagrams containing time-frequency domain energy information, and constructs a mapping relationship model between the time-frequency spectrum diagram of the microseismic waveform and the source location. Zhang Yan et al. (2023) published a microseismic source location method based on a convolutional neural network. This method uses a convolutional neural network to extract the characteristics of a single-component seismic waveform and associates it with the source location, thereby predicting the spatial location of the source. However, these methods do not utilize the location information of the station and cannot predict the magnitude of the induced earthquake. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for simultaneous inversion of the source location and magnitude of induced earthquakes based on graph neural networks, to establish an intelligent positioning system for induced earthquakes with adaptability and high learning efficiency, and to simultaneously obtain the spatial location and magnitude of the source of induced earthquakes during the injection and production process of gas storage reservoirs.

[0006] The present invention is achieved through the following technical solutions:

[0007] The present invention relates to a method for simultaneously inverting the source location and magnitude of an induced earthquake based on a graph neural network, comprising the following steps:

[0008] Step 1: Data input and preprocessing: Input the original induced seismic three-component waveform data and seismic station coordinates, and perform normalization preprocessing on them; integrate the induced seismic waveform data of different periods in the gas storage monitoring area to establish a training database and a test database;

[0009] Step 2: Create graph neural network training label data: Based on the existing catalog of induced earthquakes at the gas storage facility, extract the spatial location and magnitude of the induced earthquake events and save them in text form as training labels for the graph neural network model.

[0010] Step 3: Build a graph neural network model: Based on the convolutional neural network, a feature extractor for each node attribute of the graph neural network is established. A multi-layer perceptron is used to fuse the node attribute features with the seismic station coordinates. After maximum aggregation, the aggregated feature vector of the graph neural network model is formed. Finally, a multi-layer perceptron is used to map the aggregated feature vector to the spatial location and magnitude of the earthquake source.

[0011] Step 4: Graph neural network model training: The initialized graph neural network model is trained using an adaptive moment estimation algorithm with adaptive learning rate, high computational efficiency, and strong memory, so that it can fully learn the relationship between the induced earthquake waveform characteristics, the coordinates of the seismic station and the spatial location and magnitude of the induced earthquake, and obtain a trained graph neural network model;

[0012] Step 5: Graph neural network predicts the source location and magnitude: Based on the graph neural network model trained in Step 4, the three-component waveform data of induced earthquakes and the coordinates of seismic stations in the test database are input to obtain the predicted spatial location and magnitude of the induced earthquake source.

[0013] Preferably, in Step 1, the original induced seismic three-component waveform data is the three-component induced seismic waveform data in the training database and the test database, and the processing formula is:

[0014]

[0015] Where x is the waveform data of a component of the original induced earthquake event, x m is the maximum absolute value of the waveform of all stations corresponding to the induced seismic event, and y is the normalized waveform data of this component.

[0016] Furthermore, the induced seismic waveform data of the integrated gas storage monitoring area at different periods are divided into a training data set and a test data set according to a certain ratio.

[0017] Preferably, in Step 2, extracting the spatial position and magnitude of the induced earthquake event specifically comprises extracting the x-coordinate, y-coordinate, z-coordinate and moment magnitude corresponding to the induced earthquake event.

[0018] Preferably, in Step 3, the specific steps of establishing the graph neural network model are: using a convolutional neural network model to extract the attribute characteristics of each node in the graph neural network, inputting the original three-component induced seismic waveform data into the convolutional neural network for dimensionality reduction, and extracting waveform characteristics of each scale; fusing the extracted waveform characteristics of each scale with the seismic station coordinates (x-coordinate, y-coordinate and z-coordinate), and using a multi-layer perceptron to form a feature vector of each node containing induced seismic waveform characteristics and seismic station information.

[0019] Preferably, in Step 3, the specific steps of forming the aggregated feature vector of the graph neural network model through maximum aggregation are: merging the feature vectors of each node to form a feature matrix, and performing aggregation according to the maximum row, and the formula is:

[0020] b i =max(A ij )

[0021] Where A ij is the feature matrix formed by merging the feature vectors of each node, i and j are the row and column numbers of the feature matrix respectively, max is the maximum value, b i The aggregated feature vector of the graph neural network is formed after the feature matrix is ​​aggregated.

