An improved gatv2-based adaptive fusion geothermal anomaly detection method

Abstract

The application provides an adaptive fusion geothermal anomaly detection method based on improved GATv2, comprising: constructing an SGAF-GATv2 graph neural network architecture of "feature purification-double flow extraction-adaptive fusion-multi-task optimization". The architecture comprises: an input channel attention module to realize adaptive denoising purification of geochemical features; an improved GATv2 topological coding flow to combine geothermal reservoir forming mechanism to optimize dynamic attention, capture spatial distribution characteristics of "heat source", and extract "migration channel" features such as fault strike and structural unit boundary; an adaptive collaborative fusion module to dynamically adjust the contribution proportion of double-flow features through a learnable coefficient; and multi-task collaborative optimization to adopt Focal Loss and reconstruction regularization joint training to improve small sample generalization capability. The application realizes deep coupling of geothermal "source-migration" features and can be widely applied to exploration of middle-deep geothermal resources in a complex geological background.

Description

Technical Field

[0001] This invention relates to the field of deep geothermal resource exploration and artificial intelligence, and in particular to an adaptive fusion geothermal anomaly detection method based on an improved GATv2. Background Technology

[0002] Geothermal energy, as a stable and clean renewable energy source, is of strategic importance for promoting energy structure transformation and achieving the "dual carbon" goal through efficient exploration. Geochemical exploration is the core method for detecting deep geothermal reservoirs. By analyzing the anomalous distribution of geothermal-sensitive elements such as F, B, Li, Sr, and SiO2 in surface or shallow groundwater and soil, the location of deep heat sources, heat fluid migration channels, and reservoir distribution characteristics can be inverted, providing crucial information for geothermal exploration deployment. However, in complex geological contexts with well-developed faults and varied lithology (such as the Jizhong Depression), the anomalous signals of hidden geothermal resources in the middle and deep layers are often very weak and easily affected by surface environmental noise. Traditional manual interpretation and statistical analysis methods suffer from problems such as strong subjectivity, low efficiency, and severe ambiguity.

[0003] In recent years, Graph Neural Networks (GNNs), represented by Graph Convolutional Networks (GCN), GraphSAGE, and Graph Attention Networks (GAT), have provided new technical pathways for geothermal anomaly identification due to their ability to process non-Euclidean spatial data and capture complex topological relationships. Among them, GAT, by introducing an attention mechanism, can distinguish the importance of different neighboring nodes, and has initially acquired the ability to adaptively aggregate features. In particular, the dynamic attention mechanism proposed by the second-generation Graph Attention Network (GATv2) breaks the limitation of the first-generation GAT, where attention ranking was independent of the query node, providing a technical possibility for accurately characterizing the complex paths of geothermal fluid migration along fractures.

[0004] Despite the promising application prospects of GATv2, existing methods still face numerous challenges due to the scarcity of labeled samples, complex geological backgrounds, and weak geothermal anomaly signals commonly found in geothermal exploration scenarios. (i) Single-dimensional feature extraction, severing the "source-migration" accumulation mechanism. Existing GNN models (including GCN, GraphSAGE, and the unmodified GATv2) mainly focus on the topological correlation features of graph data, ignoring the synergistic mechanism between "heat source" (indicated by geochemical features) and "migration channel" (indicated by spatial tectonic features) in geothermal accumulation. This single-dimensional feature extraction method makes it difficult for the model to fully characterize the intrinsic laws of the geothermal system and to effectively integrate the coupling relationship between geochemical anomalies and tectonic location.

[0005] (ii) Insufficient adaptation of attention mechanisms, making it difficult to capture key geological interfaces. The static attention mechanism of the first-generation GAT, and the direct application of the dynamic attention mechanism of GATv2 without optimization of geothermal mechanisms, are unable to dynamically capture subtle geochemical gradient changes at reservoir boundaries and fault intersections when facing complex geological scenarios such as fault-controlled heat and stratified reservoirs. This "rigidity" of attention makes it impossible for the model to specifically identify core heat-controlling elements such as fluid migration interfaces and reservoir boundaries.

[0006] (III) Difficulty in extracting weak signals under small sample sizes limits generalization performance. Due to the high cost of drilling, labeled samples of known geothermal wells are extremely scarce. Traditional depth models are prone to overfitting in such small sample scenarios. At the same time, during the upwelling process of deep heat sources, they are diluted by surface water and interfered with by soil background, resulting in extremely weak anomalous signals. Existing multi-feature fusion methods mostly adopt a fixed-weight splicing mode, which not only breaks the collaborative mechanism of "heat source-migration channel" but also makes it difficult to effectively amplify these weak anomalous signals, resulting in insufficient generalization ability of the model in unknown areas.

