Geothermal prediction method based on machine learning

Abstract

The invention provides a geothermal prediction method based on machine learning, and the method comprises the steps: S11, obtaining the temperature measurement data and boundary conditions of the historical mining data of a geothermal well mining affected area, carrying out the processing of the missing points and abnormal values of the temperature measurement data and boundary conditions based on a KNN interpolation method, and adding Gaussian noise; s12, using a convolutional neural network to extract deep features of the boundary conditions, using a bidirectional long-short time memory network to capture time sequence features of the temperature measurement data, and determining target influence factors of the deep features of the boundary conditions and the time sequence features of the temperature measurement data; s13, constructing a terrestrial heat prediction model, inputting the deep features of the boundary conditions, the time sequence features of the temperature measurement data and the target influence factors into the terrestrial heat prediction model for training, and obtaining a trained terrestrial heat prediction model; and S14, mining data of an actual geothermal well are input into the geothermal prediction model, and prediction of geothermal reserves is obtained. According to the invention, based on machine learning, temperature measurement data and boundary conditions are combined, and accurate prediction of geothermal reserves is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of geothermal prediction, and particularly to a geothermal prediction method based on machine learning. Background Art

[0002] Geothermal energy, as a clean, low-carbon and widely distributed renewable energy source, has significant advantages. Although it has many favorable characteristics compared with other renewable energy sources such as wind energy and solar energy, it lags far behind in exploration, development and utilization. The exploration and development of geothermal energy are affected by the complexity of geological conditions and mainly exist deep underground, which increases the difficulty of exploration and development. In areas with insufficient exploration, drilling activities are risky and it is difficult to accurately locate high-temperature geothermal resources. The exploration and development costs of geothermal resources are high, which makes it particularly important to conduct effective exploration area selection analysis for the areas to be developed based on the existing exploration results and development cases.

[0003] However, current exploration technologies often rely on simple analysis of geothermal exploration data. This method is subjective and may overlook some key geological information. This results in insufficient accuracy of the exploration area selection results, thereby increasing the risk of exploration activities. Therefore, it is necessary to develop more advanced technologies to predict the output of geothermal energy, so as to improve the accuracy of geothermal resource exploration and reduce the development risk. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a geothermal prediction method based on machine learning, and obtain the corresponding relationship between temperature measurement data and boundary conditions based on machine learning to achieve accurate prediction of geothermal reserves.

[0005] To achieve the above invention purpose, the present invention provides a geothermal prediction method based on machine learning, and the method includes:

[0006] S11. Obtain the temperature measurement data and boundary conditions of the historical mining data in the geothermal well mining influence area, process the missing points and outliers of the temperature measurement data and boundary conditions based on the KNN interpolation method and add Gaussian noise;

[0007] S12. Use a convolutional neural network to extract the deep features of the boundary conditions, use a bidirectional long short-term memory network to capture the temporal features of the temperature measurement data, determine the corresponding relationship between the deep features of the boundary conditions and the temporal features of the temperature measurement data, and obtain the target influence factor of the boundary conditions on the temperature measurement data;

[0008] S13. Construct a geothermal prediction model, input the deep features of the boundary conditions, the temporal features of the temperature measurement data and the target influence factor into the geothermal prediction model for training, and obtain the trained geothermal prediction model;

[0009] S14. Input the extraction data of the actual geothermal well into the geothermal prediction model to obtain the prediction of the geothermal reserve.

[0010] Further, in step S11, based on the KNN interpolation method, the adjacent values of each feature missing point in the temperature measurement data and boundary conditions are automatically calculated to fill the missing values in the temperature measurement data and boundary conditions, and the outlier is replaced by the average value of the previous and next time steps of the outlier point.