[0022] Preferably, in Step 3, the specific steps of using a multilayer perceptron to map the aggregated feature vector to the spatial position and magnitude of the earthquake source are: passing the aggregated feature vector of the graph neural network through a multilayer perceptron, mapping it to the x-coordinate, y-coordinate, z-coordinate and moment magnitude of the induced earthquake source, and realizing the prediction of the spatial position and magnitude of the induced earthquake event.

[0023] Preferably, in Step 4, the activation function of the graph neural network model uses the ReLU function as follows: thereby alleviating the problem of gradient disappearance:

[0024] f(m)=max(0,m)

[0025] Where f(m) is the activation function value, and m is the difference between the predicted source attribute and the label:

[0026]

[0027] Where x ^ 、y ^ 、z ^ and are the x, y, z coordinates and moment magnitude of the induced earthquake predicted by the graph neural network model, respectively. o 、y o 、z o and are the true x, y, z coordinates and moment magnitude of the induced seismic event, respectively.

[0028] Preferably, in Step 4, the formula for training the initialized graph neural network model using the adaptive moment estimation algorithm with adaptive learning rate, high computational efficiency, and strong memory is as follows: the learning rate can be automatically adjusted in different parameter dimensions, thereby accelerating the convergence speed and improving the stability of training:

[0029]

[0030] Where gt is the gradient of the loss function J at the tth iteration, For the Hamiltonian operator, the adaptive moment estimation algorithm constructs the first-order moment c of the gradient of the loss function t and the second moment v t To update the parameter m:

[0031] c t =β1·c t-1 +(1-β1)·g t

[0032]

[0033] Where β1 and β2 are exponential decay rates, usually 0.9 and 0.999 respectively. The deviation correction of the first-order moment and the second-order moment is:

[0034]

[0035] Where, and are the deviation corrections of the first-order moment and the second-order moment, respectively. The parameters can be updated as follows:

[0036]

[0037] Where α is the learning rate and τ is a small constant, usually 1e -4 , used to ensure numerical stability.

[0038] Preferably, in Step 5, the specific steps of predicting the source location and magnitude of the earthquake using the graph neural network are: inputting the original induced earthquake three-component waveform data and seismic station coordinates in the test database into the trained graph neural network model to obtain the normalized source spatial location and magnitude, and then obtaining the source location and magnitude of the induced earthquake predicted in the gas storage monitoring area through denormalization.

[0039] Principle of the method of the present invention: The method of simultaneous inversion of the source position and magnitude of induced earthquakes based on graph neural networks involved in the present invention uses the original three-component waveform of the induced earthquakes and the station coordinates of each station as the attributes of each node of the graph neural network, adopts a convolutional neural network to perform dimensionality reduction on the attributes of each node of the graph neural network, extracts the feature vector of each node through a multi-layer perceptron, adopts the maximum aggregation method to obtain the aggregated feature vector of all nodes of the graph neural network, and uses the multi-layer perceptron to realize the attribute prediction of the graph neural network.

[0040] The present invention has the following advantages:

[0041] (1) The present invention uses an adaptive moment estimation algorithm to train the established graph neural network model, without the need for accurate prior information of the velocity model, so that the model can fully learn the relationship between the induced earthquake waveform information, station coordinates and source location and magnitude, thereby efficiently obtaining high-precision source location and magnitude inversion results, reducing the cost of human-computer interaction and the error introduced by manual picking, and providing technical support for the establishment of an intelligent monitoring system for induced earthquakes of gas storage reservoir integrity.

[0042] (2) The present invention proposes a method for simultaneous inversion of the source location and magnitude of induced earthquakes based on graph neural networks. This method is applicable to the field of long-term monitoring of induced earthquakes in gas storage reservoirs. It obtains the spatial location and magnitude of the source of induced earthquake events during the injection and production process of gas storage reservoirs, provides technical support for quantitative judgment of accidents such as rock fracture and fault activation inside gas storage reservoirs, and is of great significance for rapid and accurate dynamic monitoring and evaluation of the geological integrity and natural gas leakage of gas storage reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flowchart of inverting the source location and magnitude of induced earthquake events based on the graph neural network model;

[0044] Figure 2 This is a schematic diagram of the principle of inverting the source location and magnitude of induced earthquake events based on the graph neural network model;

[0045] Figure 3 This is the observation system diagram of the three-dimensional forward model for gas storage induced earthquake monitoring;

[0046] Figure 4 This is an xz slice demonstration diagram of the simulated longitudinal wave velocity model of the gas storage reservoir;

[0047] Figure 5 The waveform of the Vx component of an induced earthquake is obtained by simulating the elastic wave equation using a three-dimensional staggered grid finite difference method.

