Geothermal energy-oriented development decision management method and system

Abstract

The invention relates to the technical field of artificial intelligence, solves the technical problems that in the prior art, limitation of physical laws is prone to being neglected, complex coupling effects cannot be captured, large-scale strategy exploration is difficult to support, and geothermal energy overpressure mining occurs due to local optimum, and particularly relates to a development decision management method and system for geothermal energy. The method comprises the steps that geothermal multi-source data in a geothermal system is collected, a geothermal associated database is obtained through preprocessing and integration, a high-fidelity data set covering multiple scenes is generated through parameter uncertainty sampling, and a built physical sensing simulator can achieve the multi-scene high-fidelity data set while keeping the coupling physical law. The time consumption of single simulation is compressed from several hours to a second level, dynamic optimization of a development strategy under physical constraints is achieved through connection interaction of the DQN and the physical perception simulator, and the problems that the model falls into local optimum or violates the physical constraints to cause overpressure mining and the like are avoided.

Description

A development decision-making management method and system for geothermal energy. Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a development decision-making management method and system for geothermal energy. Background Technology

[0002] Artificial intelligence is a new technological science that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. Currently, the integration of artificial intelligence and geothermal energy development has entered a stage of in-depth practice. The core objective is to solve three major pain points in traditional development through technological breakthroughs: geological complexity, difficulty in dynamic prediction, and decision-making lag. Existing technologies generally use sensor networks to collect data such as temperature, pressure, and flow rate in real time. Combined with remote sensing data and geological exploration data, a multi-dimensional dataset is constructed. By standardizing or normalizing the units of measurement to unify the units of measurement, the impact of dimensional differences on model performance is eliminated. Data processing is then performed using algorithms such as random forests to analyze the relationship between geological parameters and heat production efficiency.

[0003] However, existing technologies have some problems. They are prone to ignoring the limitations of physical laws when performing model prediction processing, and they are also prone to oversimplification that fails to capture complex coupling effects. Furthermore, the long time required for a single simulation makes it difficult to support large-scale strategy exploration, leading to the problem of overpressured geothermal energy extraction due to local optima. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a development decision-making management method and system for geothermal energy. It solves the technical problems of existing technologies, which tend to overlook the limitations of physical laws, fail to capture complex coupling effects, and struggle to support large-scale strategy exploration, leading to overpressure extraction of geothermal energy due to local optima.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a development decision-making and management method for geothermal energy, comprising the following steps: collecting multi-source geothermal data from the geothermal system; obtaining a geothermal correlation database through preprocessing and integration; performing spatial analysis of reservoir parameters based on the geothermal correlation data to obtain an optimized reservoir parameter spatial distribution map; conducting dynamic simulation of the optimized reservoir parameter spatial distribution map to obtain an evolution prediction report; obtaining an optimal solution set, multi-dimensional data, and risk assessment results based on the evolution prediction report through multi-objective collaborative optimization and risk assessment; and performing interactive processing and early warning monitoring based on the optimal solution set, multi-dimensional data, and risk assessment results to obtain an interactive chart set and early warning information.

[0006] Preferably, geothermal multi-source data is collected from the geothermal system, and preprocessed and integrated, including: cleaning the geothermal multi-source data to obtain cleaned data; obtaining spatiotemporally aligned and standardized data by registration and standardization based on the cleaned data; and storing the spatiotemporally aligned and standardized data in association to obtain a geothermal association database.

[0007] Preferably, a training sample set is generated by sample pairing based on geothermal correlation data; a convolutional layer is constructed based on the training sample set, and a spatial depth feature map is obtained through forward propagation of the convolutional layer; parameter prediction and spatial kriging interpolation optimization are performed on the spatial depth feature map to obtain an optimized reservoir parameter spatial distribution map set.

[0008] Preferably, the spatial depth feature map is optimized by parameter prediction and spatial kriging interpolation, including: constructing a splicing layer, splicing data with the same regional location in the geothermal correlation data to obtain geological and geothermal splicing features; constructing a fusion layer, fusing features based on the geological and geothermal splicing features and the spatial depth feature map to obtain a fused feature vector; constructing a prediction layer, performing limit gradient processing on the fused feature vector to obtain preliminary predicted values ​​of reservoir parameters; obtaining measured geothermal values, constructing a model optimization layer, and performing spatial kriging interpolation optimization based on the preliminary predicted values ​​of reservoir parameters to obtain an optimized spatial distribution map of reservoir parameters.

[0009] Preferably, dynamic simulation of mining is performed on the optimized reservoir parameter spatial distribution map, including: performing uncertainty sampling and coupled simulation based on the optimized reservoir parameter spatial distribution map to obtain a dynamic simulation dataset; performing physical sensing neural processing on the dynamic simulation dataset to obtain a physical sensing simulator; obtaining a reservoir network model in the optimized reservoir parameter spatial distribution map, obtaining geothermal-related constraints from geothermal multi-source data, and obtaining an evolution prediction report through reward processing and model linkage simulation.

[0010] Preferably, the evolution prediction report is obtained through reward processing and model linkage extrapolation, including: spatial reward processing of the reservoir network model and geothermal-related constraints to obtain enhanced environmental conditions; constructing a DQN policy network based on the physical perception simulator and the enhanced environmental conditions to obtain an optimized policy model; and performing extrapolation analysis based on the optimized policy model and the physical perception simulator to obtain the evolution prediction report.

[0011] Preferably, the process based on evolutionary prediction reports involves multi-objective collaborative optimization and risk assessment, including: constructing a multi-objective optimization function based on the evolutionary prediction report; obtaining a non-dominated solution set through Pareto front search based on the multi-objective optimization function; acquiring historical risk data and risk logic for geothermal development, constructing a Bayesian risk assessment network, and obtaining a risk assessment model; inputting the non-dominated solution set into the risk assessment model to obtain a risk probability solution set; and performing comprehensive risk ranking on the risk probability solution set to obtain the optimal solution set, multi-dimensional data, and risk assessment results.

