Enhanced geothermal system fluid heat transfer process prediction method and proxy model based on ann network

Abstract

The present application relates to the technical field of geothermal research, and discloses an enhanced hot dry rock geothermal system fluid heat transfer process prediction method based on an ANN network, comprising the following steps: constructing an enhanced hot dry rock geothermal system fluid heat transfer process prediction proxy model; taking fracture network parameters as inputs of the prediction model; and outputting a temperature breakthrough curve from the prediction model to predict heat transfer behavior in a complex geothermal system. The present application not only solves the efficiency bottleneck problem in numerical simulation, but also provides an intelligent and economically efficient analysis tool for the development of enhanced geothermal systems. This innovative method helps to accelerate the popularization and application of geothermal energy development technology and promote the technological progress and industry development of renewable energy utilization.

Description

Technical Field

[0001] This invention relates to the field of geothermal research technology, and in particular to an enhanced method and surrogate model for predicting fluid heat transfer processes in dry hot rock geothermal systems based on ANN networks. Background Technology

[0002] Enhanced Geothermal Systems (EGS) have broad application prospects as a renewable energy utilization method. In EGS development, hot dry rock (HDR) serves as the main carrier of geothermal resources. Hot water is extracted by injecting cold water to exchange heat energy with the rock mass, thus realizing energy development. However, the fluid flow and heat transfer processes in hot dry rock geothermal systems are affected by complex geological conditions and flow field characteristics, and their dynamic characteristics require detailed analysis through numerical simulation methods.

[0003] Currently, numerical simulation technology is the primary tool for studying fluid flow and heat transfer processes in enhanced geothermal systems. By instantaneously solving the governing equations for heat transfer and fluid flow, the flow path of cold water in the fracture network, temperature changes, and its thermal impact on the surrounding rock mass can be obtained. However, these models typically require high-resolution spatial discretization and time-step calculations to capture the transient characteristics of heat transfer and cold propagation. This meticulous calculation process is not only time-consuming but also requires substantial computational resources.

[0004] In practical applications, optimizing well network layout, designing injection and production strategies, and evaluating system thermal performance often require repeated runs of numerical models. For example, in inversion analysis or optimization design, it may be necessary to run the simulation model thousands or even tens of thousands of times to cover different combinations of input parameters. This repetitive computational pattern of calling numerical models significantly increases computation time and resource consumption, becoming a bottleneck problem in EGS development.

[0005] To address the aforementioned issues, surrogate modeling techniques have gradually emerged as an effective alternative to numerical simulation in recent years. Surrogate models, by learning and fitting the input-output relationship of numerical simulation models, can provide sufficiently accurate predictions while significantly reducing computational costs. Artificial Neural Networks (ANNs), as the primary method for constructing surrogate models, have been widely applied to the modeling and optimization of complex systems due to their strong nonlinear fitting ability, good self-learning ability, and wide adaptability. However, in existing research, the development of ANN-based surrogate models for the flow and heat transfer processes of the EGS dry hot rock system is still in its early stages, lacking efficient surrogate models specifically designed for the multi-parameter and highly nonlinear characteristics of instantaneous heat transfer. 。 Summary of the Invention

[0006] To address the technical problems mentioned in the background section, this invention provides an enhanced method for predicting fluid heat transfer processes in dry hot rock geothermal systems based on an ANN network.

[0007] This invention employs the following technical solution: an enhanced method for predicting fluid heat transfer processes in a dry hot rock geothermal system based on an ANN network, comprising the following steps:

[0008] Step 1: Construct a predictive proxy model for fluid heat transfer processes in an enhanced dry hot rock geothermal system;

[0009] Step 2: Use the fracture network parameters as input to the prediction model;

[0010] Step 3: The prediction model outputs a temperature breakthrough curve to predict the heat transfer behavior in complex geothermal systems.

[0011] The construction of a surrogate model for predicting fluid heat transfer processes in an enhanced hot dry rock geothermal system includes the following steps:

[0012] Fracture network data is obtained through a fracture network generation method. The parameters of the fracture network data include the number of fractures, fracture length, fracture angle, and fracture spacing.

[0013] The flow and heat transfer process of cold water from the injection well through the fracture network to the production well are simulated by numerical modeling. The heat transfer behavior between the injection and production wells is recorded and temperature breakthrough curve data are obtained.