[0048] Figure 6 This is a graph of the loss function of a graph neural network model simulating gas storage-induced earthquake monitoring.

[0049] Figure 7 Comparison of the predicted and actual hypocenter locations of simulated gas storage-induced earthquakes based on the graph neural network model: top view (left), side view (center), and front view (right).

[0050] Figure 8 The graph neural network model is used to simulate the intersection of the predicted location and magnitude of the gas storage-induced earthquake source and the actual results: the intersection of the source x-coordinate (a), the source y-coordinate (b), the source z-coordinate (c), and the moment-magnitude (d).

[0051] Figure 9 This is the observation system and velocity model rendering of the induced seismic monitoring at a gas storage facility: P-wave velocity xy slice (left), P-wave velocity xz slice (right);

[0052] Figure 10 It is the waveform of the Vx component of an induced earthquake in a measured gas storage reservoir;

[0053] Figure 11 This is a graph of the loss function of the graph neural network model used to test gas storage-induced earthquake monitoring.

[0054] Figure 12 Comparison of the predicted and actual source locations of measured gas storage-induced earthquakes based on the graph neural network model: Source location results compared with a top view (left), a side view (center), and a front view (right).

[0055] Figure 13 These are the intersection plots of the predicted results of the measured gas storage-induced earthquake source location and magnitude based on the graph neural network model and the actual results: source x-coordinate intersection plot (a), source y-coordinate intersection plot (b), source z-coordinate intersection plot (c) and moment-magnitude intersection plot (d). DETAILED DESCRIPTION

[0056] The present invention will be described in detail below with reference to specific embodiments. It should be noted that the following embodiments are only for further explanation of the present invention, but the protection scope of the present invention is not limited to the following embodiments.

[0057] Example 1

[0058] This example simulates the induced earthquake data of gas storage reservoirs. Figure 1 As shown, the specific steps are:

[0059] (1) Input the three-component waveform data of simulated gas storage-induced earthquakes (SimulatedWave.segy) and the seismic station coordinates (Stations.csv), perform normalization preprocessing on the raw data, and establish training and testing databases.

[0060] (2) Based on the catalog of simulated gas storage-induced earthquake events, the x-coordinate, y-coordinate, z-coordinate and moment magnitude of the induced earthquake events are extracted as labels for training data and test data.

[0061] (3) Initialize the graph neural network model. Figure 2As shown in the figure, a convolutional neural network model is used to extract the characteristic attributes of each node, and a multi-layer perceptron is used to obtain the characteristic vector of each node. After maximum aggregation, the aggregated characteristic vector of the graph neural network is mapped to the spatial position and moment magnitude of the earthquake source in the training label through the multi-layer perceptron.

[0062] (4) The adaptive moment estimation method is used to train the graph neural network model, with the learning rate set to 0.0002, the iteration cycle set to 500, and the data batch size set to 10.

[0063] (5) The three-component waveform data of induced earthquakes and the seismic station data in the test database are input into the trained graph neural network model to obtain the predicted spatial location of the earthquake source and the moment magnitude corresponding to the induced event.

[0064] Figure 3 To simulate the forward model of induced earthquake monitoring in gas storage reservoirs, the triangles in the figure represent ground-based seismic stations, and the circles represent induced earthquake events distributed in three different regions. 100 induced earthquake events are set in each region, 250 of which are selected for training, and the other 50 events are selected for testing. Figure 4 The figure shows the xz slice demonstration of the longitudinal wave velocity model of the simulated gas storage reservoir. It can be seen that the velocity distribution of this forward simulation is relatively complex, including velocity anomalies such as faults and uplifts. Figure 5 To obtain the Vx component waveform of an induced earthquake using the three-dimensional staggered grid finite difference simulation of the elastic wave equation, Figure 6 This is the loss function diagram of the graph neural network model for simulating gas storage-induced earthquake monitoring. It can be seen that after 500 iterative training, the loss function value remains basically stable, that is, the training is completed. Figure 7 The predicted results (black dots) of the hypocenter locations of simulated gas storage-induced earthquakes based on the graph neural network model are compared with the actual results (circles). As can be seen from the figure, the distribution ranges of the two are basically consistent, verifying the correctness of the method. Figure 8 This is a cross-plot of the predicted results and actual results of the hypocenter location and magnitude of a simulated gas storage-induced earthquake based on the graph neural network model; Figure 8 a. Figure 8 b. Figure 8 c and Figure 8 d are the intersection diagrams of the predicted x-coordinate, y-coordinate, z-coordinate and moment magnitude of the earthquake source and the actual results. Figure 8 It can be seen that the predicted results and the actual results are basically distributed along the 45° oblique line, which further proves the correctness of the method of the present invention.