[0011] Further, in step S11, Gaussian noise is added to the temperature measurement data and boundary conditions processed by the KNN interpolation method, specifically including:

[0012] S21. Determine the probability density of the Gaussian noise. The expression of the probability density of the Gaussian noise is:

[0013]

[0014] where N(x) is the probability density of the Gaussian noise, is the normalization factor, σ is the standard deviation of the Gaussian noise, σ 2 is the variance of the Gaussian noise, x is the random variable, and μ is the expectation of the Gaussian noise, that is, the mean value of the Gaussian noise;

[0015] S22. Add the Gaussian noise to the temperature measurement data and boundary conditions after KNN interpolation. Its expression is:

[0016] G(x) = f(x) + N(x)

[0017] where G(x) is the temperature measurement data and boundary conditions with Gaussian noise added, and f(x) is the temperature measurement data and boundary conditions without Gaussian noise added.

[0018] Further, in step S12, the convolutional neural network extracts the deep features of the boundary conditions, specifically including:

[0019] S31. Convert the boundary conditions into feature vectors for representation, and input the feature vectors into the convolutional neural network;

[0020] S32. Introduce residual connections and multi-scale structures in the convolutional neural network to extract the tiny features in the feature vectors;

[0021] S33. Add 3 3*3 convolutions after the residual connection to extract the spatial features in the feature vectors;

[0022] S34. Integrate the tiny features and spatial features through the attention mechanism, and apply a non-linear activation function to capture the deep features of the boundary conditions.

[0023] Further, in step S12, a bidirectional long short-term memory network is used to capture the temporal features of the temperature measurement data, specifically including:

[0024] S41. Extract the forward and backward information of the temperature measurement data in chronological order through a bidirectional long short-term memory network;

[0025] S42. Further extract the output information of step S41 through a long short-term memory network to obtain the time dependence relationship of the temperature measurement data;

[0026] S43. Add a time attention mechanism to adaptively weight the time dependence relationship of the temperature measurement data and capture the temporal features of the temperature measurement data.

[0027] Further, in step S12, determine the corresponding relationship between the deep features of the boundary conditions and the temporal features of the temperature measurement data, and obtain the target influence factor of the boundary conditions on the temperature measurement data, specifically including:

[0028] S51. Quantitatively analyze the deep features and temporal features using the Pearson correlation coefficient to obtain the corresponding correlation matrix;

[0029] S52. Add a time attention mechanism, calculate the similarity between the elements in the corresponding correlation matrix and the labels to obtain the corresponding weights, and share the same weights at the time steps;

[0030] S53. Use the Softmax normalization function to normalize the weights in step S52 to obtain the weight matrix of the corresponding correlation matrix;

[0031] S54. Through the matrix broadcast mechanism, perform the Hadamard product of the weight matrix in step S53 and the corresponding correlation matrix in step S51 to obtain the target influence factor of the boundary conditions on the temperature measurement data.

[0032] Further, in step S13, construct a geothermal prediction model, specifically including:

[0033] S61. Optimize the hyperparameters of the LSTM model through a genetic algorithm;

[0034] S62. Train the LSTM model using the hyperparameters optimized by the genetic algorithm;

[0035] S63. Determine whether the training result meets the preset exit condition. If it meets, output the geothermal prediction model; otherwise, update the weights and biases and execute step S62;

[0036] S64. Input the deep features of the boundary conditions, the temporal features of the temperature measurement data, and the target influence factor into the geothermal prediction model for training to obtain the trained geothermal prediction model.

[0037] Further, in step S61, the hyperparameters of the LSTM model are optimized by a genetic algorithm, specifically including:

[0038] S71. Encode the hyperparameters of the LSTM model into the chromosomes of the genetic algorithm, and randomly generate a number of chromosomes to form an initial population;

[0039] S72. Take the root mean square error of the LSTM model as the fitness value of the genetic algorithm, and determine the fitness function of the genetic algorithm;