[0014] A training dataset was constructed with fracture network geometric parameters as input data and temperature breakthrough curve data as output data. The dataset was divided into a sample set and a training set. The model was trained to obtain a prediction model for the fluid heat transfer process of a strong dry hot rock geothermal system.

[0015] Optionally, the prediction model training steps are as follows: load the fracture network parameters and BTCs data from the database constructed in step 3;

[0016] Data standardization: Standardize the input parameters and output BTCs data;

[0017] Data set partitioning: divided into training set and validation set;

[0018] Construct a neural network model;

[0019] Training and validation; Training process: Batch training: Batch size is set to 2000, and model parameters are updated in each batch; Training cycle: Set to 400 epochs (the number of times the dataset is fully traversed); Validation set performance is evaluated every 10 epochs to monitor the model's learning progress in real time.

[0020] Early stopping mechanism: Set an early stopping strategy to terminate training when the performance on the validation set no longer improves. Save the parameters of the model with the best performance on the validation set to a file for later use;

[0021] Save the model; save the trained neural network model and its parameters as a binary file (e.g., .pth format); store the model structure and weight file in the file system for easy loading and application later.

[0022] Optionally, the data standardization operation is as follows: use sklearn.preprocessing.StandardScaler to standardize the input parameters and output BTC data, subtracting the mean from each feature and then dividing by the standard deviation.

[0023] Optionally, the training set is divided as follows: the first 12,000 data points are selected from the database for model learning; the validation set consists of 2,000 data points immediately following the training set for model performance evaluation.

[0024] Optionally, a neural network model can be built using the PyTorch deep learning framework, with the following specific structure:

[0025] Input layer: The input consists of parameters of the fracture network, with the specific dimensions matching the number of features in the fracture network;

[0026] Hidden layers: Set up 8 hidden layers with the following node counts: 180, 160, 120, 100, 90, 60, 60, and 20 respectively.

[0027] The activation function used was LeakyReLU (with a negative slope of 0.01) to enhance the nonlinear expressive power of the model.

[0028] Output layer: The output is standardized BTC data, with a dimension equal to the number of discrete points in the BTC data. The output layer typically does not use an activation function; it directly regresses the BTC value.

[0029] Optimizer and Loss Function: Optimizer: Using the Adam algorithm, with a learning rate set to 1×10⁻⁶. -5 Loss function: The mean squared error (MSE) is used to minimize the squared difference between the predicted and actual values.

[0030] Optionally, the fracture network data can be generated using an existing fracture network dataset or a fracture network generation method, and the fracture network generation method includes the following steps:

[0031] Step 1.1: Parameter setting and initialization in MATLAB;

[0032] Step 1.2: Generate the fracture network;

[0033] Step 1.3: Read the fracture network data generated in Step 1.2, store the data in the three-dimensional matrix DataDFN_CNN, and save it as data in DataDFN_CNN.mat format;

[0034] Optionally, the fracture network generation method includes the following steps:

[0035] The operation of step 1.2 is as follows:

[0036] Step a: Crack length generation;

[0037] Step b: Crack angle generation;

[0038] Step c: Random generation of crack locations;

[0039] Step d: Shadow area detection and removal;

[0040] Step e, requirements for crack intersection;

[0041] Step f: Determine the continuity of the fracture network;

[0042] Step g: Save the final generated fracture network data as a matrix of size Nfrac*4, where: Nfrac: represents the number of generated fractures, and each row represents one fracture; 4: represents the four key attributes of each fracture.

[0043] Optionally, the numerical model simulation is constructed using the following method;

[0044] The physical model was established; the fluid flow model was established based on Darcy's law, the cubic law, and the hydraulic coupling between the fracture and the matrix, ensuring the dynamic equilibrium of the fluid in different regions; the heat transfer model was established based on the convection-conduction coupling equation and the heat exchange model between the fracture and the matrix.

[0045] Numerical discretization and solution methods; the discretization methods include discretizing fluid and heat transfer equations using the finite element method (FEM) and employing a time-stepping scheme; the solution process solves transient flow and heat transfer equations by configuring stable iterative algorithms, such as step-by-step iteration or nonlinear solvers, and optimizes the computational efficiency of large-scale numerical models through parallel computing and appropriate convergence criteria.