[0065] Example 2

[0066] This example measures induced seismic data from a gas storage facility. The specific steps are as follows:

[0067] (1) Input the three-component waveform data of gas storage-induced earthquakes (FieldWave.segy) and the seismic station coordinates (FieldStations.csv), perform normalization preprocessing on the raw data, and establish the training and testing databases.

[0068] (2) Based on the catalog of measured gas storage induced earthquake events, the x-coordinate, y-coordinate, z-coordinate and moment magnitude of the induced earthquake events are extracted as labels for training data and test data.

[0069] (3) Initialize the graph neural network model. Figure 2 As shown in the figure, a convolutional neural network model is used to extract the characteristic attributes of each node, and a multi-layer perceptron is used to obtain the characteristic vector of each node. After maximum aggregation, the aggregated characteristic vector of the graph neural network is mapped to the spatial position and moment magnitude of the earthquake source in the training label through the multi-layer perceptron.

[0070] (4) The adaptive moment estimation method is used to train the graph neural network model, with the learning rate set to 0.0002, the iteration cycle set to 500, and the data batch size set to 10.

[0071] (5) The three-component waveform data of induced earthquakes and the seismic station data in the test database are input into the trained graph neural network model to obtain the predicted spatial location of the earthquake source and the moment magnitude corresponding to the induced event.

[0072] Figure 9 This is a P-wave velocity model for monitoring gas storage-induced seismicity. The left image shows an xy slice of P-wave velocity, and the right image shows an xz slice of P-wave velocity. The triangles in the figures represent seismic stations along the surface. 285 events of measured gas storage-induced seismicity were selected for training, and another 15 events were used for testing. Figure 10 This is the waveform of the Vx component of an induced earthquake in a gas storage reservoir. It can be seen that the measured waveform recording noise is relatively large, and several traces do not record valid signals. Figure 11 This is the loss function diagram of the graph neural network model for monitoring earthquakes induced by the test gas storage reservoir. It can be seen that after 500 iterations of training, the loss function value remains basically stable, that is, the training is completed. Figure 12 The comparison chart of the predicted results (black dots) and the actual results (circles) of the earthquake source locations induced by gas storage reservoirs based on the graph neural network model is shown in Figure 2. Figure 12 It can be seen that the distribution ranges of the two are basically the same, which verifies the correctness of this method.

[0073] Figure 13 This is a cross-plot of the predicted results of the measured gas storage-induced earthquake source location and magnitude based on the graph neural network model and the actual results; Figure 13 a. Figure 13 b. Figure 13 c and Figure 13d is the intersection plot of the predicted x-coordinate, y-coordinate, z-coordinate, and moment magnitude of the earthquake source with the actual result. As can be seen from the figure, the predicted results and the actual results are basically distributed along the 45° line, which further proves the correctness of the method of the present invention.

[0074] The present invention uses an adaptive moment estimation algorithm to train the established graph neural network model, without the need for accurate prior information of the velocity model, so that the model can fully learn the relationship between induced earthquake waveform information, station coordinates and source location and magnitude, thereby efficiently obtaining high-precision source location and magnitude inversion results, reducing the cost of human-computer interaction and the error introduced by manual picking, and providing technical support for the establishment of an intelligent monitoring system for induced earthquakes of gas storage reservoir integrity.