[0040] S73. Randomly select chromosomes for crossover and mutation operations, compare the local optimal solutions of the previous generation with the new solutions generated in the next generation, and search for the global optimal solution.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] A geothermal prediction method based on machine learning provided by the present invention can ensure that the geothermal prediction model can learn more comprehensive and accurate features during the training process by filling in the missing values and outliers of the temperature measurement data and boundary conditions of the historical mining data in the geothermal well mining influence area and adding Gaussian noise, thereby improving the accuracy and reliability of geothermal resource prediction. The deep features of the boundary conditions are extracted by a convolutional neural network, and the temporal features of the temperature measurement data are captured by a bidirectional long short-term memory network, so as to determine the corresponding relationship between the boundary conditions and the temperature measurement data, and obtain the target influence factors of the boundary conditions on the temperature measurement data. The hyperparameters of the LSTM are optimized by a genetic algorithm to construct a geothermal prediction model, and the prediction of the geothermal resource reserve is realized. In summary, the present invention realizes the accurate prediction of geothermal reserves based on machine learning combined with temperature measurement data and boundary conditions. Description of the Drawings

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0044] Figure 1 It is a schematic flow chart of a geothermal prediction method based on machine learning provided by an embodiment of the present invention. Detailed Embodiments

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only for explaining the present invention, rather than limiting the present invention. In addition, it should be noted that for the sake of description, only the parts related to the present invention rather than all the structures are shown in the drawings.

[0046] Referring to Figure 1 , this embodiment provides a geothermal prediction method based on machine learning, and the method includes:

[0047] S11. Obtain the temperature measurement data and boundary conditions of the historical mining data in the geothermal well mining influence area, and process the missing points and outliers of the temperature measurement data and boundary conditions based on the KNN interpolation method and add Gaussian noise.

[0048] In this embodiment, the KNN interpolation method can process the missing points in the temperature measurement data and boundary conditions, estimate the missing values through the information of adjacent points, thereby improving the integrity of the data. At the same time, it can identify and correct the outliers in the data, reducing their impact on the overall analysis results. Since the time span of the geothermal well mining data is large but relatively stable, with almost no mutations, it is difficult to collect subtle features. Therefore, Gaussian noise is added to expand the time series data of the mining data to enhance the prediction performance of the subsequent geothermal prediction model.

[0049] In step S11, based on the KNN interpolation method, the missing values in the temperature measurement data and boundary conditions are automatically calculated and filled according to the adjacent values of the missing points of each feature in the temperature measurement data and boundary conditions, and the outliers are replaced by the average value of the previous and next time steps of the outlier points.

[0050] In this embodiment, the KNN interpolation method is used to fill the missing values in the mining data. For each missing point of each feature column in the mining data, select the K nearest known values (adjacent values) to this point. According to these adjacent values, calculate the filling value of the missing point of each feature column in the mining data. The calculation method is: calculate the weighted average value of all adjacent values, where the weight is determined inversely according to the distance between the adjacent value and the missing point of each feature column in the original data, that is, the closer the adjacent value, the greater the weight; if there are multiple equal values among the adjacent values of the missing point of each feature column in the mining data, preferentially select the adjacent value with the closest distance as the filling value; in the case of multiple missing values, repeat the above steps until all the missing points of each feature column in the original data are filled.

[0051] Detect outliers in the mining data, where the outliers are values beyond the normal range, and the specific range is determined according to the actual application scenario. For each detected outlier, replace it with the average value of the previous and next time steps of the outlier point. The calculation method is as follows: Take the values of the previous and next time steps of the outlier point and calculate their average value; if the outlier point is located at the boundary of the data sequence (i.e., the first or last point), then only use the existing neighboring values for the average calculation.