[0046] Set model parameters and boundary conditions;

[0047] A two-dimensional geothermal reservoir model is established to simulate the fracture network and matrix region in the geothermal system. The typical size of the domain is 200m×200m. Wellbore regions are set on both sides, which are injection wells and production wells, respectively. The wellbore regions are endowed with high permeability and high porosity.

[0048] Boundary conditions:

[0049] The injection well on the left is set with constant hydraulic head and low temperature (20℃) conditions to simulate the cold water injection process.

[0050] The production well on the right is equipped with constant hydraulic head and heat outflow conditions to simulate hot water collection.

[0051] The upper and lower boundaries are set as adiabatic boundaries to avoid heat exchange with the external environment.

[0052] The initial conditions were set to a uniform temperature distribution of 100°C for both the matrix and the fracture network.

[0053] Material properties:

[0054] Fractured regions: high conductivity, low porosity, emphasizing that fractures are the main channels for fluid.

[0055] Matrix region: low conductivity and high porosity ensure that heat conduction is the main mode of heat transfer;

[0056] The thermal properties, such as thermal conductivity, specific heat capacity, and density, are set based on the typical lithological parameters of thermal reservoirs.

[0057] Optionally, the output of the numerical model simulation includes: the dynamic distribution of the temperature field; the temperature breakthrough curves at each monitoring point; and the path and extent of cold water flow in the fracture network.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention can improve computational efficiency: the high computational cost of traditional numerical simulation severely restricts the application of optimization design and inversion analysis, while the prediction model proposed in this invention can complete the same prediction task in a very low running time, which significantly improves computational efficiency.

[0060] The prediction model proposed in this invention supports large-scale optimization: This invention can quickly complete thousands or even tens of thousands of simulation predictions, making system design parameter optimization and sensitivity analysis more feasible, and providing important support for the design and operation of enhanced geothermal systems.

[0061] This invention can reduce resource consumption: It significantly reduces reliance on high-performance computing equipment while saving operating energy and computing time, which helps to promote the economy and sustainability of geothermal development.

[0062] This invention can expand application scenarios: The prediction model proposed in this invention, with its high efficiency and ease of use, is suitable for real-time prediction, site simulation and rapid performance evaluation, providing the possibility for dynamic decision-making in geothermal resource development.

[0063] This invention not only solves the efficiency bottleneck problem in numerical simulation, but also provides an intelligent and cost-effective analysis tool for the development of enhanced geothermal systems. This innovative method helps to accelerate the promotion and application of geothermal energy development technology and promote the technological progress and industry development of renewable energy utilization. Attached Figure Description

[0064] Figure 1 This is a sample file of the fracture network data proposed in this invention;

[0065] Figure 2 This is a dynamic distribution diagram of the temperature field simulated in this invention;

[0066] Figure 3 These are the temperature breakthrough curves for each monitoring point simulated in this invention;

[0067] Figure 4 This is a diagram illustrating the effect of using this model to denormalize the prediction results and restore them to the true BTC value.

[0068] Figure 5 This is a flowchart of the prediction method proposed in this invention. Detailed Implementation

[0069] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0070] Example 1:

[0071] Reference Figure 1-5 This paper proposes an enhanced method for predicting fluid heat transfer processes in hot dry rock geothermal systems based on ANN networks, which includes the following steps:

[0072] Step 1: Construct a predictive proxy model for fluid heat transfer processes in an enhanced dry hot rock geothermal system;

[0073] Step 2: Use the fracture network parameters as input to the prediction model;

[0074] Step 3: The prediction model outputs a temperature breakthrough curve to predict the heat transfer behavior in complex geothermal systems.

[0075] The construction of a surrogate model for predicting fluid heat transfer processes in an enhanced hot dry rock geothermal system includes the following steps:

[0076] Fracture network data is obtained through a fracture network generation method. The parameters of the fracture network data include the number of fractures, fracture length, fracture angle, and fracture spacing.

[0077] The flow and heat transfer process of cold water from the injection well through the fracture network to the production well are simulated by numerical modeling. The heat transfer behavior between the injection and production wells is recorded and temperature breakthrough curve data are obtained.

[0078] Specifically, for each group of fracture networks, temperature change curves at multiple monitoring points were recorded. This included:

[0079] Time series: Temperature distribution at different time points during the simulation;

[0080] Monitoring point setup: Temperature sensors are placed at the outlet of the fracture network to monitor the thermal response during the cold water propulsion process.

