Geothermal resource evaluation method and system based on machine learning

Abstract

The invention discloses a geothermal resource assessment method and system based on machine learning, and relates to the technical field of resource development, and the method comprises the following steps: S1, obtaining, preprocessing and uploading multi-source data; s2, constructing a machine learning model containing a geologic feature branch, a dynamic change branch and an attention fusion layer, and obtaining a geothermal resource evaluation model through hierarchical training and multi-dimensional verification; s3, inputting real-time data to evaluate reserves, and sequencing and positioning the maximum reserve position; and S4, the development value is analyzed based on the whole-cycle cost income measurement and calculation and the return on investment rate. The method correspondingly comprises a data acquisition module, a model construction module, a preliminary evaluation module and a value evaluation module, solves the problems of erroneous judgment and resource waste caused by dependence on empirical formulas, incapability of capturing complex features and lack of economic association analysis in a traditional method, realizes accurate evaluation through a multi-branch attention fusion model, and improves the evaluation accuracy. The economic evaluation closed loop is combined to ensure the development feasibility, reduce the investment risk and improve the project success rate.

Description

Technical Field

[0001] This invention relates to the field of resource development technology, specifically to a method and system for geothermal resource assessment based on machine learning. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, geothermal resources, as a rich, stable, and sustainable renewable energy source, hold significant strategic importance for alleviating energy supply and demand imbalances and reducing carbon emissions. Geothermal resource assessment is a core prerequisite for exploration and development, requiring multi-source data integration, reservoir characteristic analysis, and reserve calculation to provide a scientific basis for development decisions. However, current geothermal resource assessment methods and related systems still suffer from numerous technical bottlenecks, failing to meet the demands for efficient and accurate assessment. Traditional methods often employ empirical formulas or single statistical models (such as regression analysis and interpolation), which cannot effectively capture the complex characteristics of geothermal resources. After obtaining reserve data, traditional methods merely determine development priorities through simple ranking without correlating reserves with development value, often resulting in misjudgments of "high reserves but low development value," leading to wasted exploration and development resources. To address this, we propose a machine learning-based geothermal resource assessment method and system. Summary of the Invention

[0003] To address the aforementioned technical issues, a machine learning-based geothermal resource assessment method and system are provided. This technical solution resolves the problem of misjudgment that arises from blindly developing areas with high reserves but high costs.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a geothermal resource assessment method based on machine learning, comprising the following steps:

[0005] S1. Acquire multi-source data, including geological data, geophysical data, geochemical data, and drilling data, and upload the data after preprocessing.

[0006] S2. Construct a machine learning model, optimize and train the model, verify the model results, and obtain a geothermal resource assessment model. The model introduces a lithological semantic segmentation algorithm to transform geological profile maps into quantifiable feature vectors. For dynamic data on geothermal gradient and groundwater flow velocity, a temporal convolutional network is used to capture the long-term trend of data changes. The weights of the two branches are dynamically adjusted based on the attention fusion layer.

[0007] S3. Input the real-time acquired data into the geothermal resource assessment model to assess the geothermal resources, obtain the current geothermal resource storage volume, sort the storage volumes, and obtain the location of the geothermal resource with the largest storage volume.

[0008] S4. Based on the return on investment, analyze the development feasibility of the current geothermal resource location and determine its development value.

[0009] S5. If geothermal resources have value, they will be acquired; otherwise, development value will be acquired sequentially based on the ranking results until a geothermal resource location with development value is obtained.

[0010] Preferably, in step S1, geological data is obtained through outdoor exploration and indoor analysis; geophysical data is obtained by using a high-precision gravimeter, laying out survey lines in a grid method, measuring surface gravity anomalies, and simultaneously conducting magnetic surveys with a proton magnetometer to capture the distribution of magnetic rock layers, the thickness of the bottom layer, the depth of fault zones, and the boundaries of thermal reservoirs; geochemical data is obtained by analyzing the chemical components in groundwater, soil, and gas; and drilling data is obtained based on the construction team's report.