[0052] In step S11, add Gaussian noise to the temperature measurement data and boundary conditions processed by the KNN interpolation method, which specifically includes:

[0053] S21. Determine the probability density of the Gaussian noise. The expression for the probability density of the Gaussian noise is:

[0054]

[0055] where N(x) is the probability density of the Gaussian noise, is the normalization factor, σ is the standard deviation of the Gaussian noise, σ 2 is the variance of the Gaussian noise, x is the random variable, and μ is the expectation of the Gaussian noise, that is, the mean value of the Gaussian noise;

[0056] S22. Add the Gaussian noise to the temperature measurement data and boundary conditions after KNN interpolation. Its expression is:

[0057] G(x) = f(x) + N(x)

[0058] where G(x) is the temperature measurement data and boundary conditions with added Gaussian noise, and f(x) is the temperature measurement data and boundary conditions without added Gaussian noise.

[0059] In this embodiment, the probability spectrum of the Gaussian noise follows a uniform distribution. Therefore, the noise amplitude at any position in the mining data is random. In areas with low exploration levels, the mining data processed by this method can more accurately predict the geothermal reserves and reduce the exploration risk.

[0060] S12. Use a convolutional neural network to extract the deep features of the boundary conditions, use a bidirectional long short-term memory network to capture the temporal features of the temperature measurement data, determine the corresponding relationship between the deep features of the boundary conditions and the temporal features of the temperature measurement data, and obtain the target influence factor of the boundary conditions on the temperature measurement data.

[0061] In this embodiment, the CNN is good at automatically learning and extracting deep features from the input data of boundary conditions. The bidirectional long short-term memory network can capture the time series features of the temperature measurement data in two directions (forward and backward). By determining the correspondence between the deep features of the boundary conditions and the temporal features of the temperature measurement data, the key factors affecting geothermal resources can be more accurately identified, the prediction of geothermal reserves can be realized, and thus the exploration and development strategies of geothermal resources can be optimized.

[0062] In step S12, the convolutional neural network extracts the deep features of the boundary conditions, specifically including:

[0063] S31. Convert the boundary conditions into feature vectors for representation, and input the feature vectors into the convolutional neural network.

[0064] S32. Introduce residual connections and multi-scale structures in the convolutional neural network to extract the tiny features in the feature vectors.

[0065] S33. Add 3 3*3 convolutions after the residual connection to extract the spatial features in the feature vectors.

[0066] S34. Integrate the tiny features and spatial features through the attention mechanism, and apply a non-linear activation function to capture the deep features of the boundary conditions.

[0067] In this embodiment, converting the boundary conditions into feature vectors provides a standardized vector for network processing to represent different input data, enabling the network to more easily process and learn the features of the boundary conditions. By adding residual connections and multi-scale structures in the convolutional neural network, which consists of multiple convolutional layers and pooling layers, it can help solve the gradient vanishing problem in deep networks and allow the network to capture features of different scales in the boundary conditions simultaneously. The 3 3*3 convolutional layers provide a good balance between the number of parameters and the receptive field for extracting the spatial features in the feature vectors. The attention mechanism can automatically assign weights to the tiny features and spatial features in the feature vectors, thereby integrating the tiny features and spatial features. Applying a non-linear activation function increases the non-linear expression ability, enabling the network to capture more complex deep features in the boundary conditions.

[0068] In step S12, the bidirectional long short-term memory network is used to capture the temporal features of the temperature measurement data, specifically including:

[0069] S41. Extract the information of the temperature measurement data in the forward and backward directions in chronological order through the bidirectional long short-term memory network.

[0070] S42. Further extract the output information of step S41 through the long short-term memory network to obtain the time dependence relationship of the temperature measurement data.

[0071] S43. Incorporate a temporal attention mechanism to adaptively weight the temporal dependence relationship of the temperature measurement data and capture the temporal characteristics of the temperature measurement data.

[0072] In this embodiment, a bidirectional long short-term memory network performs sequence modeling on the temperature measurement data to capture the temporal characteristics of temperature changes. The bidirectional long short-term memory network can effectively extract context information and enhance the understanding of temperature dynamic changes. The long short-term memory network is suitable for multivariate time series and can consider multiple influencing factors simultaneously to extract the temporal dependence relationship of the temperature measurement data. At the same time, to enhance the significant expression of different temporal characteristics in the current temperature change prediction sequence, a temporal attention mechanism is designed and incorporated into the above bidirectional long short-term memory network to adaptively weight the deep time series features and capture the temporal characteristics of the temperature measurement data.