[0081] The numerical model outputs multiple temperature breakthrough curves, and the key features of the curves (such as the pre-cooling breakthrough time and cooling rate) directly reflect the geometric and heat transfer characteristics of the fracture network.

[0082] A training dataset was constructed with fracture network geometric parameters as input data and temperature breakthrough curve data as output data. The dataset was divided into a sample set and a training set. The model was trained to obtain a prediction model for the fluid heat transfer process of a strong dry hot rock geothermal system.

[0083] Optionally, the prediction model training steps are as follows: load the fracture network parameters and BTCs data from the database constructed in step 3;

[0084] Data standardization: Standardize the input parameters and output BTCs data;

[0085] Data set partitioning: divided into training set and validation set;

[0086] Construct a neural network model;

[0087] Training and validation; Training process: Batch training: Batch size is set to 2000, and model parameters are updated in each batch; Training cycle: Set to 400 epochs (the number of times the dataset is fully traversed); Validation set performance is evaluated every 10 epochs to monitor the model's learning progress in real time.

[0088] Early stopping mechanism: Set an early stopping strategy to terminate training when the performance on the validation set no longer improves. Save the parameters of the model with the best performance on the validation set to a file for later use;

[0089] Save the model; save the trained neural network model and its parameters as a binary file (e.g., .pth format); store the model structure and weight file in the file system for easy loading and application later.

[0090] Optionally, the data standardization operation is as follows: use sklearn.preprocessing.StandardScaler to standardize the input parameters and output BTC data, subtracting the mean from each feature and then dividing by the standard deviation.

[0091] Optionally, the training set is divided as follows: the first 12,000 data points are selected from the database for model learning; the validation set consists of 2,000 data points immediately following the training set for model performance evaluation.

[0092] Optionally, a neural network model can be built using the PyTorch deep learning framework, with the following specific structure:

[0093] Input layer: The input consists of parameters of the fracture network, with the specific dimensions matching the number of features in the fracture network;

[0094] Hidden layers: Set up 8 hidden layers with the following node counts: 180, 160, 120, 100, 90, 60, 60, and 20 respectively.

[0095] The activation function used was LeakyReLU (with a negative slope of 0.01) to enhance the nonlinear expressive power of the model.

[0096] Output layer: The output is standardized BTC data, with a dimension equal to the number of discrete points in the BTC data. The output layer typically does not use an activation function; it directly regresses the BTC value.

[0097] Optimizer and Loss Function: Optimizer: Using the Adam algorithm, with a learning rate set to 1×10⁻⁶. -5 Loss function: The mean squared error (MSE) is used to minimize the squared difference between the predicted and actual values.

[0098] Alternatively, the fracture network data may be an existing fracture network dataset.

[0099] Optionally, the numerical model simulation is constructed using the following method;

[0100] The physical model was established; the fluid flow model was established based on Darcy's law, the cubic law, and the hydraulic coupling between the fracture and the matrix, ensuring the dynamic equilibrium of the fluid in different regions; the heat transfer model was established based on the convection-conduction coupling equation and the heat exchange model between the fracture and the matrix.

[0101] Numerical discretization and solution methods; the discretization methods include discretizing fluid and heat transfer equations using the finite element method (FEM) and employing a time-stepping scheme; the solution process solves transient flow and heat transfer equations by configuring stable iterative algorithms, such as step-by-step iteration or nonlinear solvers, and optimizes the computational efficiency of large-scale numerical models through parallel computing and appropriate convergence criteria.

[0102] Set model parameters and boundary conditions;

[0103] A two-dimensional geothermal reservoir model is established to simulate the fracture network and matrix region in the geothermal system. The typical size of the domain is 200m×200m. Wellbore regions are set on both sides, which are injection wells and production wells, respectively. The wellbore regions are endowed with high permeability and high porosity.

[0104] Boundary conditions:

[0105] The injection well on the left is set with constant hydraulic head and low temperature (20℃) conditions to simulate the cold water injection process.

[0106] The production well on the right is equipped with constant hydraulic head and heat outflow conditions to simulate hot water collection.

[0107] The upper and lower boundaries are set as adiabatic boundaries to avoid heat exchange with the external environment.

[0108] The initial conditions were set to a uniform temperature distribution of 100°C for both the matrix and the fracture network.