[0007] To address the aforementioned technical bottlenecks, existing technologies have mainly developed three improvement approaches, but all have inherent limitations: improvement methods based on traditional GNNs are limited by the model architecture itself and struggle to handle complex nonlinear correlations in geothermal data; attention-based optimization methods remain at a static or semi-dynamic level, exhibiting poor adaptability to complex geological scenarios; and existing multi-feature fusion methods lack dynamic adjustment capabilities, failing to adaptively adjust fusion strategies according to the hydrocarbon accumulation characteristics of different geological regions, thus limiting model generalization performance. Therefore, this invention proposes an adaptive fusion geothermal anomaly detection method based on an improved GATv2. Through the collaborative design of feature purification, dual-stream extraction, adaptive fusion, and multi-task optimization, it achieves deep coupling of geothermal "source-transport" features and accurate extraction of weak anomaly signals, effectively overcoming the shortcomings of existing technologies. Summary of the Invention

[0008] Anomaly detection of medium-deep concealed geothermal resources in complex geological backgrounds is a core aspect of geothermal exploration. To address the problems of feature fragmentation, rigid attention mechanisms, and difficulty in extracting weak anomalies with small samples in existing technologies, this invention constructs a spatial-geochemical adaptive fusion (SGAF-GATv2) graph neural network architecture that combines feature purification, dual-stream extraction, adaptive fusion, and multi-task optimization. This architecture achieves deep coupling between geochemical features and spatial structural features, thereby improving the prediction accuracy and generalization ability of medium-deep concealed geothermal resources in complex geological backgrounds.

[0009] The first aspect of this invention provides an adaptive fusion geothermal anomaly detection method based on an improved GATv2, the specific steps of which include: S1. Data Input and Preprocessing: Obtain multidimensional geochemical data and spatial coordinate data of sampling points, perform logarithmic transformation and Z-score standardization on geochemical data, and perform Cartesian coordinate transformation and normalization on spatial coordinates. S2. Graph Structure Construction: A K-nearest neighbor graph is constructed based on the spatial distance of sampling points. Sampling points are used as nodes, and the edges between adjacent nodes are used to simulate the lateral transport path of thermal fluid, forming graph structure data. S3. Geochemical feature purification: Input the preprocessed geochemical data into the channel attention module, evaluate the importance of features channel by channel through the learnable weight vector, adaptively enhance the weight of geothermal sensitive elements, suppress non-geothermal interference factors, and output the purified geochemical features. S4. Geochemical Topological Feature Extraction: The purified geochemical features and graph structure data are input into the improved GATv2 topological coding stream. A two-layer stacked convolutional layer and a multi-head attention mechanism are used. The dynamic attention mechanism is optimized in combination with the geothermal accumulation mechanism. The attention of neighboring nodes is dynamically adjusted according to the geological features of the query node to capture the spatial distribution characteristics of the geothermal "heat source" and output the geochemical topological features. S5. Spatial structural feature extraction: Input the normalized plane rectangular coordinates of the sampling points into a fully connected multilayer perceptron to extract spatial structural features of geothermal "migration channels" including but not limited to fault strikes and tectonic unit boundaries, and output the spatial structural features. S6. Adaptive Feature Fusion: Geomorphological topological features and spatial structural features are input into the adaptive collaborative fusion module, which uses learnable fusion coefficients to achieve the desired fusion. A gated fusion unit is constructed, and the contribution ratio of geochemical topological features and spatial tectonic features is dynamically adjusted according to the hydrocarbon accumulation characteristics of different geological regions. Weighted fusion is then performed, and the final fused features are output. S7. Multi-task collaborative optimization and prediction: The final fused features are input into the classification task head and the reconstruction task head respectively. The classification task uses Focal Loss as the loss function to adapt to the sample imbalance problem, and the reconstruction task uses MSE Loss as the regularization term. The model parameters are optimized collaboratively by the loss weight ratio of the classification task and the reconstruction task. In the prediction stage, the geothermal anomaly probability of each node obtained from the classification task head is directly output.

[0010] Furthermore, in step S2, the value of K in the K-nearest neighbor graph is 8.

[0011] Furthermore, the attention score formula for the improved GATv2 topologically encoded stream described in step S4 is as follows: In the formula, To query node characteristics, Features of neighboring nodes W Let be the feature transformation matrix, and 'a' be the attention weight vector. This indicates a feature splicing operation.

[0012] Furthermore, in the improved GATv2 topology coding stream, the number of heads in the multi-head attention mechanism is 8, the hidden layer dimension is 64, and the output layer dimension is 32.

[0013] Furthermore, the spatially fully connected multilayer perceptron mentioned in step S5 is a 3-layer network with hidden layer dimensions of 64 and 32 respectively, and the activation function is LeakyReLU.

[0014] Furthermore, the feature fusion formula of the adaptive collaborative fusion module in step S6 is as follows: In the formula, For geomorphic topological features, As a spatial structural feature, For the final fusion features, The learnable fusion coefficient is 0 < <1.