[0075] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various variations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for simultaneous inversion of earthquake source location and magnitude based on graph neural network, characterized by: The steps include: Step 1: Data input and preprocessing: Input the original induced seismic three-component waveform data and seismic station coordinates, and perform normalization preprocessing on them; integrate the induced seismic waveform data of different periods in the gas storage monitoring area to establish a training database and a test database; Step 2: Create graph neural network training label data: Based on the information in the training and test databases, extract the spatial location and magnitude of the induced earthquake event and save it in text form as training labels for the graph neural network model; Step 3: Build a graph neural network model: Based on the convolutional neural network, a feature extractor for each node attribute of the graph neural network is established. A multi-layer perceptron is used to fuse the node attribute features with the seismic station coordinates. After maximum aggregation, the aggregated feature vector of the graph neural network model is formed. Finally, a multi-layer perceptron is used to map the aggregated feature vector to the spatial location and magnitude of the earthquake source. The specific steps of establishing the graph neural network model are as follows: using the convolutional neural network model to extract the attribute features of each node in the graph neural network, inputting the original three-component induced seismic waveform data into the convolutional neural network for dimensionality reduction, and extracting waveform features at each scale; fusing the extracted waveform features at each scale with the coordinates of the seismic station (x-coordinate, y-coordinate, z-coordinate), and using a multi-layer perceptron to form a feature vector for each node containing the induced seismic waveform features and seismic station information; Step 4: Graph neural network model training: The initialized graph neural network model is trained using an adaptive moment estimation algorithm with adaptive learning rate, high computational efficiency, and strong memory, so that it can fully learn the relationship between the induced earthquake waveform characteristics, the coordinates of the seismic station and the spatial location and magnitude of the induced earthquake, and obtain a trained graph neural network model; Step 5: Graph neural network predicts the source location and magnitude: Based on the graph neural network model trained in Step 4, the three-component waveform data of induced earthquakes and the coordinates of seismic stations in the test database are input to obtain the predicted spatial location and magnitude of the induced earthquake source.

2. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 1, the original induced seismic three-component waveform data is the three-component induced seismic waveform data in the training database and the test database, and the processing formula is: Where x is the waveform data of a component of the original induced earthquake event, x m is the maximum absolute value of the waveform of all stations corresponding to the induced seismic event, and y is the normalized waveform data of this component.

3. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 2, the extraction of the spatial position and magnitude of the induced earthquake event specifically includes: extracting the x-coordinate, y-coordinate, z-coordinate and moment magnitude corresponding to the induced earthquake event.

4. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 3, the specific steps of forming the aggregated feature vector of the graph neural network model through maximum aggregation are: merging the feature vectors of each node to form a feature matrix, and aggregating according to the maximum row, the formula is: b i =max(A ij ) Where A ij is the feature matrix formed by merging the feature vectors of each node, i and j are the row and column numbers of the feature matrix respectively, max is the maximum value, b i The aggregated feature vector of the graph neural network is formed after the feature matrix is ​​aggregated.

5. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 3, the specific steps of using a multi-layer perceptron to map the aggregated feature vector to the spatial position and magnitude of the earthquake source are as follows: the aggregated feature vector of the graph neural network is passed through a multi-layer perceptron and mapped to the x-coordinate, y-coordinate, z-coordinate and moment magnitude of the induced earthquake source, so as to realize the prediction of the spatial position and magnitude of the induced earthquake event.

6. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 4, the activation function of the graph neural network model uses the Relu function as follows: f(m)=max(0,m) Where f(m) is the activation function value, and m is the difference between the predicted source attribute and the label: Where x^, y^, z^ and are the x, y, z coordinates and moment magnitude of the induced earthquake predicted by the graph neural network model, respectively. o 、y o 、z o and are the true x, y, z coordinates and moment magnitude of the induced seismic event, respectively.

7. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 4, the formula for training the initialized graph neural network model using the adaptive moment estimation algorithm with adaptive learning rate, high computational efficiency, and strong memory is as follows: Where g t is the gradient of the loss function J at the tth iteration, For the Hamiltonian operator, the adaptive moment estimation algorithm constructs the first-order moment c of the gradient of the loss function t and the second moment v t To update the parameter m: c t =β1·c t-1 +(1-β1)·g t Where β1 and β2 are exponential decay rates, which are 0.9 and 0.999 respectively. The deviation correction of the first-order moment and the second-order moment is: Where, and are the deviation corrections of the first-order moment and the second-order moment, respectively, and the parameters are updated according to the following formula: Where α is the learning rate and τ is a small constant, usually 1e -4 , used to ensure numerical stability.

8. The method for simultaneous inversion of earthquake source location and magnitude based on graph neural network according to claim 1, characterized in that: In Step 5, the specific steps of the graph neural network to predict the source location and magnitude are as follows: the original induced earthquake three-component waveform data and seismic station coordinates in the test database are input into the trained graph neural network model to obtain the normalized source spatial location and magnitude, and then the source location and magnitude of the induced earthquake predicted in the gas storage monitoring area are obtained through denormalization.

Images