[0012] Preferably, interactive processing and early warning monitoring are performed based on the optimal solution set, multidimensional data, and risk assessment results, including: data mapping of the optimal solution set, multidimensional data, and risk assessment results to obtain an interactive chart set; calculation of real-time deviation rate and deviation threshold based on data deviation of the interactive chart set; obtaining early warning information based on the real-time deviation rate and deviation threshold, and sending the early warning information to the geothermal management center.

[0013] This technical solution also provides a system applied to the aforementioned geothermal energy development decision-making and management method. The system includes: a database module for collecting multi-source geothermal data from the geothermal system and obtaining a geothermal correlation database through preprocessing and integration; a reservoir parameter module for performing spatial analysis of reservoir parameters based on the geothermal correlation data to obtain an optimized reservoir parameter spatial distribution map; a dynamic simulation module for performing dynamic simulations of the optimized reservoir parameter spatial distribution map to obtain an evolution prediction report; a risk assessment module for obtaining the optimal solution set, multi-dimensional data, and risk assessment results based on the evolution prediction report through multi-objective collaborative optimization and risk assessment; and an early warning monitoring module for interactive processing and early warning monitoring based on the optimal solution set, multi-dimensional data, and risk assessment results to obtain an interactive chart set and early warning information.

[0014] By means of the above technical solution, the present invention provides a development decision management method and system for geothermal energy, which has at least the following beneficial effects: 1. The present invention innovatively combines convolutional neural networks and extreme gradient boosting trees for comprehensive processing. First, deep spatial features are automatically extracted from geophysical data. Then, the deep spatial features are spliced ​​with traditional geological and geothermal features as input to the extreme gradient boosting tree. This method has both powerful spatial feature learning capabilities and the excellent structured data processing and prediction accuracy of the extreme gradient boosting tree, realizing continuous and high-precision prediction of reservoir parameters, significantly improving the accuracy and spatial resolution of prediction, and reducing exploration uncertainty.

[0015] 2. This invention utilizes parameter uncertainty sampling to generate a high-fidelity dataset covering multiple scenarios. The constructed physical perception simulator, while maintaining the coupling with physical laws, reduces the time of a single simulation from several hours to seconds, providing an efficient environment for reinforcement learning. Through the connection and interaction between DQN and the physical perception simulator, the development strategy is dynamically optimized under physical constraints, avoiding problems such as the model getting stuck in local optima or violating physical constraints and resulting in over-pressure mining.

[0016] 3. This invention combines a non-dominated sorting genetic algorithm with a Bayesian network. The genetic algorithm efficiently searches in a vast solution space, directly optimizing multiple objectives and generating a non-dominated solution set. Meanwhile, a risk assessment model based on a Bayesian network evaluates the robustness of each solution under multiple risks in real time. This combination significantly improves the accuracy of risk quantification and enhances the scientific nature, comprehensiveness, and risk resistance of decision-making.

[0017] 4. This invention uses tools such as parallel coordinate graphs and radar charts to intuitively present the performance risk trade-off relationship of the solution, breaking through the limitations of traditional two-dimensional tables. Real-time deviation calculation can accurately capture the dynamic deviation trend of indicators such as instantaneous heat generation and pressure, significantly improving the response efficiency of operation and maintenance management. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 is a flowchart of a development decision management method for geothermal energy according to the present invention; Figure 2 is a structural block diagram of a development decision management system for geothermal energy according to the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0020] Because existing technologies tend to overlook the limitations of physical laws, they cannot capture complex coupling effects and are difficult to support large-scale strategy exploration. This leads to technical problems such as overpressure extraction of geothermal energy due to local optima. Referring to Figure 1, this embodiment provides a development decision-making and management method for geothermal energy that can capture complex coupling effects based on the limitations of physical laws, significantly shorten simulation time, and prevent problems such as local optima and overpressure extraction. The steps of this method are as follows: S1. Collect geothermal multi-source data from the geothermal system, and obtain a geothermal related database through preprocessing and integration. In existing technologies, geothermal data is difficult to use directly due to its diverse sources and different formats. To solve the above problems, the specific implementation steps are as follows: S11. Clean the geothermal multi-source data to obtain cleaned data. Geothermal multi-source data includes geological structure data, geophysical exploration data, geothermal gradient data, historical drilling data, surface environment data, and geothermal-related constraints, such as maximum number of wells and injection-production pressure limits. In this step, the 3σ principle in statistics is first adopted, that is, by calculating the mean and standard deviation of each data feature, the reasonable range of normal data distribution is determined, that is, the mean plus or minus three. Data within a range equal to or exceeding one standard deviation is considered outlier and removed. Simultaneously, a box plot method is used. By calculating the upper quartile, median, lower quartile, and the difference between the interquartile ranges of the upper and lower quartiles, data within the box plot's bend range are identified. Typically, this range is between the lower quartile minus 1.5 times the interquartile range and the upper quartile plus 1.5 times the interquartile range. Data outside this range are considered outliers and removed. For example, for a geothermal gradient dataset with a mean of 30°C per 100 meters and a standard deviation of 5°C per 100 meters, then according to... The 3σ principle states that the normal data range is 15°C to 45°C per 100 meters, i.e., 30°C plus or minus 3 times 5°C. Geothermal gradient values ​​outside this range will be removed. Furthermore, if the lower quartile of the dataset is 25°C per 100 meters, the upper quartile is 35°C per 100 meters, and the interquartile range is 10°C per 100 meters, then the box plot must cover a range of 10°C to 50°C per 100 meters, i.e., 25°C minus 1.5 times 10°C to 35°C plus 1.5 times 10°C. Geothermal gradient values ​​outside this range will also be identified as anomalies and removed. The 3σ principle is a commonly used statistical method in statistical data and will not be elaborated upon here.

[0021] S12. Based on the cleaned data, spatiotemporally aligned standardized data is obtained through registration and standardization. In this step, georegistration is performed first. This step aims to unify geospatial data from different sources and coordinate systems into the same spatiotemporal coordinate system, ensuring that all data are accurately aligned in geospatial location and eliminating data misalignment caused by coordinate differences. After georegistration is completed, each feature in the dataset is then standardized using the Z-Score standardization method. The specific steps include first calculating the average of all data values ​​for the feature, then calculating its standard deviation, then subtracting the previously calculated average from each original data value of the feature to obtain the difference, and then dividing this difference by the standard deviation. That is, multiplying the result of subtracting the average from the original data value by the reciprocal of the standard deviation, which is equivalent to dividing by the standard deviation, and finally obtaining the standardized value.