[0015] Furthermore, in step S7, the loss function for the classification task is Focal Loss with a focusing parameter γ=2, and the weight ratio of the loss for the classification task to the reconstruction task is 1:0.3.

[0016] Furthermore, the model training employed the AdamW optimizer with a weight decay coefficient of 1e-5, an initial learning rate of 0.001 that was adaptively decayed using a cosine annealing function, 300 training epochs, an early stopping strategy to avoid overfitting, and a batch size of 16.

[0017] A second aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor as described above in an adaptive fusion geothermal anomaly detection method based on an improved GATv2.

[0018] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described adaptive fusion geothermal anomaly detection method based on improved GATv2.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1) Mechanism innovation, coupling the "source-transport" mechanism: For the first time, a dual-flow architecture of geochemical topological flow and spatial location flow is constructed. Combined with the improved GATv2 dynamic attention and ASF adaptive fusion mechanism, the deep coupling of geothermal "source-transport" characteristics is realized, which is in complete agreement with the geological formation law and fundamentally solves the core bottleneck of feature fragmentation in existing methods.

[0020] 2) Excellent performance and strong weak anomaly extraction capability: Through the synergistic effect of ICA module for denoising and purification, GATv2 dynamic attention enhancement, and reconstruction regularization constraints, the ability to identify weak anomaly signals and generalize performance in small samples are significantly improved. On the Jizhong Depression dataset, the F1 score reaches 0.9091, the accuracy reaches 0.9667, and the AUC reaches 0.9792, which is significantly better than mainstream models such as GCN, GraphSAGE, GAT, and GATv2.

[0021] 3) Adaptive and versatile: The ASF module can dynamically adjust feature weights according to the reservoir characteristics of different geological regions. It can adapt to complex geological backgrounds without manual intervention. It is not only applicable to the Jizhong Depression, but can also be extended to other medium-deep geothermal resource exploration scenarios with fault development and complex strata, and has broad engineering application value.

[0022] 4) High interpretability and fits engineering needs: The explicit dual-stream architecture and adaptive fusion logic can clearly trace the basis for anomaly identification (geochemical features or spatial structure), providing geological experts with interpretable prediction results, directly serving the delineation of geothermal drilling target areas, effectively reducing exploration costs and improving exploration efficiency. Attached Figure Description

[0023] Figure 1 The flowchart illustrates the steps of an adaptive fusion geothermal anomaly detection method based on an improved GATv2, as provided in this embodiment of the invention.

[0024] Figure 2 The diagram shows the overall architecture of the SGAF-GATv2 model, illustrating the hierarchical relationship, data flow, and collaborative logic of the ICA module, the improved GATv2 topology coding flow, the spatial location coding flow, the ASF module, and multi-task collaborative optimization.

[0025] Figure 3 This is a diagram comparing the GATv2 dynamic attention mechanism with the traditional GAT.

[0026] Figure 4 This diagram illustrates the working principle of the spatial-geochemical dual-flow collaboration, explaining the feature extraction logic of geochemical flow (simulated heat source) and spatial flow (simulated transport channel), as well as the dynamic weight adjustment mechanism of the ASF module.

[0027] Figure 5 This figure shows a performance comparison of our method with four mainstream graph neural network models—GCN, GraphSAGE, GAT, and GATv2—in the Jizhong Depression experimental area.

[0028] Figure 6The image shows the geothermal anomaly detection results and target area prediction map on the training set of the Jizhong Depression geothermal field. The colors in the image, from light to dark, represent the increasing confidence of the model prediction from 0.5 to 1.0. The high-confidence areas, from orange to dark red, are the key geothermal anomaly target areas that have been identified. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0030] Please see Figure 1 , Figure 1 A flowchart illustrating the steps of an adaptive fusion geothermal anomaly detection method based on an improved GATv2, provided in this embodiment of the invention, includes the following specific steps: S1. Data Input and Preprocessing: Obtain multidimensional geochemical data and spatial coordinate data of sampling points, perform logarithmic transformation and Z-score standardization on geochemical data, and perform Cartesian coordinate transformation and normalization on spatial coordinates. S2. Graph Structure Construction: A K-nearest neighbor graph is constructed based on the spatial distance of sampling points. Sampling points are used as nodes, and the edges between adjacent nodes are used to simulate the lateral transport path of thermal fluid, forming graph structure data. S3. Geochemical feature purification: Input the preprocessed geochemical data into the channel attention module, evaluate the importance of features channel by channel through the learnable weight vector, adaptively enhance the weight of geothermal sensitive elements, suppress non-geothermal interference factors, and output the purified geochemical features. S4. Geochemical Topological Feature Extraction: The purified geochemical features and graph structure data are input into the improved GATv2 topological coding stream. A two-layer stacked convolutional layer and a multi-head attention mechanism are used. The dynamic attention mechanism is optimized in combination with the geothermal accumulation mechanism. The attention of neighboring nodes is dynamically adjusted according to the geological features of the query node to capture the spatial distribution characteristics of the geothermal "heat source" and output the geochemical topological features. S5. Spatial structural feature extraction: Input the normalized plane rectangular coordinates of the sampling points into a fully connected multilayer perceptron to extract spatial structural features of geothermal "migration channels" including but not limited to fault strikes and tectonic unit boundaries, and output the spatial structural features. S6. Adaptive Feature Fusion: Geomorphological topological features and spatial structural features are input into the adaptive collaborative fusion module, which uses learnable fusion coefficients to achieve the desired fusion. A gated fusion unit is constructed, and the contribution ratio of geochemical topological features and spatial tectonic features is dynamically adjusted according to the hydrocarbon accumulation characteristics of different geological regions. Weighted fusion is then performed, and the final fused features are output. S7. Multi-task collaborative optimization and prediction: The final fused features are input into the classification task head and the reconstruction task head respectively. The classification task uses Focal Loss as the loss function to adapt to the sample imbalance problem, and the reconstruction task uses MSE Loss as the regularization term. The model parameters are optimized collaboratively by the loss weight ratio of the classification task and the reconstruction task. In the prediction stage, the geothermal anomaly probability of each node obtained from the classification task head is directly output.