[0109] Material properties:

[0110] Fractured regions: high conductivity, low porosity, emphasizing that fractures are the main channels for fluid.

[0111] Matrix region: low conductivity and high porosity ensure that heat conduction is the main mode of heat transfer;

[0112] The thermal properties, such as thermal conductivity, specific heat capacity, and density, are set based on the typical lithological parameters of thermal reservoirs.

[0113] Optionally, the output of the numerical model simulation includes: the dynamic distribution of the temperature field (see example results). Figure 2 Temperature breakthrough curves at each monitoring point (see BTCs results for examples). Figure 3 ); the path and extent of cold water flow in the fracture network.

[0114] Example 2: In this example, referring to Figure 1 The fracture network can be generated using a fracture network generation method, which includes the following steps:

[0115] Step 1.1: Parameter setting and initialization in MATLAB;

[0116] Step 1.2: Generate the fracture network;

[0117] Step 1.3: Read the fracture network data generated in Step 1.2, store the data in the three-dimensional matrix DataDFN_CNN, and save it as data in DataDFN_CNN.mat format;

[0118] Optionally, the fracture network generation method includes the following steps:

[0119] The operation of step 1.2 is as follows:

[0120] Step a: Crack length generation;

[0121] Step b: Crack angle generation;

[0122] Step c: Random generation of crack locations;

[0123] Step d: Shadow area detection and removal;

[0124] Step e, requirements for crack intersection;

[0125] Step f: Determine the continuity of the fracture network;

[0126] Step g: Save the final generated fracture network data as a matrix of size Nfrac*4, where: Nfrac: represents the number of generated fractures, and each row represents one fracture; 4: represents the four key attributes of each fracture.

[0127] Specifically, the parameter setting and initialization operations are as follows:

[0128] Number of fractures: The number of fractures in each fracture network is randomly generated. The range of the number of fractures can be set according to the requirements, such as between [14, 20].

[0129] Crack angles: Set the crack direction parameters angle1 = -10 and angle2 = 65, representing the two main crack directions. std = 2 is the standard deviation of the crack angles.

[0130] Spatial dimensions: Define the spatial dimensions of the fracture network, xL = 200, yL = 200, and set the minimum fracture length lmin = 50 and the maximum length lbig = 140.

[0131] The process of generating a fracture network includes the following steps.

[0132] Crack length generation

[0133] The crack length is generated based on a statistical model of fractal geometry, and the specific steps are as follows:

[0134] Parameter settings: Set the range between the minimum and maximum crack length values ​​lminl and lbigl. The length of each crack is randomly generated within this range.

[0135] Distribution generation: The distribution of crack length is controlled using a fractional dimension D and a power a. The probability distribution of crack length is calculated using the following formula.

[0136] N fra (L)=c·L -a

[0137] Where L is the length of the fracture, a is the power controlling the distribution of the fracture length, and c is a normalization constant. The resulting lengths follow a power-law distribution, simulating the distribution of fractures of different lengths in nature.

[0138] Random length generation: The length of each fracture is generated using the cumulative distribution function (CDF). The fracture length is obtained by sampling random numbers to ensure the diversity of the fracture network.

[0139] Crack angle generation; the crack angle is generated based on a normal distribution, randomly deviating from two main directions:

[0140] Principal Angles: Set two principal directions, θ1 = -10° and θ2 = 65°.

[0141] Random bias: The actual angle of each fracture is generated using a normal distribution with a standard deviation σ = 2°. Some fractures will randomly deviate from the main direction to simulate directional changes in the fracture network.

[0142] Angle normalization: Ensure that the generated angles are between 0° and 180°, avoiding angles that are too large or too small.

[0143] Random generation of crack locations;

[0144] Random center point generation: The center point of each fracture is randomly generated within the defined domain [0, xL] × [0, yL]. This position follows a uniform distribution.

[0145] Specific region generation: For the first few cracks, a specific y-direction distribution can be set so that their centers are located in a specific region (such as near the center of the domain) to simulate a more ordered crack distribution.

[0146] Shadow area detection and removal;

[0147] After a crack is generated, it is necessary to check whether the newly generated crack encroaches on the shadow area of ​​an existing crack:

[0148] Shadow area generation: Generate a shadow area for each crack, which is usually a certain width (w, 2 units in length) around the crack. This area represents the range of influence of the crack.