[0011] Preferably, the machine learning module construction step in step S2 is as follows:

[0012] The model construction adopts a geological feature branch + dynamic change branch + attention fusion layer architecture. The geological feature branch uses a lithological semantic segmentation algorithm, which, after data preprocessing, annotation and enhancement, transforms the geological profile into a 64-dimensional static feature vector containing lithological spatial and physical parameters.

[0013] The dynamic branch uses a temporal convolutional network to process geothermal gradient and groundwater flow velocity data from more than 5 years after preprocessing, capture long-term trends, and output a 32-dimensional dynamic feature vector.

[0014] The attention fusion layer dynamically adjusts the weights of the two branches through a dual mechanism to obtain a final 96-dimensional feature vector.

[0015] Preferably, in step S2, the optimization training adopts layered training, training the geological feature branch and the dynamic change branch separately. After the validation set index of the two branches reaches the preset threshold, the attention fusion layer is trained together to accelerate model convergence.

[0016] During model training, Dropout layers and L2 regularization are introduced to suppress model overfitting.

[0017] The validation includes static feature branch validation, dynamic change branch validation, and overall model validation. Static feature branch validation uses the Dice coefficient and intersection-union ratio of semantic segmentation as evaluation metrics.

[0018] Dynamic branch validation uses the mean absolute error and root mean square error of time series prediction as evaluation indicators.

[0019] The overall model validation uses the relative error between the predicted geothermal resource reserves and the actual drilling verification values ​​as the core indicator to verify the overall evaluation performance of the model.

[0020] Preferably, in step S3, the preprocessed data is input into the model. The acquired lithological data and the surface geological structure data interpreted by remote sensing are converted into image format and input into the geological feature branch of the model. The model performs real-time segmentation of lithology, extracts static features such as the area ratio, thickness and porosity of geothermal reservoirs, generates static feature vectors, and determines the supporting capacity of the geological structure of the current area for geothermal reservoirs.

[0021] The real-time collected geothermal gradient and groundwater flow velocity time series data are input into the dynamic change branch. The short-term fluctuations and long-term trends of the data are captured by the temporal convolutional network to generate dynamic feature vectors that reflect the current dynamic change state of the geothermal reservoir.

[0022] By fusing static and dynamic feature vectors, the geothermal resource storage capacity of the current region is calculated using the formula of reservoir volume × rock heat capacity × temperature difference. The storage value of each monitoring unit is output, forming a regional geothermal resource storage distribution map.

[0023] Preferably, drilling data is pre-converted into depth-lithology curves and aligned with the coordinates of remotely sensed surface geological structure vector maps, and then the two are merged into a raster image using GIS software;

[0024] The converted image uses adaptive histogram equalization to enhance lithological boundaries and median filtering to remove drilling data noise, outputting a 512×512 pixel grayscale image.

[0025] The calculation of the thermal reservoir volume involves using eight-neighbor connected domain analysis in the segmentation results to remove isolated thermal reservoir fragments with an area of ​​less than 500 pixels, and then calculating the effective thermal reservoir pixel ratio.

[0026] Preferably, in step S4, the determination of the power generation and heating revenue model is made by calculating the costs of the entire life cycle of preliminary detailed survey, construction and operation, calculating the annual net cash flow, calculating static and dynamic indicators, and making a preliminary assessment of feasibility based on industry benchmark values. Then, the impact of parameter fluctuations on the return on investment (ROI) is tested through sensitivity analysis. Combining the ROI with the risk level, the development value is classified as high, medium and low. Finally, a feasibility report is output to clarify the development value.

[0027] Preferably, the preliminary detailed survey includes detailed survey fees, approval fees, and land acquisition fees; construction costs include drilling engineering fees, heat exchange system equipment fees, and power and heat transmission pipeline network fees; operating costs include annual maintenance fees, water and electricity fees, labor costs, and water treatment fees; power generation revenue is calculated based on installed capacity * annual utilization hours * grid connection price, and heating revenue is calculated based on coverage area * unit area charge. The annual net cash flow calculation formula is as follows:

[0028] Net cash flow = Total annual revenue - Annual operating costs - Fixed asset depreciation

[0029] The initial investment is negative in year 0, and operating cash flow is generated from year 1 onwards, taking into account the impact of inflation on costs and revenues.