[0031] In one specific implementation of this embodiment, the data preparation and graph structure construction process includes: (1) Dataset selection This invention validates the effectiveness of the SGAF-GATv2 model on a dataset from the Jizhong Depression geothermal field. This dataset covers the core area of ​​the Jizhong Depression, with a total area of ​​approximately 2000 km². The region features complex geological structures, well-developed faults, and varied stratigraphic lithology, making it a typical area where carbonate and sandstone geothermal reservoirs coexist. 326 shallow groundwater samples were collected from the region, and the geographical coordinates (latitude and longitude) and stratigraphic lithology background information of each sampling point were recorded simultaneously. Geochemical analysis was performed on each sample to obtain the content data of geothermal-sensitive elements such as F (fluorine), B (boron), Li (lithium), Sr (strontium), SiO2 (silicon dioxide), and As (arsenic) as original geochemical characteristics.

[0032] The dataset was divided into training, validation, and test sets in a ratio of 7:1.5:1.5. There were 200 known geothermal wells (labeled samples), which were distributed to each set in the same proportion: 140 in the training set, 30 in the validation set, and 30 in the test set, to ensure the consistency and effectiveness of training, validation, and testing.

[0033] (2) Data preprocessing Geochemical data preprocessing: Since geochemical data usually follows a log-normal distribution and the dimensions of different elements vary greatly, the content of each element is first logarithmically transformed to reduce data skewness and strengthen weak anomalous signals. Then, the logarithmically transformed data is Z-score standardized to eliminate dimensional differences and make the characteristics of each element have the same scale, finally obtaining a standardized geochemical feature matrix.

[0034] Spatial coordinate preprocessing: The latitude and longitude coordinates of the sampling points are projected into Cartesian coordinates (such as UTM projection) to obtain the coordinates of each point. Then, the coordinates are min-max normalized and scaled to the [0,1] interval to obtain the coordinate matrix, which provides input for the spatial location coding stream.

[0035] (3) Graph structure construction A K-nearest neighbor (KNN) graph is constructed based on the spatial distance between sampling points to simulate the lateral migration path of deep geothermal fluids along fractures or pores. Specifically, the Euclidean distance between any two sampling points is calculated, and for each node, the K nearest neighbors are selected to establish undirected edges. In this embodiment, K=8 is chosen, a value that has been experimentally verified to achieve a good balance between information sufficiency and computational efficiency. The final graph structure data is G=(V,E), where V is the set of nodes (326 nodes in total) and E is the set of edges. This graph structure will provide topological support for subsequent graph convolution operations and adapt to the nonlinear topological correlation characteristics of geothermal data.

[0036] In one specific implementation of this embodiment, the model construction and parameter setting process includes: This embodiment implements the SGAF-GATv2 network based on the PyTorch framework and deploys it on a server equipped with an NVIDIA RTX 1650 GPU (4GB VRAM) and 16GB of RAM. The overall architecture of the SGAF-GATv2 model is as follows: Figure 2 As shown, the specific module and parameter settings are as follows: (1) Input Channel Attention (ICA) module To address the common problems in geochemical data, such as high-dimensional redundancy, multi-element collinearity, and surface environmental noise interference, this invention employs an input channel attention module to achieve adaptive denoising and purification of geochemical features. The core of this module lies in its data-driven approach to automatically learn the importance weights of each element channel, enhancing signals of sensitive elements closely related to deep heat sources without manual intervention, while suppressing interference factors of non-geochemical origin.