[0073] In step S12, determine the correspondence between the deep features of the boundary conditions and the temporal characteristics of the temperature measurement data, and obtain the target influence factor of the boundary conditions on the temperature measurement data, specifically including:

[0074] S51. Quantitatively analyze the deep features and temporal characteristics using the Pearson correlation coefficient to obtain the corresponding correlation matrix.

[0075] S52. Incorporate a temporal attention mechanism, calculate the similarity between the elements in the corresponding correlation matrix and the labels to obtain the corresponding weights, and share the same weights at the time steps.

[0076] S53. Use the Softmax normalization function to normalize the weights in step S52 to obtain the weight matrix of the corresponding correlation matrix.

[0077] S54. Through the matrix broadcast mechanism, perform the Hadamard product of the weight matrix in step S53 and the corresponding correlation matrix in step S51 to obtain the target influence factor of the boundary conditions on the temperature measurement data.

[0078] In this embodiment, the Pearson correlation coefficient is used to quantitatively analyze the deep features and temporal features to accurately measure the linear relationship between different features, thereby obtaining the corresponding correlation matrix. The time attention mechanism is added to calculate the similarity between the elements in the correlation matrix and the labels to obtain the corresponding weights, which allows the model to give different attentions to the information at different time steps when processing sequence data. This helps to capture the key time dynamic features, thereby improving the network's understanding of time dependence. The Softmax normalization function is used to normalize the weights to ensure that the sum of the elements in the weight matrix is 1, which helps to maintain numerical stability in subsequent calculations. Through the matrix broadcasting mechanism, the Hadamard product is performed on the weight matrix and the correlation matrix to integrate the weight information into the correlation degree of each feature, thereby obtaining the target influence factor of the boundary condition on the temperature measurement data, strengthening the influence of important features, and suppressing the interference of unimportant features at the same time.

[0079] S13. Construct a geothermal prediction model, input the deep features of the boundary conditions, the temporal features of the temperature measurement data, and the target influence factor into the geothermal prediction model for training to obtain a trained geothermal prediction model.

[0080] In this embodiment, by constructing a geothermal prediction model, the prediction of the reserves of geothermal resources is improved, thereby ensuring the exploration and development efficiency of geothermal resources, and it can also provide a solid foundation for the sustainable utilization of geothermal energy. By integrating the deep features of the boundary conditions, the model can identify the key geological factors affecting the distribution and flow of geothermal resources; by integrating the temporal features of the temperature measurement data, the model can understand the changing trend of geothermal resources over time, providing insights in the time dimension for prediction; by combining the target influence factor, the model can more accurately predict the distribution and changes of geothermal resources, improving the reliability and accuracy of the prediction.

[0081] In step S13, constructing a geothermal prediction model specifically includes:

[0082] S61. Optimize the hyperparameters of the LSTM model through a genetic algorithm.

[0083] S62. Train the LSTM model using the hyperparameters optimized by the genetic algorithm.

[0084] S63. Determine whether the training result meets the preset exit condition. If it meets, output the geothermal prediction model; otherwise, update the weights and biases, and execute step S62.

[0085] S64. Input the deep features of the boundary conditions, the temporal features of the temperature measurement data, and the target influence factor into the geothermal prediction model for training to obtain a trained geothermal prediction model.

[0086] In this embodiment, there is a problem that it is difficult to determine the optimal parameters during the training process of the LSTM model. Generally, the number of neurons in the hidden layer and the learning rate are selected according to experience or literature references. Inappropriate selected values will directly affect the prediction accuracy of the geothermal prediction model. Therefore, the genetic algorithm is used to optimize the number of neurons in the hidden layer and determine the initial learning amount of the LSTM neural network model to construct a geothermal prediction model.