[0022] S13. The spatiotemporally aligned and standardized data is associated and stored to obtain a geothermal association database. This step first matches data records with the same spatial location identifier from different data sources, much like finding pieces of a jigsaw puzzle that correspond to the same location and piecing them together. Through table joins, related data that were originally scattered across different data tables but belong to the same spatial location are integrated into a single record. Although no calculation formula is involved here, from the perspective of data integration logic, it can be understood as adding data items corresponding to the same spatial location from different data sources, that is, merging various related data into one place, ultimately obtaining a complete data record containing rich information. This integrated data is then stored in the geothermal association database. Table joins can be performed using a full join, a common data integration method, which will not be elaborated upon here. This invention constructs a unified, high-quality database through spatiotemporal alignment and standardization processing, providing reliable input data for subsequent algorithm models.

[0023] S2. Perform spatial analysis of reservoir parameters based on geothermal correlation data to obtain an optimized spatial distribution map of reservoir parameters. Existing technologies typically rely on single geophysical inversion or statistical interpolation. Single geophysical inversion solutions are not unique and have high uncertainty, while statistical interpolation struggles to capture complex nonlinear relationships. To address these issues, the specific implementation steps are as follows: S21. Generate a training sample set based on geothermal correlation data through sample pairing. In this step, the overall data is first divided into training, validation, and training sets according to certain rules, including factors such as well location, formation depth range, or data acquisition time. The dataset is divided into different parts, such as the test set. For each well location, the geophysical attribute data, such as seismic wave velocity and resistivity, is combined with the geological attribute data, such as lithology and porosity. This can be understood as generating a comprehensive feature vector in the form of superposition of these different types of feature data, which represents the characteristics of the well location. The measured reservoir parameters corresponding to the well location, such as permeability and oil saturation, are used as the paired labels. Finally, each set of feature vectors and corresponding labels are combined into a sample pair, and all sample pairs are collected into a training sample set.

[0024] S22. Convolutional layers are constructed based on the training sample set. Spatial depth feature maps are obtained through the forward propagation of the convolutional layers. The data in the training sample set is organized into a data grid suitable for processing by the CNN model, such as a grid. The convolutional layer is the pre-trained CNN model, which contains structures such as convolutional layers and pooling layers. During the forward propagation of the model, taking the convolutional layer as an example, the convolutional kernel in the convolutional layer slides on the data grid with a certain stride. During the sliding process, the convolutional kernel performs multiplication operations with the data at the corresponding positions on the data grid. Then, the results of all multiplication operations are added together to obtain a convolutional value. This convolutional value represents the local depth feature map extracted after the convolution operation. Different convolutional kernels extract different local features. For example, some convolutional kernels may be better at capturing edge information in geological structures, while others may be more sensitive to geological texture features, such as the Sobel convolutional kernel. Pooling layers mainly play the role of dimensionality reduction and feature selection. They perform downsampling operations on the feature maps output by the convolutional layers. For example, max pooling selects the maximum value in each local region as the representative value of that region. This can reduce the amount of data while retaining the most significant feature information. After multiple layers of convolution and pooling operations, the original data grid is gradually transformed into a set of more representative and abstract feature maps. These feature maps are the final extracted spatial depth feature maps.

[0025] S23. Perform parameter prediction and spatial kriging interpolation optimization on the spatial depth feature map to obtain an optimized reservoir parameter spatial distribution map set; S231. Construct a splicing layer, splicing data with the same regional location in the geothermal correlation data to obtain geological and geothermal splicing features; In this step, according to a certain analysis method, such as rock sample analysis to obtain geological attributes such as lithology and mineral composition, the corresponding geothermal correlation data is collected at each measurement point. The measurement point is the location of the geothermal well where the geothermal correlation data is obtained. Geological attribute data such as lithology, porosity, and permeability, as well as geothermal data such as temperature values ​​at different underground depths, are collected and combined to obtain the geological and geothermal splicing features of the data points in the same regional location.

[0026] S232. Construct a fusion layer. Based on the geological and geothermal splicing features and the spatial depth feature map, feature fusion is performed to obtain a fused feature vector. In this step, feature information is first aggregated from the spatial neighborhood of a certain location in the spatial depth feature map. Here, feature aggregation can be understood as comprehensively considering the feature data of multiple locations in the neighborhood. For example, during aggregation, a weighted summation method can be used to assign different weights to the features of different locations in the neighborhood. The weights can be determined based on factors such as the distance from the predicted location. The closer the distance, the greater the weight may be. The features of each location in the neighborhood are multiplied by their corresponding weights and then added together to obtain the aggregated features. Then, the aggregated features are spliced ​​with the geological and geothermal splicing features. The final fused feature vector integrates the spatial depth features and the point-like geological and geothermal splicing features.

[0027] S233. Construct a prediction layer and perform extreme gradient processing on the fused feature vector to obtain preliminary predicted values ​​of reservoir parameters. Obtain labels from the training sample set, which contain reservoir parameter data for geothermal well points. First, train an XGBoost regression model using the fused feature vectors and corresponding labels of known well points. The XGBoost regression model consists of multiple regression trees. During training, the model continuously learns the relationship between the fused feature vector and the reservoir parameter labels. Specifically, each regression tree judges and calculates based on the input fused feature vector. By comparing and analyzing different feature values, the data is divided into different branch nodes, eventually reaching the leaf nodes. Each leaf node outputs a value. During prediction, for any input fused feature vector, the model allows all regression trees to process it separately. Each regression tree provides a predicted value. Finally, the predicted values ​​of all regression trees are added together to obtain the model's prediction of the reservoir parameters corresponding to the fused feature vector. For example, assuming there are 3 regression trees in the model, for a certain fused feature vector, the first regression tree gives a predicted value of 2, the second is 3, and the third is 4. Then the final predicted value of the model is 2+3+4=9. At the same time, the trained prediction layer is obtained. The preliminary predicted values ​​of reservoir parameters include three categories: rock physical properties, fluid occurrence state, and geological structural characteristics. Rock physical properties include the three-dimensional distribution of porosity obtained by micro-CT scanning or nuclear magnetic resonance technology, the anisotropic permeability measured by core flow experiments, and the elastic modulus and Poisson's ratio obtained based on rock mechanics tests. Fluid parameters include the crude oil viscosity-temperature curve measured by high-pressure physical property experiments, the spatial variation characteristics of formation water salinity, and the dynamic changes of oil saturation obtained by well logging interpretation and production data analysis. Geological structural parameters integrate the fault sealing of three-dimensional seismic interpretation, the sand body connectivity controlled by sedimentary microfacies, and the quantitative indicators of reservoir heterogeneity based on geological modeling. These data are also reflected in the training sample set.