[0037] Specifically, the module receives the standardized geothermal feature matrix and introduces a learnable weight vector. For each node's multidimensional feature vector, the weight vector is multiplied element-wise to obtain the weighted features. The weight vector is constrained to the (0,1) interval using the Sigmoid function to ensure that the weight of each channel has a clear physical meaning (close to 1 indicates that the element significantly contributes to geothermal anomaly identification, while close to 0 indicates that the element is mainly noise or interference). This weight vector is automatically optimized and updated through backpropagation during model training.

[0038] Through this channel-by-channel importance assessment mechanism, the module can adaptively enhance the feature weights of geothermal-sensitive elements such as Li, Sr, B, F, and SiO2, while effectively suppressing the influence of interfering factors such as non-geothermal elements, surface water dilution effects, and soil background noise. Compared to traditional methods of manually selecting features or using fixed weight combinations, this module solves the problems of strong subjectivity in feature selection and incomplete noise filtering, making it particularly suitable for complex geothermal exploration scenarios with multi-element collaborative indicators.

[0039] (2) Improve the GATv2 topology coding stream As the core improvement module of this invention, the improved GATv2 topology coding stream aims to simulate the "heat source" characteristics in geothermal systems and accurately characterize the spatial distribution and transmission patterns of deep heat sources on fracture networks. This module takes the geochemical features purified by the input channel attention module and the graph structure generated by KNN dynamic graph construction as input. It uses the second-generation graph attention network GATv2 as the graph convolution backbone and optimizes the dynamic attention mechanism in conjunction with geothermal reservoir formation mechanisms to adapt to the nonlinear topological correlation characteristics of geothermal data.

[0040] like Figure 3 The diagram illustrates a comparison between the GATv2 dynamic attention mechanism and the traditional GAT. It visually demonstrates the differences in neighbor node weight allocation under the two attention mechanisms. Addressing the static limitation of the traditional GAT, where attention ranking is independent of the query node, the dynamic attention mechanism introduced by GATv2 makes the attention score dependent on the characteristics of the query node itself. This module further integrates with geothermal exploration scenarios, enabling attention calculation to dynamically adjust the attention level to neighbor nodes based on the geological characteristics of the query node (sampling point) (such as geochemical element content and its tectonic location), thereby sensitively capturing subtle geochemical gradient changes at key locations such as fault intersections and reservoir boundaries. The attention score calculation formula is as follows: (1) In the formula, To query the features of nodes (sampling points), Features of neighboring nodes W Let be the feature transformation matrix, and 'a' be the attention weight vector. This indicates a feature splicing operation.

[0041] Specifically, the module employs two stacked GATv2 convolutional layers, each using an 8-head multi-head attention mechanism. The hidden layer dimension is set to 64, and the output layer dimension is set to 32. Through the two convolutional layers, each node can aggregate information from multi-hop neighbors, expanding its receptive field. The multi-head attention mechanism learns the association patterns between nodes from multiple perspectives, further strengthening the topological association mining of multi-dimensional features and improving the ability to perceive weak anomalous signals.

[0042] The module ultimately outputs the geomorphic topology features of each node, which encodes the spatial distribution pattern of the "heat source" on the fractured network. This effectively solves the technical bottleneck that traditional static attention cannot adapt to complex hydrothermal systems, and provides a high-quality "heat source" representation for subsequent adaptive fusion.

[0043] (3) Spatial location coding stream To overcome the inherent limitation of graph neural networks in their insensitivity to absolute location information, this invention designs a parallel spatial location encoding stream specifically to capture the spatial structural features of geothermal "migration channels," aligning with the complete geothermal "source-migration-reservoir" accumulation theory. This module aims to simulate the "channel" characteristics within a geological system, extracting macroscopic structural background information such as fault strikes, tectonic unit boundaries, and the distribution of uplifts and depressions, providing high-quality spatial feature support for subsequent adaptive fusion.

[0044] Specifically, this module receives normalized Cartesian coordinates of the sampled points and inputs them into a 3-layer fully connected multilayer perceptron (Spatial MLP). The network structure is designed as follows: 2D input layer, 64D first hidden layer, 32D second hidden layer, and 32D output layer. The activation function is uniformly LeakyReLU (with a negative slope of 0.2) to ensure in-depth mining and effective representation of spatial structural features. Through this multi-layer nonlinear mapping, the MLP can automatically learn and extract implicit fault structure background field features from simple two-dimensional coordinates. Although the model is not directly told the fault location, during training, through supervision by geothermal anomaly labels, the model can gradually learn to map the coordinate space to the structural semantic space, making points located on the same fault strike or similar structural units closer to each other in the feature space.