[0030] Preferably, the return on investment (ROI) is calculated using the following formula:

[0031] Return on Investment (ROI) = (Average Annual Net Income ÷ Total Investment) × 100%

[0032] The industry benchmark values ​​are based on big data and range from 8% to 12%.

[0033] By combining return on investment (ROI) and risk level, value ranges are defined, and the ROI of the current geothermal resource location is matched to the defined value range, and a development feasibility report is generated.

[0034] A machine learning-based geothermal resource assessment system, the assessment system comprising:

[0035] The data acquisition module is used to acquire data from multiple sources, preprocess it, and then upload it.

[0036] The model building module is used to build a geothermal resource assessment model, train and optimize it, and then use it for evaluation.

[0037] The preliminary assessment module inputs real-time acquired data into the geothermal resource assessment model to assess the geothermal resources, obtain the current geothermal resource storage capacity, sort them, and obtain the location of the geothermal resource with the largest storage capacity.

[0038] The value assessment module analyzes the developability based on the return on investment and determines the development value of the current geothermal resource location.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention proposes to transform visualized geological profile maps into quantifiable feature vectors, replacing subjective human judgments of lithology and avoiding experience bias. It accurately captures the long-term trends of geothermal gradients and groundwater flow rates, dynamically adjusts branch weights, and makes model evaluations more closely aligned with actual geological scenarios. It performs full-cycle calculations from both cost and benefit perspectives, clearly defining key indicators such as the project's payback period, net present value, and internal rate of return. By combining industry benchmarks to determine development value, it avoids blindly developing areas with high reserves but high costs, ensuring that each proposed development location is economically feasible, reducing corporate investment risks, and increasing project success rates. Attached Figure Description

[0041] Figure 1This is a flowchart of the evaluation steps of the present invention. Detailed Implementation

[0042] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0043] Reference Figure 1 As shown, a machine learning-based geothermal resource assessment method includes the following steps:

[0044] S1. Acquire multi-source data, including geological data, geophysical data, geochemical data, and drilling data, and upload the data after preprocessing.

[0045] S2. Construct a machine learning model, optimize and train the model, verify the model results, and obtain a geothermal resource assessment model. The model introduces a lithological semantic segmentation algorithm to transform geological profile maps into quantifiable feature vectors. For dynamic data on geothermal gradient and groundwater flow velocity, a temporal convolutional network is used to capture the long-term trend of data changes. The weights of the two branches are dynamically adjusted based on the attention fusion layer.

[0046] S3. Input the real-time acquired data into the geothermal resource assessment model to assess the geothermal resources, obtain the current geothermal resource storage volume, sort the storage volumes, and obtain the location of the geothermal resource with the largest storage volume.

[0047] S4. Based on the return on investment, analyze the development feasibility of the current geothermal resource location and determine its development value.

[0048] S5. If geothermal resources have value, they will be acquired; otherwise, development value will be acquired sequentially based on the ranking results until a geothermal resource location with development value is obtained.

[0049] This application integrates multi-source data from geology, geophysics, and other sources into its data layer. Preprocessing reduces noise, laying the foundation for accurate assessment. The model layer combines lithological semantic segmentation, temporal convolutional networks, and attention fusion layers to quantify geological graphics, capture long-term trends in dynamic data, and dynamically balance feature weights, significantly improving assessment accuracy. The location layer quickly identifies high-potential areas through quantitative storage assessment and sorting, avoiding disorderly exploration. The decision-making layer introduces ROI analysis, balancing technical and economic feasibility, filtering out ineffective development projects, and reducing investment risk. The process layer's cyclical screening mechanism forms a closed loop, ensuring efficient identification of valuable development areas and achieving an upgrade from experience-driven to data-driven approaches, providing scientific support for efficient exploration and precise development of geothermal resources.

[0050] In step S1, geological data is acquired through outdoor exploration and indoor analysis. Geophysical data is acquired by using a high-precision gravimeter, laying out survey lines in a grid method, measuring surface gravity anomalies, and simultaneously conducting magnetic surveys with a proton magnetometer to capture the distribution of magnetic rock layers, the thickness of the bottom layer, the depth of fault zones, and the boundaries of thermal reservoirs. Geochemical data is obtained by analyzing the chemical components in groundwater, soil, and gas. Drilling data is obtained based on the construction team's reports.