[0028] S234. Obtain geothermal measured values, construct a model optimization layer, and perform spatial kriging interpolation optimization based on the preliminary predicted values ​​of reservoir parameters to obtain an optimized spatial distribution map of reservoir parameters. In this step, the preliminary predicted values ​​of reservoir parameters are input to provide an initial prediction result covering the entire study area, reflecting the approximate spatial distribution trend of reservoir parameters. Obtaining the geothermal measured values ​​for each geothermal well point provides accurate actual measurement data, which is crucial for verifying and optimizing the prediction results. Geological structural boundary data can be obtained through training sample sets combined with geological exploration methods, providing important geological constraints for subsequent interpolation processing. In terms of processing methods, the preliminary predicted values ​​of reservoir parameters are used as the trend surface. Then, within each geological zone divided by the geological structural boundary data, spatial kriging interpolation optimization is employed. Spatial kriging interpolation optimization comprehensively considers the relationship between known well point measured values ​​and the trend surface, combining the precise information of the measured values ​​with the overall trend of the trend surface through a specific weight allocation mechanism. The detailed steps are as follows: During the interpolation process, for... For each interpolation point, a weighting coefficient is calculated based on its spatial distance from known well points and its similarity to geological structural features. For example, the weighting coefficient is calculated by multiplying the spatial distance of a single well point by its similarity and then summing the spatial distances of all well points by their similarities. This weighting coefficient determines the degree of influence of the measured values ​​of known well points on the predicted value of that point. Ultimately, the predicted value of that point can be seen as the result of a weighted sum of the trend surface value at that point and the measured values ​​of known well points according to the weighting coefficient. The results of all well points are then integrated into an optimized reservoir parameter spatial distribution map. This invention innovatively combines convolutional neural networks and extreme gradient boosting trees for comprehensive processing. First, deep spatial features are automatically extracted from geophysical data. Then, these deep spatial features are concatenated with traditional geological and geothermal features as input to the extreme gradient boosting tree. This method combines powerful spatial feature learning capabilities with the excellent structured data processing and prediction accuracy of extreme gradient boosting trees, achieving continuous and high-precision prediction of reservoir parameters. This significantly improves the accuracy and spatial resolution of predictions and reduces exploration uncertainty.

[0029] S3. Perform dynamic simulation of the exploitation process using the optimized reservoir parameter spatial distribution map and obtain an evolution prediction report. Existing technologies tend to overlook the limitations of physical laws, are prone to oversimplification and failure to capture complex coupling effects, and are difficult to support large-scale strategy exploration due to the long simulation time per run. This can lead to geothermal energy overpressure exploitation problems due to local optima. To solve these problems, the specific implementation steps are as follows: S31. Perform uncertainty sampling and coupled simulation based on the optimized reservoir parameter spatial distribution map to obtain a dynamic simulation dataset. Since actual reservoir parameters have uncertainties, such as permeability fluctuating within a certain range, we use parameter uncertainty sampling methods to extract multiple representative parameter sets from the optimized reservoir parameter spatial distribution map. Numerical combinations are employed. For example, for each key parameter, such as permeability and porosity, within its known reasonable range, multiple values ​​are selected according to certain sampling rules, such as uniform sampling or normal sampling. Then, the selected values ​​of different parameters are combined to form multiple different parameter sets. For each parameter set, a small number of coupled numerical simulations of the three physical quantities of heat, water, and force are run. During the simulation, each parameter set is used as the initial condition of the reservoir and input into the numerical simulation model. At the same time, different development schemes are set, such as different injection and production rates and well layouts. The numerical simulation model comprehensively considers the interaction between heat conduction, fluid flow, and rock mechanical deformation in the reservoir. Through complex mathematical calculations and physical process simulations, such as the TOUGH FLAC model, an open-source numerical simulation platform developed by Lawrence Berkeley National Laboratory in the United States, it predicts the production dynamics of the reservoir under specific parameter sets and development schemes, including indicators such as well production, pressure changes, and water cut. Finally, we combine each parameter set, the corresponding development scheme, and the simulated production dynamic data to form a dynamic simulation dataset.

[0030] S32. Perform physical perception neural processing on the dynamic simulation dataset to obtain a physical perception simulator. In this step, a physical perception neural network, namely PINN, is first built. The parameter set and development scheme in the dynamic simulation dataset are used as input features to generate dynamic indicators, such as oil well production, pressure, and water cut, as potential output targets. During the training phase, the accuracy of prediction is measured not only by calculating the mean square error between the network's predicted values ​​and the real simulation data, i.e., the data fitting loss, but also by constructing control equations based on physical laws such as energy conservation and mass conservation. The residuals of these equations are calculated within the computational domain, i.e., the physical constraint loss. The data fitting loss and the physical constraint loss are added together by weighted coefficients to form a total loss function to guide the network parameter update. Through continuous iterative optimization, the total loss is minimized, and finally a well-trained PINN proxy simulator is obtained. It can accurately fit the existing simulation data and perform generalized predictions in accordance with physical laws.

[0031] S33. Obtain the reservoir network model from the optimized reservoir parameter spatial distribution map set, obtain geothermal-related constraints from geothermal multi-source data, and obtain an evolution prediction report through reward processing and model linkage inference; the optimized reservoir parameter spatial distribution map set predicts and generates numerical spatial distributions of parameters such as permeability and temperature. These data are stored and managed in the system in a discrete three-dimensional grid, i.e., in the reservoir grid model. Each grid cell contains a set of predicted attribute values. The geothermal-related constraints in the geothermal multi-source data are the physical constraints initially set based on the geological characteristics of the geothermal well points.