[0087] In step S61, the hyperparameters of the LSTM model are optimized by the genetic algorithm, specifically including:

[0088] S71. Encode the hyperparameters of the LSTM model as the chromosomes of the genetic algorithm, and randomly generate several chromosomes to form an initial population.

[0089] S72. Take the minimum root mean square error of the LSTM model as the fitness value of the genetic algorithm, and determine the fitness function of the genetic algorithm.

[0090] S73. Randomly select chromosomes for crossover and mutation operations, compare the local optimal solutions of the previous generation with the new solutions generated in the next generation, and search for the global optimal solution.

[0091] In this embodiment, as a global optimization method, the genetic algorithm can find the optimal solution in a large search space, effectively avoid falling into local optima. The genetic algorithm does not rely on gradient information and can handle nonlinear, multimodal, and discontinuous optimization problems, showing strong adaptability in the optimization of LSTM hyperparameters. By randomly selecting chromosomes for crossover and mutation operations, the genetic algorithm can improve the search efficiency while maintaining the diversity of the population, increasing the possibility of finding high-quality solutions. Taking the minimum root mean square error of the LSTM model as the fitness value helps to guide the algorithm to find the hyperparameter combination that can improve the model's prediction performance. RMSE can be naturally mapped to the fitness function, where a smaller RMSE value corresponds to a higher fitness. By using the minimum RMSE as the fitness value, the genetic algorithm can more effectively search the hyperparameter space and find the hyperparameter combination that can improve the prediction performance of the LSTM model. Optimizing the hyperparameters of the LSTM model by the genetic algorithm can improve the model's performance while maintaining the flexibility and robustness of the optimization process.

[0092] S14. Input the exploitation data of the actual geothermal well into the geothermal prediction model to obtain the prediction of the geothermal reserves.

[0093] In this embodiment, the actual geothermal well exploitation data includes but is not limited to exploitation volume, water levels of production wells and reinjection wells, water outlet temperature of production wells, thermal conductivity, porosity, injection and production depth, fracture permeability, injection-production well spacing, etc.; the collected data is cleaned and preprocessed, including removing outliers, filling missing values, data normalization, etc.; the temperature measurement data and boundary conditions are extracted from the preprocessed data and input into the geothermal prediction model to obtain the prediction of the geothermal reserves of the geothermal well.

[0094] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A geothermal prediction method based on machine learning, characterized in that: The method comprises: S11, obtaining temperature measurement data and boundary conditions of historical mining data in the geothermal well mining impact area, processing missing points and abnormal values ​​of the temperature measurement data and boundary conditions based on the KNN interpolation method and adding Gaussian noise; S12, using a convolutional neural network to extract deep features of the boundary conditions, using a bidirectional long short-term memory network to capture the time series features of the temperature measurement data, determining the correspondence between the deep features of the boundary conditions and the time series features of the temperature measurement data, and obtaining the target impact factor of the boundary conditions on the temperature measurement data; S13, constructing a geothermal prediction model, inputting deep characteristics of boundary conditions, time series characteristics of temperature measurement data, and target influencing factors into the geothermal prediction model for training, and obtaining a trained geothermal prediction model; S14. Input the actual geothermal well mining data into the geothermal prediction model to obtain a prediction of geothermal reserves.

2. The geothermal prediction method based on machine learning according to claim 1, characterized in that: In step S11, based on the KNN interpolation method, the missing values ​​in the temperature measurement data and boundary conditions are automatically calculated and filled according to the adjacent values ​​of each feature missing point in the temperature measurement data and boundary conditions, and the abnormal value is replaced by the average value of one time step before and after the abnormal point.