[0051] The various data acquisition methods in step S1 of this application ensure the quality and value of the data required for geothermal resource assessment from the source, laying a solid foundation for subsequent model construction and resource assessment. In terms of data accuracy, geophysical data acquisition uses a high-precision gravimeter to lay out survey lines in a grid method, which can systematically and standardizedly measure surface gravity anomalies. Combined with synchronous magnetic measurement by a proton magnetometer, it can accurately capture the distribution of magnetic rock layers, stratum thickness, fault zone depth and thermal reservoir boundaries, avoiding the errors of traditional random measurements and making the data related to underground geological structures more accurate.

[0052] In terms of data comprehensiveness, geological data is obtained by combining outdoor exploration with indoor analysis, covering surface and deep geological information; geochemical data supplements the chemical dimension information not covered by geological and geophysical data by analyzing the chemical composition of groundwater, soil and gas; drilling data provides direct information on the actual underground geological conditions. These four types of data cover the information required for geothermal resource assessment from different dimensions, thus achieving data comprehensiveness.

[0053] The steps for building the machine learning module in step S2 are as follows:

[0054] The model construction adopts a geological feature branch + dynamic change branch + attention fusion layer architecture. The geological feature branch uses a lithological semantic segmentation algorithm, which, after data preprocessing, annotation and enhancement, transforms the geological profile into a 64-dimensional static feature vector containing lithological spatial and physical parameters.

[0055] The dynamic branch uses a temporal convolutional network to process geothermal gradient and groundwater flow velocity data from more than 5 years after preprocessing, capture long-term trends, and output a 32-dimensional dynamic feature vector.

[0056] The attention fusion layer dynamically adjusts the weights of the two branches through a dual mechanism to obtain a final 96-dimensional feature vector.

[0057] In step S2, the optimization training adopts layered training. The geological feature branch and the dynamic change branch are trained separately. After the validation set index of the two branches reaches the preset threshold, the attention fusion layer is trained together to accelerate the model convergence.

[0058] During model training, Dropout layers and L2 regularization are introduced to suppress model overfitting.

[0059] The validation includes static feature branch validation, dynamic change branch validation, and overall model validation. Static feature branch validation uses the Dice coefficient and intersection-union ratio of semantic segmentation as evaluation metrics.

[0060] Dynamic branch validation uses the mean absolute error and root mean square error of time series prediction as evaluation indicators.

[0061] The overall model validation uses the relative error between the predicted geothermal resource reserves and the actual drilling verification values ​​as the core indicator to verify the overall evaluation performance of the model.

[0062] At the architectural design level, this application employs a highly targeted architecture consisting of a geological feature branch, a dynamic change branch, and an attention fusion layer. The geological feature branch uses a lithological semantic segmentation algorithm to transform geological profiles into 64-dimensional static feature vectors, accurately extracting core static information such as lithological spatial distribution and physical parameters. The dynamic change branch utilizes a temporal convolutional network to process geothermal gradient and groundwater flow velocity data spanning more than 5 years, outputting 32-dimensional dynamic feature vectors to effectively capture long-term data trends. The two branches focus on key static and dynamic features respectively, and the attention fusion layer dynamically adjusts the weights to obtain a final 96-dimensional feature vector. This ensures that neither type of core feature is omitted, while flexibly balancing their impact based on the actual data characteristics. This makes the model more aligned with the complex needs of geothermal resource assessment, which combines static geological foundations with dynamic environmental impacts, avoiding assessment biases caused by a single feature dimension.