[0149] Shadow area detection: Using a geometric intersection algorithm, the system detects whether newly generated cracks overlap with existing crack shadow areas. If overlap occurs, it indicates that the new crack's generation location is inappropriate, and the center position or angle needs to be regenerated.

[0150] Crack intersection requirements;

[0151] The connectivity of a fracture network is achieved by requiring new fractures to intersect with existing fractures:

[0152] Forced intersection: For each newly generated crack, it is required to intersect with at least one existing crack. This process is achieved by detecting the geometric intersection point of two cracks.

[0153] Geometric intersection detection: This algorithm detects whether two cracks intersect. If the intersection requirement is not met, the algorithm will regenerate the center or angle of the crack until the intersection condition is met.

[0154] Determining the continuity of a fractured network;

[0155] The resulting fracture network must be able to extend from the left to the right of the domain:

[0156] Left-right connectivity check: After the split network is generated, the algorithm checks whether there exists a connected path from the left side of the domain (x=0) to the right side (x=xL). Only split networks that satisfy this connectivity are considered valid split networks.

[0157] Continuity determination: If the fracture network does not meet the continuity requirement, the generation process will restart until a continuous fracture network is generated.

[0158] The generated fracture network data will ultimately be saved as a matrix of size Nfrac*4. Wherein:

[0159] Nfrac: Represents the number of generated fractures, with each row representing one fracture. 4: Represents the four key attributes of each fracture: x1: x-coordinate of the fracture's starting point; y1: y-coordinate of the fracture's starting point; x2: x-coordinate of the fracture's ending point; y2: y-coordinate of the fracture's ending point. A total of 9000 such fracture network samples were generated as training data.

[0160] Examples are shown in the table below, for details as follows: Figure 1 As shown.

[0161]

[0162] This invention significantly improves the efficiency of numerical simulation of the heat transfer process in dry hot rock of enhanced geothermal systems by developing a highly efficient surrogate model based on artificial neural networks (ANNs). This surrogate model leverages the powerful nonlinear fitting capabilities of ANNs to learn and approximate the complex relationship between the input parameters and output results of the numerical simulation model, thereby significantly reducing computational costs while maintaining accuracy. Compared to traditional numerical simulation methods that require complex transient solutions, the surrogate model of this invention is much more efficient. A complete numerical simulation that originally took 2 to 5 minutes (the exact time depends on the model size and computational conditions) can be completed in only about 0.001 to 0.01 seconds using the ANN surrogate model. The extremely high efficiency of the surrogate model enables large-scale parameter optimization, well network layout design, and thermal efficiency analysis.