[0032] S331. Spatial reward processing is applied to the reservoir network model and geothermal-related constraints to obtain enhanced environmental conditions. In this step, key physical field information, such as pressure and temperature values ​​of each grid block, is extracted based on the reservoir grid model. Combined with geothermal-related constraints, such as the maximum number of wells and injection-production pressure limits, an action space is defined. For example, the action of opening or closing wells is limited to not exceeding the maximum number of wells, and the adjustment range of flow rate adjustments is constrained within the injection-production pressure limits. Simultaneously, the current pressure field, temperature field, and other physical quantities are combined into a state vector as the observation input for the agent. The reward function design integrates development goals and constraint satisfaction: the cumulative power generation is used as the basic reward item; if the action causes pressure or temperature to deviate from the safe level, a reward is given. For the entire range, the total reward is adjusted by multiplying by a penalty coefficient, such as multiplying the reward by 0.8 when the pressure exceeds the limit, or by adding a negative reward, such as deducting 5 points for every 1 MPa of pressure exceeding the limit. This results in a comprehensive reward value that balances efficiency and safety. For example, in the development of an oil reservoir, if the current status shows that the pressure in a certain area is close to the upper limit, and the agent chooses to reduce the water injection volume in that area, the system will first check whether the adjusted pressure still meets the constraints. If it is within a reasonable range, the system will calculate the basic reward, such as the score corresponding to the daily oil production, and then add extra points for the high pressure maintenance. If the pressure exceeds the limit after adjustment, the basic reward will be multiplied by a discount coefficient and the overpressure penalty points will be deducted. The final output includes the enhanced environmental conditions that include the status, the legality of the action, and the comprehensive reward.

[0033] S332. Construct a DQN policy network based on the physical perception simulator and enhanced environmental conditions to obtain an optimized policy model. This step first constructs an initial DQN policy network. In each training round, the initial DQN policy network generates an action probability distribution based on the current state, such as reservoir pressure distribution and well location status. Actions are selected according to probability, such as adjusting the water injection rate of a well. The actions are then input into the PINN physical perception simulator. The simulator, based on physical conservation laws such as mass conservation and energy conservation, advances the reservoir state evolution and returns the new state and corresponding reward value. For example, if the action increases oil production without exceeding pressure constraints, the reward is the base oil production gain multiplied by the efficiency improvement coefficient; if overpressure occurs, the reward is the base profit multiplied by the safety discount coefficient, and an overpressure penalty is deducted. The initial DQN policy network... The network compares the error between the Q-value of the current state action pair and the target Q-value. The Q-value is predicted by the neural network, and the target Q-value is calculated by adding the maximum predicted Q-value of the new state to the reward returned by the simulator. Gradient descent is used to update the network parameters, gradually reducing the prediction error. After multiple iterations, the initial DQN strategy network can directly output the optimal action based on the input state, such as the optimal injection and production scheme, achieving a dynamic balance between development benefits and physical constraints. For example, in the development of a low-permeability oil reservoir, the initial DQN strategy network caused local pressure exceedance due to excessive water injection. By reducing the probability of this action through a penalty mechanism, a segmented water injection combined with a periodic shut-in strategy was explored. The final trained strategy network reduced the pressure exceedance time by 60% while maintaining stable daily oil production, and the optimized initial DQN strategy network was used as the optimized strategy model.

[0034] S333. Based on the optimization strategy model and the physical sensing simulator, an evolution prediction report is obtained through inference and analysis. The dynamic simulation dataset contains initial development schemes, which include different well layouts, such as combinations of horizontal and vertical wells, injection and production strategies, such as comparisons between continuous and periodic water injection, or development scales, such as phased development and one-time development. These schemes must meet geothermal-related constraints. During the inference, for schemes requiring intelligent optimization, the optimization strategy model outputs the optimal action based on the current state, such as the initial pressure field and temperature field. The optimal action includes adjusting the injection and production rate of a certain well. The physical sensing simulator advances the state evolution based on physical conservation laws, such as the heat convection and conduction equation, and records the production at each step, such as daily oil production or power generation and temperature field data. For fixed strategy schemes, the physical sensing simulator is run directly for prediction. Finally, the production at all time steps under each scheme is accumulated to obtain the cumulative production. At the same time, the temperature field distribution at key time nodes is extracted and integrated into a system containing production curves, temperature curves, and temperature field data. This invention provides an evolution prediction report with temperature field evolution diagrams and comparative analysis. For example, in the development of a geothermal field, the initial scheme set includes two modes: single-well circulation development and multi-well linkage development. The multi-well linkage scheme guided by DQN dynamically adjusts the injection and production rates, extending the thermal breakthrough time from 8 years to 12 years under a fixed strategy, and increasing the cumulative power generation by 35%. In contrast, the single-well circulation scheme under a fixed strategy experiences a sharp drop in power generation after the 5th year due to rapid cooling of the thermal reservoir. The final report provides quantitative basis for decision-makers by comparing the production curves and temperature field evolution of the two schemes, and forms an evolution prediction report. This invention utilizes parameter uncertainty sampling to generate a high-fidelity dataset covering multiple scenarios. The constructed physical perception simulator, while maintaining the coupling of physical laws, compresses the simulation time from several hours to seconds, providing an efficient environment for reinforcement learning. Through the connection and interaction between DQN and the physical perception simulator, the development strategy is dynamically optimized under physical constraints, avoiding problems such as the model getting trapped in local optima or violating physical constraints, resulting in overpressure mining.