[0045] The output of this module is spatial tectonic features, with each node's 32-dimensional vector encoding its macroscopic tectonic location information. This design enables the model to simultaneously possess the dual capabilities of "fluid property perception" (geochemical flow) and "tectonic location perception" (spatial flow), solving the technical problem that existing methods cannot fully characterize geothermal systems due to their single feature dimension. It also provides a "migration channel" feature representation that complements the "heat source" features for subsequent adaptive collaborative fusion modules.

[0046] (4) Adaptive Synergistic Fusion (ASF) module To achieve deep coupling between geochemical topological features and spatial tectonic features, this invention designs an adaptive collaborative fusion module that simulates the logic of "multi-dimensional comprehensive judgment" by geological experts, breaking through the limitations of traditional fixed-weight splicing. For example... Figure 4 As shown, the ASF module, as the convergence hub of the dual-stream architecture, achieves dynamic weighted fusion through a gating mechanism, enabling the fusion strategy to adaptively adjust according to the hydrocarbon accumulation characteristics of different geological regions.

[0047] Specifically, the ASF module receives localized topological features from the output of the improved GATv2 topology-coded stream. Spatial construction features of spatial location encoded stream output Learnable fusion coefficients are computed through a lightweight gating network. This achieves point-by-point adaptive dynamic fusion. The fusion formula is: (2) In the formula, For geomorphic topological features, As a spatial structural feature, To achieve the final fusion of information from both the "heat source" and the "migration pathway," The learnable fusion coefficient is 0 < <1.

[0048] The core innovation of the ASF module lies in its adaptive adjustment capability driven by geological mechanisms: fusion coefficient It can automatically perceive the geological background of the current node based on input features and dynamically adjust the contribution ratio of the two major feature flows. In areas where fractures significantly control heat (such as uplift edges and fracture intersections), the gating mechanism automatically increases the spatial flow weight (i.e., To reduce the influence of tectonic channels on geothermal enrichment, the control of tectonic channels on geothermal accumulation should be strengthened. In stratabound reservoir areas (such as depression centers and areas with gentle strata), the geochemical flow weight should be emphasized (i.e., This mechanism (increases the model's ability to highlight the indicative role of heat source characteristics) allows the model to adapt to complex and ever-changing hydrocarbon accumulation environments without human intervention.

[0049] Through the above design, the ASF module has achieved a leap from "fixed weight splicing" to "geological mechanism-driven adaptive fusion", which significantly improves the model's adaptability in complex geological regions and provides a comprehensive feature representation rich in geological semantics for subsequent multi-task collaborative optimization.

[0050] (5) Multi-task collaborative optimization module To address the problem of small-sample learning and overfitting caused by the scarcity of known geothermal well samples in geothermal exploration, this invention introduces a multi-task collaborative optimization module. By jointly training the main classification task and the unsupervised feature reconstruction task, the classification accuracy is improved while the generalization ability of the model is enhanced.

[0051] During the training phase, the fused features output by the adaptive collaborative fusion module are input into two parallel task heads: the classification task head serves as the main task, where a linear layer maps the 32-dimensional features to 2 dimensions (geothermal anomaly / non-anomaly), followed by a Softmax function that outputs the anomaly probability for each node. To address the severe imbalance between positive and negative samples in geothermal exploration, Focal Loss is used as the loss function, with a focusing parameter γ=2. This loss function reduces the weight of easily classified samples, allowing the model to focus more on the sparse and difficult-to-classify geothermal anomaly samples.

[0052] The reconstruction task head serves as an auxiliary regularization term, with a decoder network reconstructing the fused features back to the original input space. The decoder structure is symmetric to the MLP in the spatial location coding stream, and the output dimension is consistent with the original input feature dimension (in this embodiment, the reconstruction target includes 6-dimensional geodesic features and 2-dimensional coordinates, for a total of 8 dimensions). The loss function uses mean squared error loss (MSELoss) to measure the difference between the reconstructed data and the original input. This auxiliary task forces the model to retain sufficient original input information in the learned hidden layer features, thereby effectively constraining the model parameter learning process, preventing overfitting, and improving sensitivity to weak anomalous signals.

[0053] The total loss function is expressed as follows: In the formula, FocalLoss represents the classification task. This represents the MSE Loss of the refactoring task. The weights represent the MSE Loss. In this embodiment, the optimal balance between classification accuracy and generalization ability can be achieved when the weight ratio of the classification task loss to the reconstruction task is 1:0.3.

[0054] In the prediction phase, only the classification task head is used to directly output the geothermal anomaly probability of each node. Through the above multi-task collaborative design, the model can still maintain excellent generalization performance in small sample scenarios with scarce labeled samples, effectively solving the technical problems of traditional deep models being prone to overfitting and difficult to extract weak anomalies.