[0063] Optimizing the training strategy significantly improves efficiency and effectiveness. The hierarchical training mode first trains the two branches separately, and then trains the attention fusion layer together after the validation set index reaches the preset threshold. This avoids mutual interference when the two branches and the fusion layer are trained simultaneously. Each branch focuses on learning its own corresponding features first, and the fusion logic is optimized after laying a solid foundation. This greatly reduces the complexity of model training, accelerates the convergence process, and reduces the overall training time cost. The introduction of the Dropout layer and L2 regularization specifically addresses the overfitting problem. The Dropout layer avoids the model's over-reliance on local features by randomly "shutting down" some neurons. L2 regularization prevents the model from "over-memorizing" the training data by limiting the parameter size. The combination of the two allows the model to maintain stable performance in new scenarios outside the training data and improves generalization ability.

[0064] In step S3, the preprocessed data is input into the model. The acquired lithological data and the surface geological structure data interpreted by remote sensing are converted into image format and input into the geological feature branch of the model. The model performs real-time segmentation of lithology, extracts static features such as the area ratio, thickness and porosity of geothermal reservoirs, generates static feature vectors, and determines the supporting capacity of the geological structure of the current area for geothermal reservoirs.

[0065] The real-time collected geothermal gradient and groundwater flow velocity time series data are input into the dynamic change branch. The short-term fluctuations and long-term trends of the data are captured by the temporal convolutional network to generate dynamic feature vectors that reflect the current dynamic change state of the geothermal reservoir.

[0066] By fusing static and dynamic feature vectors, the geothermal resource storage capacity of the current region is calculated using the formula of reservoir volume × rock heat capacity × temperature difference. The storage value of each monitoring unit is output, forming a regional geothermal resource storage distribution map.

[0067] Assuming that real-time drilling lithology data and remote sensing geological structure data are converted into a format to form an input image matrix H represents the image height, W represents the width, and C represents the number of channels. Lithological semantic segmentation is performed using a U-Net++ network, outputting a lithology classification matrix. S i,j This represents the lithology category corresponding to image pixel (i,j);

[0068] The formula for calculating the area ratio R of the thermal reservoir is:

[0069]

[0070] in This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise. It represents the ratio of the total number of pixels in the thermal reservoir to the total number of pixels in the image.

[0071] The formula for calculating the thickness h of a thermal reservoir is:

[0072]

[0073] Among them, z max z min These represent the maximum and minimum depths of the thermal reservoir in the vertical direction, respectively. This represents the average thickness of the thermal reservoir within the region.

[0074] The formula for calculating the porosity φ of a thermal reservoir is:

[0075]

[0076] Where, φ k denoted as porosity of the kth core sample from the thermal reservoir obtained in real-time drilling, where n is the number of core samples.

[0077] After standardizing the static indices of thermal reservoir area ratio, thickness, porosity, and other derived parameters, a static feature vector X is constructed. static The calculation formula is:

[0078] X static =[x static,1 ,xstatic,2 ,…,x static,64 ] T

[0079] in p k The original parameter value, μ k σ k These are the mean and standard deviation of the parameters in the training set, respectively.

[0080] Let the real-time collected geothermal gradient time series data be: T(t)={T1,T2,…,T m}, T t Let m be the geothermal gradient at time t, m = 72 hours / monitoring interval, i.e., the total number of high-frequency monitoring data points. The groundwater flow velocity time series data is V(t) = {V1, V2, ..., V...} m}, V t Let be the water flow velocity at time t;

[0081] rate of change of geothermal gradient r T The calculation formula is:

[0082]

[0083] Where Δt is the time interval of the time series data, representing the average rate of change of the geothermal gradient over the past 72 hours;

[0084] Groundwater flow velocity stability δ V The calculation formula is:

[0085]

[0086] in V t The average water flow velocity over the past 72 hours, δ V The coefficient of variation;

[0087] After standardizing the rate of change of geothermal gradient, water flow velocity stability, and other dynamic indicators, temporal features are extracted using a temporal convolutional network to construct a dynamic feature vector X. dynamic The formula is expressed as:

[0088] X dynamic =[x dynamic,1 ,x dynamic,2 ,…,x dynamic,32 ] T

[0089] Let the regional geological complexity coefficient be λ, then the dynamic branch weight ω d With static branch weight ω s satisfy:

[0090] ω d =0.4+0.2λ, ωs =0.6-0.2λ, and ω d +ω s =1;

[0091] By weighted fusion of static and dynamic feature vectors X fusion The expression is:

[0092]

[0093] Vector operations involve weighted summation of elements in corresponding dimensions, ultimately yielding...