[0035] S4. Based on the evolutionary prediction report, the optimal solution set, multidimensional data, and risk assessment results are obtained through multi-objective collaborative optimization and risk assessment. Existing multi-objective optimization technologies rely on weight allocation, are highly subjective, and struggle to handle multiple contradictory issues in the search process. They also lack dynamic modeling capabilities for causal relationships and uncertainties, potentially leading to solutions that deviate from the actual optimal. To address these issues, the specific steps are as follows: S41. Construct a multi-objective optimization function based on the evolutionary prediction report. This step first extracts core indicators from the prediction report, such as cumulative net present value, total heat recovery, formation pressure drop, and temperature maintenance. Then, the objective function is defined according to engineering requirements. For example, maximizing net present value and minimizing formation pressure drop can be combined into a bi-objective optimization problem, or a weighted sum of total heat recovery and temperature maintenance can be introduced as a comprehensive objective. The objective function is often constructed using a linear combination form, for example: comprehensive objective value... The multi-objective optimization function is defined as follows: maximizing the comprehensive objective equals 0.5 times the cumulative net present value, plus 0.3 times the reservoir life coefficient, minus 0.2 times the investment cost. Solving this model, Scheme C, which balances heat recovery and reservoir protection, becomes the optimal solution. Its comprehensive objective value is 12% higher than Scheme A, validating the multi-objective optimization function in complex development decisions. For example, Scheme A, with high-intensity mining, yields 5 million gigajoules of heat recovery over 20 years, but the formation pressure drop reaches 15 MPa, shortening the reservoir life. Scheme B, with moderate mining, yields 4 million gigajoules of heat recovery with a pressure drop of only 8 MPa. Scheme C, with intelligent control mining, dynamically adjusts the injection and production rates, achieving 4.8 million gigajoules of heat recovery while controlling the pressure drop to 10 MPa. The multi-objective optimization function is defined as: maximizing the comprehensive objective equals 0.5 times the heat recovery, plus 0.3 times the reservoir life coefficient, minus 0.2 times the investment cost. By solving this model, Scheme C becomes the optimal solution because it balances heat recovery and reservoir protection, with its comprehensive objective value increasing by 12% compared to Scheme A. This verifies the effectiveness of multi-objective optimization in complex development decisions.

[0036] S42. Based on the multi-objective optimization function, a non-dominated solution set is obtained through Pareto front search. This step uses non-dominated sorting and reference point guidance mechanisms to iteratively generate a candidate solution group within the geothermal-related constraints. In each iteration, the multi-objective function value corresponding to each solution is calculated, and non-dominated solutions are selected based on the dominance relationship, i.e., no other solution is better than all objectives. This gradually approaches the Pareto front, and the final output non-dominated solution set is a set of solutions that cannot be further optimized in the objective space. For example, in the development of a medium-deep geothermal field, the geothermal-related constraints are defined as: well location coordinates (X∈[1000,2000],Y∈[500,1500]), in meters, and injection-production rate Q∈[50,150]m. 3 / d、The development cycle T∈[5,20] years, the multi-objective function is to maximize the comprehensive benefit = 0.5 multiplied by the cumulative heat recovery over 20 years, plus 0.3 multiplied by the reservoir life, minus 0.2 multiplied by the total investment. After 100 iterations using the NSGA-Ⅲ algorithm, a Pareto front containing 15 non-dominated solutions is generated. Solution A contains 4.8 million GJ of heat recovery, 18 years of reservoir life, and 800 million RMB of investment. Solution B contains 4.5 million GJ of heat recovery, 25 years of reservoir life, and 600 million RMB of investment, representing two extreme strategies: efficient development and sustainable development, respectively. Solution C contains 4.65 million GJ of heat recovery, 22 years of reservoir life, and 700 million RMB of investment. Because it balances heat recovery efficiency and reservoir protection, it is selected as the recommended solution. Integrating all the optimized solutions yields the non-dominated solution set.

[0037] S43. Obtain historical risk data and risk logic for geothermal development, construct a Bayesian risk assessment network, and obtain a risk assessment model. Historical risk data typically comes from engineering records of developed geothermal fields, such as the time, location, and consequences of events like wellbore leaks, reservoir collapses, and equipment failures. This data can be obtained through enterprise databases or industry reports. Risk logic is collected using the Delphi method, covering the causal logic between risk factors. For example, excessively high injection-production rates may lead to formation pressure imbalances, resulting in a chain-like logic of wellbore integrity failure. When constructing the Bayesian network, historical risk data is first used as nodes, such as geological uncertainties, equipment aging, and operational violations. Then, directed edges between nodes are determined based on the risk logic. Finally, a conditional probability table is learned using the event frequencies in the historical risk data. For example, in the development of a high-temperature geothermal field, historical data shows that there were 3 wellbore leak events in the past 5 years, 2 of which were related to injection-production rates exceeding 120m³ / h. 3 / d related, 1 case was caused by formation heterogeneity; the risk logic points out that improper pressure management will simultaneously aggravate wellbore corrosion and reservoir collapse; based on this, the constructed Bayesian network contains 4 key nodes: injection-production rate (high or low), formation heterogeneity (strong or weak), pressure management (compliant or non-compliant), and wellbore leakage risk (high or low). Through learning from historical data, conditional probabilities are obtained: if the injection-production rate is high and pressure management is non-compliant, the wellbore leakage risk rises to 70%; if only formation heterogeneity is strong, the risk is 30%. Thus, a risk assessment model capable of risk assessment is obtained. The Delphi method is a structured expert consultation method that integrates expert knowledge through multiple rounds of anonymous feedback mechanisms to ultimately form a collective consensus prediction or decision-making tool, which will not be elaborated here.

[0038] S44. Input the non-dominated solution set into the risk assessment model to obtain the risk probability solution set. In this step, the characteristic parameters of each scheme in the non-dominated solution set, such as injection-production rate, well location distribution, and development cycle, are first input into the risk assessment model as evidence variables. The risk assessment model transmits risk information layer by layer based on the causal relationship between nodes, such as high injection-production rate leading to formation pressure imbalance and subsequent wellbore leakage, and the corresponding probability table of conditions. For example, if a certain scheme adopts an injection-production rate of 150m... 3 / d Furthermore, given the high heterogeneity of the formation, the risk assessment model first calculates the prior probability of formation pressure imbalance, then combines it with the conditional probability of wellbore leakage caused by pressure imbalance, and finally outputs the posterior probability of wellbore leakage risk. After traversing all risk nodes, the model obtains the quantitative probability values ​​of various risks under this scheme, and finally forms a risk probability solution set.