[0055] (6) Training parameter settings The model was trained using the following parameter configuration: the optimizer was AdamW, with a weight decay coefficient of 1e-5 to mitigate overfitting; the initial learning rate was set to 0.001, and cosine annealing was used to achieve adaptive learning rate decay. The total training epochs were 300, and an early stopping mechanism was introduced to monitor the validation set loss. If the loss did not decrease for 30 consecutive epochs, training was terminated early to further improve generalization ability. Due to GPU memory limitations, the batch size was set to 16 to balance training efficiency and stability.

[0056] In terms of model structure, the improved GATv2 topology encoding flow employs two GATv2 convolutional layers, each using an 8-head attention mechanism, with a hidden layer dimension of 64 and an output layer dimension of 32. The spatial location encoding flow uses a three-layer fully connected network, with hidden layer dimensions of 64 and 32 respectively. LeakyReLU is used uniformly as the activation function to ensure deep extraction of spatially constructed features. Regarding loss function settings, Focal Loss with a focusing parameter γ=2 is used for the main classification task to adapt to the imbalance problem of geothermal samples; mean squared error loss (MSE Loss) is used for the auxiliary reconstruction task. The loss weight ratio between the main and auxiliary tasks was experimentally optimized to 1:0.3, effectively improving generalization ability under small sample sizes while ensuring classification accuracy.

[0057] In one specific implementation of this embodiment, the model is trained on the training set of the Jizhong Depression geothermal field according to the above parameter settings. After each round, the performance is evaluated on the validation set, and the learning rate and early stopping are adjusted based on the validation set loss. After training is complete, a final evaluation is performed on the test set. The experimental results and analysis are as follows: (1) Comparison of benchmark models To verify the effectiveness of the method of this invention, four mainstream graph neural network models, GCN (2017), GraphSAGE (2017), GAT (2018), and GATv2 (2021), were selected as benchmarks and comparative experiments were conducted on the same dataset. The results are as follows: Figure 5 As shown in Table 1, the core performance indicators are as follows.

[0058] Table 1 Comparison of Model Indicators (2) Results Analysis ① Improvements compared to the baseline model GATv2: Compared to the baseline model GATv2, the SGAF-GATv2 method of this invention improves the F1 score from 0.7500 to 0.9091, an improvement of 21.21%; the accuracy from 0.8667 to 0.9667, an improvement of 11.54%; and the AUC value from 0.9306 to 0.9792, an improvement of 5.22%. This significant improvement fully verifies the effectiveness of the spatial location encoding stream and ASF adaptive fusion module—the pure GATv2 model only focuses on the topological correlation features of graph data, ignoring key spatial structural information (such as stratigraphic distribution and well location spatial correlation) in geothermal exploration scenarios. In contrast, this invention achieves deep fusion of topological and spatial features through the collaboration of two modules, effectively improving the model's adaptability to complex geological regions. It should be noted that the recall rate of this invention is 0.8333, which is lower than that of GATv2 (1.0000). This is because the model reduces false positive predictions by optimizing accuracy (avoiding misclassifying non-thermal reservoir areas as thermal reservoirs), and is more in line with the actual needs of geothermal drilling deployment (reducing exploration costs and risks).

[0059] ② Overall Performance Advantages: Among all compared models, SGAF-GATv2 leads in core performance metrics. Specifically, its F1 score (0.9091) is 13.64% higher than the second-best model GAT (0.8000) and 54.56% higher than the worst-performing GCN (0.5882) (nearly 60%); its accuracy (0.9667) is 7.41% higher than GAT and GraphSAGE (both 0.9000), and 26.09% higher than GCN; its AUC (0.9792) is 2.92% higher than the second-best model GAT (0.9514), further demonstrating its superior classification and discrimination capabilities. These results indicate that the method of this invention effectively solves the problems of traditional graph convolution (such as GCN) failing to capture complex nonlinear topological patterns and GraphSAGE's insufficient adaptability to sparse topologies, achieving a breakthrough in the comprehensiveness and accuracy of feature extraction.

[0060] ③ Small-sample generalization ability: This experiment only used 200 labeled samples for model training and validation. Under this small-sample scenario, SGAF-GATv2 still maintained extremely high generalization performance, with an AUC value of 0.9792, significantly higher than the comparison model. This is due to the model's built-in reconstruction regularization auxiliary task, which effectively suppressed overfitting by constraining the feature reconstruction of unlabeled samples. It perfectly adapts to the actual scenario of scarce labeled samples in the field of geothermal exploration and has strong practical value.

[0061] In summary, the adaptive fusion geothermal anomaly detection method based on the improved GATv2 proposed in this invention achieves deep coupling of geothermal "source-transport" characteristics and accurate extraction of weak anomaly signals through the collaborative design of five innovative modules (such as...). Figure 6 As shown, this is the geothermal anomaly detection result and target area prediction map of the Jizhong Depression geothermal field training set in the embodiment of the present invention. It has excellent prediction accuracy and small sample generalization ability under complex geological background, and the model has strong interpretability and can directly serve geothermal resource exploration practice.