3. The geothermal prediction method based on machine learning according to claim 1, characterized in that: In step S11, Gaussian noise is added to the temperature measurement data and boundary conditions processed by the KNN interpolation method, specifically including: S21. Determine the probability density of Gaussian noise. The expression of the probability density of Gaussian noise is: Where N(x) is the probability density of Gaussian noise, is the normalization factor, σ is the standard deviation of Gaussian noise, σ 2 is the variance of Gaussian noise, x is a random variable, μ is the expectation of Gaussian noise, that is, the mean of Gaussian noise; S22. Add Gaussian noise to the temperature measurement data and boundary conditions after KNN interpolation. The expression is: G(x)=f(x)+N(x) Among them, G(x) is the temperature measurement data and boundary conditions with Gaussian noise added, and f(x) is the temperature measurement data and boundary conditions without Gaussian noise added.

4. The geothermal prediction method based on machine learning according to claim 1, characterized in that: In step S12, the convolutional neural network extracts deep features of the boundary conditions, specifically including: S31, converting the boundary condition into a feature vector for representation, and inputting the feature vector into a convolutional neural network; S32, introduce residual connection and multi-scale structure into convolutional neural network to extract tiny features in feature vector; S33, add three 3*3 convolutions after the residual connection to extract the spatial features in the feature vector; S34. Integrate small features and spatial features through the attention mechanism, and apply nonlinear activation functions to capture the deep features of boundary conditions.

5. The geothermal prediction method based on machine learning according to claim 1, characterized in that: In step S12, a bidirectional long short-term memory network is used to capture the temporal characteristics of the temperature measurement data, specifically including: S41, extracting forward and reverse information from the temperature measurement data according to the chronological order through a bidirectional long short-term memory network; S42, further extracting the output information of step S41 through the long short-term memory network to obtain the time dependency of the temperature measurement data; S43. Add a time attention mechanism to achieve adaptive weighting of the time dependency of the temperature measurement data and capture the temporal characteristics of the temperature measurement data.

6. The geothermal prediction method based on machine learning according to claim 1, characterized in that: In step S12, the correspondence between the deep characteristics of the boundary conditions and the time series characteristics of the temperature measurement data is determined to obtain the target impact factor of the boundary conditions on the temperature measurement data, which specifically includes: S51, using Pearson correlation coefficient to quantitatively analyze deep features and time series features, and obtain the corresponding correlation matrix; S52, adding a time attention mechanism, calculating the similarity between the elements in the corresponding correlation matrix and the label to obtain the corresponding weights, and sharing the same weights at the time step; S53, using the Softmax normalization function to normalize the weights of step S52 to obtain a weight matrix corresponding to the correlation matrix; S54. Through the matrix broadcast mechanism, the weight matrix in step S53 is Hadamard-producted with the corresponding correlation matrix in step S51 to obtain the target impact factor of the boundary condition on the temperature measurement data.

7. The geothermal prediction method based on machine learning according to claim 1, characterized in that: In step S13, a geothermal prediction model is constructed, which specifically includes: S61. Optimize the hyperparameters of the LSTM model using genetic algorithm. S62. Train the LSTM model using the hyperparameters optimized by genetic algorithm. S63, judging whether the training result meets the preset exit condition, if so, outputting the geothermal prediction model, otherwise, updating the weight and bias, and executing step S62; S64, inputting the deep characteristics of the boundary conditions, the time series characteristics of the temperature measurement data and the target influencing factors into the geothermal prediction model for training, and obtaining a trained geothermal prediction model.

8. The geothermal prediction method based on machine learning according to claim 7, characterized in that: In step S61, the hyperparameters of the LSTM model are optimized by a genetic algorithm, specifically including: S71, encode the hyperparameters of the LSTM model into chromosomes of the genetic algorithm, and randomly generate a number of chromosomes to form an initial population; S72, taking the minimum root mean square error of the LSTM model as the fitness value of the genetic algorithm, and determining the fitness function of the genetic algorithm; S73. Randomly select chromosomes for crossover and mutation operations, liberate the local optimal solution of the previous generation to the new solution generated by the next generation for comparison, and find the global optimal solution.

Images