[0039] S45. Perform comprehensive risk ranking on the risk probability solution set to obtain the optimal solution set, multidimensional data, and risk assessment results. This step first constructs a multidimensional assessment system including performance indicators such as thermal extraction efficiency, development cost, and resource utilization rate, and risk indicators such as wellbore leakage probability, reservoir collapse risk, and equipment failure rate. Then, an approximation-to-ideal-solution ranking method is used. This involves first determining the optimal and worst values ​​of each indicator to constitute the ideal and negative ideal solutions, then calculating the closeness of each scheme to the ideal solution, and comprehensively scoring the schemes. A higher closeness indicates a better balance between performance and risk. Finally, the schemes are ranked from highest to lowest score to select the optimal solution that balances efficiency and safety, and its performance indicators and risk assessment results are output simultaneously. Risk assessment results, for example, in the development of a high-temperature geothermal field, four candidate schemes were obtained through preliminary optimization, with the following risk probability solution sets and performance indicators: Scheme A: Thermal extraction efficiency 92%, development cost 120 million yuan, wellbore leakage risk 45%, reservoir collapse risk 30%; Scheme B: Thermal extraction efficiency 85%, development cost 90 million yuan, wellbore leakage risk 15%, reservoir collapse risk 20%; Scheme C: Thermal extraction efficiency 88%, development cost 100 million yuan, wellbore leakage risk 30%, reservoir collapse risk 25%; Scheme D: Thermal extraction efficiency 95%, development cost 150 million yuan, wellbore leakage risk 60%, reservoir collapse risk 40%. Through multi-dimensional evaluation: the ideal solution is determined: thermal extraction... The optimal solutions are: Highest efficiency (95%), lowest cost (0.9 billion yuan), lowest risk (15% leakage, 20% collapse); Proximity calculation: Solution C achieves the best balance between 88% efficiency, 100 million yuan cost, and risks of 30% leakage and 25% collapse, with a proximity score of 0.82; Solution B, while having the lowest risk, has an efficiency of only 85%, resulting in a proximity score of 0.75; Solutions A and D have excessively high risks, with proximity scores of only 0.68 and 0.60 respectively; The final optimal solution set is: Recommended Solution C: 88% thermal efficiency, 100 million yuan cost, 30% wellbore leakage risk, 25% reservoir collapse risk; Alternative Solution B: 85% thermal efficiency, 0.9 billion yuan cost, 15% wellbore leakage risk, 20% reservoir collapse risk. 0%, where the formula for calculating the proximity is the core formula used in the approximation-to-ideal-solution ranking method to measure the comprehensive similarity between the scheme and the ideal solution. The approximation-to-ideal-solution ranking method is a commonly used method that comprehensively ranks candidate schemes by quantifying the geometric distance between the scheme and the ideal solution and the negative ideal solution, which will not be elaborated here. This invention combines a non-dominated ranking genetic algorithm with a Bayesian network. The genetic algorithm efficiently searches in a huge scheme space, directly optimizes multiple objectives, and generates a non-dominated solution set. At the same time, the risk assessment model based on the Bayesian network evaluates the robustness of each scheme under multiple risks in real time. This combination significantly improves the accuracy of risk quantification and enhances the scientific nature, comprehensiveness, and risk resistance of decision-making.

[0040] S5. Based on the optimal solution set, multidimensional data, and risk assessment results, interactive processing and early warning monitoring are performed to obtain an interactive chart set and early warning information. Existing technologies generally use static charts to display solution data, which cannot support interactive exploration or dynamic updates, and the charts are often monotonous and lack comprehensive display of multiple chart types. To solve these problems, the specific implementation method is as follows: S51. Data mapping is performed on the optimal solution set, multidimensional data, and risk assessment results to obtain an interactive chart set. First, the data is preprocessed by normalization to eliminate dimensional differences. Then, appropriate graphic dimensions are selected according to the indicator type: for continuous indicators, such as... Thermal efficiency, cost, and risk probability are mapped using parallel coordinate axes to display the distribution of values ​​for each indicator. For multi-dimensional comprehensive comparisons, radar charts are used to project the solutions onto the vertices of regular polygons, and the area of ​​the closed region is used to intuitively reflect the comprehensive performance. Considering the unique geological spatial characteristics of geothermal development, risk probability and performance indicators are superimposed on a three-dimensional geological model, and dynamic coloring with color gradients and transparency is used to achieve a three-dimensional presentation from stratigraphic profiles to well site deployment. Finally, interactive chart sets with multi-granularity visualization are generated through interactive controls, which can be implemented using the TeeChart chart control.

[0041] S52. Calculate the real-time deviation rate and deviation threshold based on the data deviation using the interactive chart set. In this step, first determine the actual monitoring value and the expected benchmark value corresponding to the current or specific time point in the interactive chart set. Subtract the expected value from the actual value to obtain the difference. Then divide the difference by the expected value to eliminate the influence of dimensions. Finally, multiply the result by the percentage conversion factor to obtain the deviation rate presented as a percentage. The deviation threshold can be obtained by processing the mean of the difference and the mean of the expected value based on the historical data of geothermal energy. Other calculation methods are also possible, such as using the threshold commonly used in geothermal energy collection, which will not be elaborated here.

[0042] S53. Based on the real-time deviation rate and deviation threshold, early warning information is obtained and sent to the geothermal management center. The early warning information includes abnormal early warning information and normal information. If the real-time deviation rate is greater than or equal to the deviation threshold, abnormal early warning information is generated; otherwise, normal information is generated. After receiving the abnormal early warning information, the geothermal management center can conduct detection and resolve geothermal anomalies. This invention uses tools such as parallel coordinate graphs and radar charts to intuitively present the performance risk trade-off relationship of the solution, breaking through the limitations of traditional two-dimensional tables. Real-time deviation calculation can accurately capture the dynamic deviation trend of instantaneous heat generation, pressure, and other indicators, significantly improving the response efficiency of operation and maintenance management.