[0094] Based on the fused feature vector X fusion Input the trained random forest storage regression model f(·), and calculate using the core formula for geothermal storage:

[0095] Q = f(X) fusion )×ρ×c×ΔT×V res

[0096] Where f(X) fusion ) represents the reservoir correction coefficient output by the model; ρ represents the density of the thermal reservoir rock; c represents the specific heat capacity of the rock; ΔT represents the temperature difference between the thermal reservoir and ambient temperature; V res The volume of the thermal reservoir;

[0097] Finally, the reserves Q of each 1km×1km monitoring unit are output, forming a regional geothermal resource reserve distribution map.

[0098] Drilling data is pre-converted into depth-lithology curves and aligned with the coordinates of the surface geological structure vector map interpreted by remote sensing. The two are then merged into a raster image using GIS software.

[0099] The converted image uses adaptive histogram equalization to enhance lithological boundaries and median filtering to remove drilling data noise, outputting a 512×512 pixel grayscale image.

[0100] The choice of 512×512 pixel resolution is based on three technical considerations: scale adaptation for geothermal lithology characteristics, as the typical scale of geothermal lithology boundaries and reservoir distribution ranges from several meters to hundreds of meters, and 512×512 resolution can accurately depict lithological details at a scale of 1:1000, balancing feature accuracy and computational efficiency; model training compatibility, as mainstream semantic segmentation networks (such as U-Net) have mature optimizations for 512×512 input, and this resolution can balance memory usage and feature extraction capabilities, avoiding the loss of key lithological boundaries due to excessively small resolutions or the decrease in training efficiency due to excessively large resolutions; and industry data standard alignment, matching the resolution of remote sensing images commonly used in the geothermal exploration field (such as 50cm-2m / pixel), facilitating scale alignment during multi-source data fusion.

[0101] The calculation of the thermal reservoir volume involves using eight-neighbor connected domain analysis in the segmentation results to remove isolated thermal reservoir fragments with an area of ​​less than 500 pixels, and then calculating the effective thermal reservoir pixel ratio.

[0102] This application improves the efficiency and accuracy of geothermal research from three aspects: data integration, image quality, and thermal reservoir calculation. Coordinate alignment and fusion of drilling and remote sensing data can link underground lithological information with surface geological structure, breaking the isolation of data and providing a complete and three-dimensional geological background for subsequent analysis, reducing the bias of single data. Adaptive histogram equalization strengthens lithological boundaries, and median filtering eliminates noise and outputs standardized grayscale images, which not only highlight key geological features but also improve image clarity and consistency, making it easier to accurately identify lithological distribution and reduce human interpretation errors.

[0103] In step S4, the feasibility is determined by calculating the costs of the entire lifecycle of preliminary detailed exploration, construction and operation, identifying the revenue model for power generation and heating, calculating the annual net cash flow, and measuring static and dynamic indicators. The feasibility is initially judged based on industry benchmark values. Then, sensitivity analysis is used to test the impact of parameter fluctuations on the return on investment (ROI). Combining the ROI with the risk level, the development value is classified as high, medium and low. Finally, a feasibility report is output to clarify the development value.

[0104] This application provides a scientific and comprehensive basis for project development decisions through full-cycle cost calculation, multi-dimensional indicator analysis, and risk assessment. Full-cycle cost and benefit calculation lays a solid foundation for decision-making, covering cost calculations for preliminary detailed surveys, construction, and operation. Combined with the calculation of annual net cash flow based on power generation and heating revenue models, it can accurately grasp the economic dynamics of the project, avoid decision-making deviations caused by cost omissions or misjudgments of benefits, and make the economic profile of the project clearly visible. It assesses the project's profitability potential from different time dimensions. Sensitivity analysis tests the impact of parameter fluctuations on ROI, which can identify key risk points in advance and avoid decision-making errors caused by relying on a single indicator or ignoring parameter changes.