[0062] This invention also provides a computer-readable storage medium storing a computer program thereon, the computer program being executed by a processor of steps in an adaptive fusion geothermal anomaly detection method based on an improved GATv2, as described in any of the above embodiments.

[0063] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of an adaptive fusion geothermal anomaly detection method based on an improved GATv2 as described in any of the above embodiments.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions conceived without inventive effort should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims.

[0065] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An improved GATv2-based adaptive fusion geothermal anomaly detection method, characterized in that, A graph neural network architecture of SGAF-GATv2, consisting of "feature purification, dual-stream extraction, adaptive fusion, and multi-task optimization," is constructed to achieve accurate detection of geothermal anomalies through the collaborative efforts of five modules. Specific steps include: S1. Data Input and Preprocessing: Obtain multidimensional geochemical data and spatial coordinate data of sampling points, perform logarithmic transformation and Z-score standardization on geochemical data, and perform Cartesian coordinate transformation and normalization on spatial coordinates. S2. Graph Structure Construction: A K-nearest neighbor graph is constructed based on the spatial distance of sampling points. Sampling points are used as nodes, and the edges between adjacent nodes are used to simulate the lateral transport path of thermal fluid, forming graph structure data. S3. Geochemical feature purification: Input the preprocessed geochemical data into the channel attention module, evaluate the importance of features channel by channel through the learnable weight vector, adaptively enhance the weight of geothermal sensitive elements, suppress non-geothermal interference factors, and output the purified geochemical features. S4. Geochemical Topological Feature Extraction: The purified geochemical features and graph structure data are input into the improved GATv2 topological coding stream. A two-layer stacked convolutional layer and a multi-head attention mechanism are used. The dynamic attention mechanism is optimized in combination with the geothermal accumulation mechanism. The attention of neighboring nodes is dynamically adjusted according to the geological features of the query node to capture the spatial distribution characteristics of the geothermal "heat source" and output the geochemical topological features. S5. Spatial structural feature extraction: Input the normalized Cartesian coordinates of the sampling points into a fully connected multilayer perceptron to extract spatial structural features of geothermal "migration channels" including but not limited to fault strikes and tectonic unit boundaries, and output the spatial structural features. S6, adaptive feature fusion: input the geochemical topological features and spatial structure features into an adaptive collaborative fusion module, and dynamically adjust the contribution proportion of the geochemical topological features and the spatial structure features according to different geological regions A gating fusion unit is constructed, the contribution proportion of the geochemical topological features and the spatial structure features is dynamically adjusted according to the reservoir forming characteristics of different geological regions, weighted fusion is performed, and finally fused features are output. S7. Multi-task collaborative optimization and prediction: The final fused features are input into the classification task head and the reconstruction task head respectively. The classification task uses Focal Loss as the loss function to adapt to the sample imbalance problem, and the reconstruction task uses MSE Loss as the regularization term. The model parameters are optimized collaboratively by the loss weight ratio of the classification task and the reconstruction task. In the prediction stage, the geothermal anomaly probability of each node obtained from the classification task head is directly output.

2. The method of claim 1, wherein, In step S2, the value of K in the K-nearest neighbor graph is 8.

3. The method of claim 1, wherein, The attention score formula of the improved GATv2 topology encoding stream in step S4 is: , wherein, is a query node feature, is a neighbor node feature, W is a feature conversion matrix, and a is an attention weight vector, denotes a feature splicing operation.

4. The method of claim 1, wherein, In the improved GATv2 topology coding stream, the multi-head attention mechanism has 8 heads, a hidden layer dimension of 64, and an output layer dimension of 32.

5. The method according to claim 1, characterized in that, The spatially fully connected multilayer perceptron mentioned in step S5 is a 3-layer network with hidden layer dimensions of 64 and 32 respectively, and the activation function is LeakyReLU.

6. The method of claim 1, wherein, The feature fusion formula of the adaptive collaborative fusion module in step S6 is: In the formula, For geomorphic topological features, As a spatial structural feature, For the final fusion feature, The learnable fusion coefficient is 0 < <1.

7. The method of claim 1, wherein, In step S7, the loss function for the classification task is Focal Loss with a focusing parameter γ=2, and the weight ratio of the loss for the classification task to that for the reconstruction task is 1:0.

3.

8. The method of claim 1, wherein, The model was trained using the AdamW optimizer with a weight decay coefficient of 1e-5, an initial learning rate of 0.001 that was adaptively decayed using a cosine annealing function, 300 training epochs, an early stopping strategy to avoid overfitting, and a batch size of 16.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive fusion geothermal anomaly detection method based on the improved GATv2 as described in any one of claims 1-8.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the adaptive fusion geothermal anomaly detection method based on the improved GATv2 as described in any one of claims 1-8.

Images