[0043] Please refer to Figure 2, which shows a structural block diagram of a geothermal energy development decision management system provided in this embodiment. The system includes: a database module, a reservoir parameter module, a dynamic simulation module, a risk assessment module, and an early warning monitoring module. The database module is used to collect multi-source geothermal data from the geothermal system and obtain a geothermal correlation database through preprocessing and integration. The reservoir parameter module is used to perform spatial analysis of reservoir parameters based on the geothermal correlation data to obtain an optimized reservoir parameter spatial distribution map. The dynamic simulation module is used to perform dynamic simulation of exploitation based on the optimized reservoir parameter spatial distribution map and obtain an evolution prediction report. The risk assessment module is used to obtain the optimal solution set, multi-dimensional data, and risk assessment results based on the evolution prediction report through multi-objective collaborative optimization and risk assessment. The early warning monitoring module is used to perform interactive processing and early warning monitoring based on the optimal solution set, multi-dimensional data, and risk assessment results to obtain an interactive chart set and early warning information.

[0044] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0045] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A development decision-making and management method for geothermal energy, characterized in that, The method involves the following steps: collecting multi-source geothermal data from the geothermal system; preprocessing and integrating the data to obtain a geothermal correlation database; performing spatial analysis of reservoir parameters based on the geothermal correlation data to obtain an optimized reservoir parameter spatial distribution map; conducting dynamic simulation of the optimized reservoir parameter spatial distribution map to obtain an evolution prediction report; and based on the evolution prediction report, obtaining the optimal solution set, multi-dimensional data, and risk assessment results through multi-objective collaborative optimization and risk assessment. Interactive processing and early warning monitoring are performed based on the optimal solution set, multidimensional data, and risk assessment results to obtain interactive chart sets and early warning information.

2. The development decision-making and management method for geothermal energy according to claim 1, characterized in that, Geothermal multi-source data from the geothermal system is collected, and preprocessed and integrated, including: cleaning the geothermal multi-source data to obtain cleaned data; obtaining spatiotemporally aligned and standardized data from the cleaned data through registration and standardization; and storing the spatiotemporally aligned and standardized data in association to obtain a geothermal association database.

3. The development decision-making and management method for geothermal energy according to claim 1, characterized in that, Spatial analysis of reservoir parameters based on geothermal correlation data includes: generating a training sample set by sample pairing based on geothermal correlation data; constructing a convolutional layer based on the training sample set and obtaining a spatial depth feature map through forward propagation of the convolutional layer; performing parameter prediction and spatial kriging interpolation optimization on the spatial depth feature map to obtain an optimized spatial distribution map set of reservoir parameters.

4. The development decision-making and management method for geothermal energy according to claim 3, characterized in that, The process of parameter prediction and spatial kriging interpolation optimization of the spatial depth feature map includes: constructing a splicing layer by splicing data with the same regional location in the geothermal correlation data to obtain geological and geothermal splicing features; constructing a fusion layer by fusing features based on the geological and geothermal splicing features and the spatial depth feature map to obtain a fused feature vector; constructing a prediction layer by performing limit gradient processing on the fused feature vector to obtain preliminary predicted values ​​of reservoir parameters; and obtaining measured geothermal values ​​and constructing a model optimization layer by performing spatial kriging interpolation optimization based on the preliminary predicted values ​​of reservoir parameters to obtain an optimized spatial distribution map of reservoir parameters.

5. The development decision-making and management method for geothermal energy according to claim 1, characterized in that, Dynamic simulation of exploitation is performed on the optimized reservoir parameter spatial distribution map, including: performing uncertainty sampling and coupled simulation based on the optimized reservoir parameter spatial distribution map to obtain a dynamic simulation dataset; performing physical sensing neural processing on the dynamic simulation dataset to obtain a physical sensing simulator; obtaining the reservoir network model in the optimized reservoir parameter spatial distribution map, obtaining geothermal-related constraints from geothermal multi-source data, and obtaining an evolution prediction report through reward processing and model linkage simulation.

6. The development decision-making and management method for geothermal energy according to claim 5, characterized in that, The evolution prediction report is obtained through reward processing and model linkage extrapolation, including: spatial reward processing of reservoir network model and geothermal related constraints to obtain enhanced environmental conditions; construction of DQN policy network based on physical perception simulator and enhanced environmental conditions to obtain optimized policy model; and extrapolation analysis based on optimized policy model and physical perception simulator to obtain evolution prediction report.

7. The development decision-making and management method for geothermal energy according to claim 1, characterized in that, Based on the evolutionary prediction report, multi-objective collaborative optimization and risk assessment are conducted, including: constructing a multi-objective optimization function based on the evolutionary prediction report; obtaining the non-dominated solution set through Pareto front search based on the multi-objective optimization function; acquiring historical risk data and risk logic for geothermal development, constructing a Bayesian risk assessment network, and obtaining a risk assessment model; inputting the non-dominated solution set into the risk assessment model to obtain a risk probability solution set; and performing comprehensive risk ranking on the risk probability solution set to obtain the optimal solution set, multi-dimensional data, and risk assessment results.

8. The development decision-making and management method for geothermal energy according to claim 1, characterized in that, The system performs interactive processing and early warning monitoring based on the optimal solution set, multidimensional data, and risk assessment results. This includes: mapping the optimal solution set, multidimensional data, and risk assessment results to obtain an interactive chart set; calculating the real-time deviation rate and deviation threshold based on the data deviation from the interactive chart set; and generating early warning information based on the real-time deviation rate and deviation threshold, which is then sent to the geothermal management center.

9. A system applied to the development decision-making and management method for geothermal energy as described in any one of claims 1-8, characterized in that, The system includes: a database module for collecting multi-source geothermal data from the geothermal system and obtaining a geothermal correlation database through preprocessing and integration; a reservoir parameter module for performing spatial analysis of reservoir parameters based on the geothermal correlation data and obtaining an optimized reservoir parameter spatial distribution map; a dynamic simulation module for performing dynamic simulation of exploitation based on the optimized reservoir parameter spatial distribution map and obtaining an evolution prediction report; a risk assessment module for obtaining the optimal solution set, multi-dimensional data, and risk assessment results based on the evolution prediction report through multi-objective collaborative optimization and risk assessment; and an early warning monitoring module for interactive processing and early warning monitoring based on the optimal solution set, multi-dimensional data, and risk assessment results, obtaining an interactive chart set and early warning information.

Images