[0105] The preliminary detailed survey includes survey fees, approval fees, and land acquisition fees; construction costs include drilling costs, heat exchange system equipment costs, and power and heat transmission pipeline costs; operating costs include annual maintenance fees, water and electricity fees, labor costs, and water treatment fees; power generation revenue is calculated based on installed capacity * annual utilization hours * grid connection price, and heating revenue is calculated based on coverage area * unit area charge. The annual net cash flow calculation formula is as follows:

[0106] Net cash flow = Total annual revenue - Annual operating costs - Fixed asset depreciation

[0107] The initial investment is negative in year 0, and operating cash flow is generated from year 1 onwards, taking into account the impact of inflation on costs and revenues.

[0108] The return on investment (ROI) is calculated using the following formula:

[0109] Return on Investment (ROI) = (Average Annual Net Income ÷ Total Investment) × 100%

[0110] The industry benchmark values ​​are based on big data and range from 8% to 12%.

[0111] By combining return on investment (ROI) and risk level, value ranges are defined, and the ROI of the current geothermal resource location is matched to the defined value range, and a development feasibility report is generated.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A machine learning-based method for geothermal resource assessment, characterized in that, Includes the following steps: S1. Acquire multi-source data, including geological data, geophysical data, geochemical data, and drilling data, and upload the data after preprocessing. S2. Construct a machine learning model, optimize and train the model, verify the model results, and obtain a geothermal resource assessment model. The model introduces a lithological semantic segmentation algorithm to transform geological profile maps into quantifiable feature vectors. For dynamic data on geothermal gradient and groundwater flow velocity, a temporal convolutional network is used to capture the long-term trend of data changes. The weights of the two branches are dynamically adjusted based on the attention fusion layer. S3. Input the real-time acquired data into the geothermal resource assessment model to assess the geothermal resources, obtain the current geothermal resource storage volume, sort the storage volumes, and obtain the location of the geothermal resource with the largest storage volume. S4. Based on the return on investment, analyze the development feasibility of the current geothermal resource location and determine its development value. S5. If geothermal resources have value, they will be acquired; otherwise, development value will be acquired sequentially based on the ranking results until a geothermal resource location with development value is obtained.

2. The geothermal resource assessment method based on machine learning according to claim 1, characterized in that: In step S1, geological data is obtained through outdoor exploration and indoor analysis. Geophysical data is acquired by using a high-precision gravimeter, laying out survey lines in a grid method, measuring surface gravity anomalies, and simultaneously conducting magnetic surveys with a proton magnetometer to capture the distribution of magnetic rock layers, the thickness of the bottom layer, the depth of fault zones, and the boundaries of thermal reservoirs. Geochemical data is obtained by analyzing the chemical components in groundwater, soil, and gas. Drilling data is obtained based on the construction team's report.

3. The geothermal resource assessment method based on machine learning according to claim 1, characterized in that: The steps for building the machine learning module in step S2 are as follows: The model construction adopts a geological feature branch + dynamic change branch + attention fusion layer architecture. The geological feature branch uses a lithological semantic segmentation algorithm, which, after data preprocessing, annotation and enhancement, transforms the geological profile into a 64-dimensional static feature vector containing lithological spatial and physical parameters. The dynamic branch uses a temporal convolutional network to process geothermal gradient and groundwater flow velocity data from more than 5 years after preprocessing, capture long-term trends, and output a 32-dimensional dynamic feature vector. The attention fusion layer dynamically adjusts the weights of the two branches through a dual mechanism to obtain a final 96-dimensional feature vector.

4. The geothermal resource assessment method based on machine learning according to claim 1, characterized in that: In step S2, the optimization training adopts layered training. The geological feature branch and the dynamic change branch are trained separately. After the validation set index of the two branches reaches the preset threshold, the attention fusion layer is trained together to accelerate the model convergence. During model training, Dropout layers and L2 regularization are introduced to suppress model overfitting. The validation includes static feature branch validation, dynamic change branch validation, and overall model validation. Static feature branch validation uses the Dice coefficient and intersection-union ratio of semantic segmentation as evaluation metrics. Dynamic branch validation uses the mean absolute error and root mean square error of time series prediction as evaluation indicators. The overall model validation uses the relative error between the predicted geothermal resource reserves and the actual drilling verification values ​​as the core indicator to verify the overall evaluation performance of the model.