[0163] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting fluid heat transfer processes in an enhanced dry hot rock geothermal system based on an ANN network, characterized in that, Includes the following steps: Step 1: Construct a predictive model for fluid heat transfer processes in an enhanced dry hot rock geothermal system; Step 2: Use the fracture network parameters as input to the prediction model; Step 3: The prediction model outputs a temperature breakthrough curve to predict the heat transfer behavior in complex geothermal systems. Constructing the prediction model includes the following steps: Fracture network data is obtained through a fracture network generation method. The parameters of the fracture network data include the number of fractures, fracture length, fracture angle, and fracture spacing. The flow and heat transfer process of cold water from the injection well through the fracture network to the production well are simulated by numerical modeling. The heat transfer behavior between the injection and production wells is recorded and temperature breakthrough curve data are obtained. A training dataset was constructed with the geometric parameters of the fracture network as input data and the temperature breakthrough curve data as output data. The dataset was divided into a sample set and a training set. The model was trained to obtain a prediction model for the fluid heat transfer process of a strong dry hot rock geothermal system. The training steps for the prediction model are as follows: Load the fracture network parameters and temperature breakthrough curve data from the database; Data standardization: Standardize the input parameters and output temperature breakthrough curve data; Data set partitioning: divided into training set and validation set; Build a neural network model; Training and validation; Training process: Batch training: Batch size is set to 2000, and model parameters are updated in each batch; Training cycle: Set to 400 epochs; Validation set performance is evaluated every 10 epochs, and the learning progress of the model is monitored in real time. Early stopping mechanism: Set an early stopping strategy to terminate training when the performance on the validation set no longer improves; save the parameters of the model with the best performance on the validation set to a file for later use; Save the model; save the trained neural network model and its parameters as a binary file; store the model structure and weight file in the file system for easy loading and application later; The method for generating the fracture network includes the following steps: Step 1.1: Parameter setting and initialization in MATLAB; Step 1.2: Generate the fracture network; Step 1.3: Read the fracture network data generated in Step 1.2, store the data in the three-dimensional matrix DataDFN_CNN, and save it as data in DataDFN_CNN.mat format; The operation of step 1.2 is as follows: Step a: Crack length generation; Step b: Crack angle generation; Step c: Random generation of crack locations; Step d: Shadow area detection and removal; Step e, requirements for crack intersection; Step f: Determine the continuity of the fracture network; Step g: Save the final generated fracture network data as a matrix of size Nfrac*4, where: Nfrac: represents the number of generated fractures, with each row representing one fracture; 4: represents the four key attributes of each fracture. The specific structure of a neural network model built using the PyTorch deep learning framework is as follows: Input layer: The input consists of parameters of the fracture network, with the specific dimensions matching the number of features in the fracture network; Hidden layers: Set up 8 hidden layers with the following node counts: 180, 160, 120, 100, 90, 60, 60, and 20 respectively. The LeakyReLU activation function was chosen to enhance the model's nonlinear expressive power; Output layer: The output is the standardized temperature breakthrough curve data, and the dimension is equal to the number of discrete points of the temperature breakthrough curve data; The output layer typically does not use an activation function and directly regresses the temperature breakout curve value; Optimizer and Loss Function: Optimizer: Using the Adam algorithm, with a learning rate set to 1×10⁻⁶. - 5. Loss function: The mean squared error is used to minimize the squared difference between the predicted and actual values; Numerical model simulations are constructed using the following methods; The physical model was established; the fluid flow model was established based on Darcy's law, the cubic law, and the hydraulic coupling between the fracture and the matrix, ensuring the dynamic equilibrium of the fluid in different regions; the heat transfer model was established based on the convection-conduction coupling equation and the heat exchange model between the fracture and the matrix. Numerical discretization and solution methods; the discretization methods include discretizing fluid and heat transfer equations using the finite element method and employing a time stepping scheme; the solution process uses a stable iterative algorithm, which can be a step-by-step iterative method or a nonlinear solver, to solve transient flow and heat transfer equations, and optimizes the computational efficiency of large-scale numerical models through parallel computing and appropriate convergence criteria. Set model parameters and boundary conditions; A two-dimensional geothermal reservoir model is established to simulate the fracture network and matrix region in the geothermal system. The typical size of the domain is 200m×200m. Wellbore regions are set on both sides, which are injection wells and production wells, respectively. The wellbore regions are endowed with high permeability and high porosity. Boundary conditions: The left injection well is equipped with a constant hydraulic head and low temperature conditions to simulate the cold water injection process; The right-side production well is equipped with constant hydraulic head and heat outflow conditions to simulate hot water collection. The upper and lower boundaries are set as thermally insulating boundaries to avoid heat exchange with the external environment; The initial conditions were set to a uniform temperature distribution of 100°C for both the matrix and the fracture network. Material properties: Fractured regions: high conductivity, low porosity, emphasizing that fractures are the main channels for fluid flow; Matrix region: low conductivity and high porosity ensure that heat conduction is the main mode of heat transfer; The thermal properties parameters are set based on typical thermal reservoir lithological parameters, namely thermal conductivity, specific heat capacity, and density.

2. The method for predicting fluid heat transfer processes in enhanced dry hot rock geothermal systems based on ANN networks as described in claim 1, characterized in that, The specific operation of data standardization is as follows: use sklearn.preprocessing.StandardScaler to standardize the input parameters and output temperature breakthrough curve data, and subtract the mean from each feature and then divide by the standard deviation.

3. The method for predicting fluid heat transfer processes in enhanced dry hot rock geothermal systems based on ANN networks as described in claim 1, characterized in that, The training set is divided as follows: the first 12,000 data points are selected from the database for model learning; the validation set consists of 2,000 data points immediately following the training set, used for model performance evaluation.

4. The method for predicting fluid heat transfer processes in enhanced dry hot rock geothermal systems based on ANN networks as described in claim 1, characterized in that, The output of the numerical model simulation includes: the dynamic distribution of the temperature field; the temperature breakthrough curves at each monitoring point; and the path and extent of cold water flow in the fracture network.

Images