5. The geothermal resource assessment method based on machine learning according to claim 1, characterized in that: In step S3, the preprocessed data is input into the model. The acquired lithological data and the surface geological structure data interpreted by remote sensing are converted into image format and input into the geological feature branch of the model. The model performs real-time segmentation of lithology, extracts static features such as the area ratio, thickness and porosity of geothermal reservoirs, generates static feature vectors, and determines the supporting capacity of the geological structure of the current area for geothermal reservoirs. The real-time collected geothermal gradient and groundwater flow velocity time series data are input into the dynamic change branch. The short-term fluctuations and long-term trends of the data are captured by the temporal convolutional network to generate dynamic feature vectors that reflect the current dynamic change state of the geothermal reservoir. By fusing static and dynamic feature vectors, the geothermal resource storage capacity of the current region is calculated using the formula of reservoir volume × rock heat capacity × temperature difference. The storage value of each monitoring unit is output, forming a regional geothermal resource storage distribution map.

6. The geothermal resource assessment method based on machine learning according to claim 5, characterized in that: Drilling data is pre-converted into depth-lithology curves and aligned with the coordinates of the surface geological structure vector map interpreted by remote sensing. The two are then merged into a raster image using GIS software. The converted image uses adaptive histogram equalization to enhance lithological boundaries and median filtering to remove drilling data noise, outputting a 512×512 pixel grayscale image. The calculation of the thermal reservoir volume involves using eight-neighbor connected domain analysis in the segmentation results to remove isolated thermal reservoir fragments with an area of ​​less than 500 pixels, and then calculating the effective thermal reservoir pixel ratio.

7. The geothermal resource assessment method based on machine learning according to claim 1, characterized in that: In step S4, the feasibility is determined by calculating the costs of the entire lifecycle of preliminary detailed exploration, construction and operation, identifying the revenue model for power generation and heating, calculating the annual net cash flow, and measuring static and dynamic indicators. The feasibility is initially judged based on industry benchmark values. Then, sensitivity analysis is used to test the impact of parameter fluctuations on the return on investment (ROI). Combining the ROI with the risk level, the development value is classified as high, medium and low. Finally, a feasibility report is output to clarify the development value.

8. The geothermal resource assessment method based on machine learning according to claim 7, characterized in that: The preliminary detailed survey includes survey fees, approval fees, and land acquisition fees; construction costs include drilling costs, heat exchange system equipment costs, and power and heat transmission pipeline costs; operating costs include annual maintenance fees, water and electricity fees, labor costs, and water treatment fees; power generation revenue is calculated based on installed capacity * annual utilization hours * grid connection price, and heating revenue is calculated based on coverage area * unit area charge. The annual net cash flow calculation formula is as follows: Net cash flow = Total annual revenue - Annual operating costs - Fixed asset depreciation The initial investment is negative in year 0, and operating cash flow is generated from year 1 onwards, taking into account the impact of inflation on costs and revenues.

9. The geothermal resource assessment method based on machine learning according to claim 7, characterized in that, The return on investment (ROI) is calculated using the following formula: Return on Investment (ROI) = (Average Annual Net Income ÷ Total Investment) × 100% The industry benchmark values ​​are based on big data and range from 8% to 12%. By combining return on investment (ROI) and risk level, value ranges are defined, and the ROI of the current geothermal resource location is matched to the defined value range, and a development feasibility report is generated.

10. A machine learning-based geothermal resource assessment system, characterized in that, A machine learning-based geothermal resource assessment method applicable to any one of claims 1 to 9, the assessment system comprising: The data acquisition module is used to acquire data from multiple sources, preprocess it, and then upload it. The model building module is used to build a geothermal resource assessment model, train and optimize it, and then use it for evaluation. The preliminary assessment module inputs real-time acquired data into the geothermal resource assessment model to assess the geothermal resources, obtain the current geothermal resource storage capacity, sort them, and obtain the location of the geothermal resource with the largest storage capacity. The value assessment module analyzes the developability based on the return on investment and determines the development value of the current geothermal resource location.

Images