A machine learning-based method for predicting favorable geothermal areas

Abstract

The present invention proposes a method for predicting favorable geothermal regions based on machine learning, which relates to the field of predicting favorable geothermal regions. The prediction process of favorable geothermal regions proposed by the present invention includes constructing a prediction dataset for favorable geothermal regions, automatically adjusting the time window size by using the dynamic time window normalization method to normalize time series data, using an adaptive Fourier convolution module to fuse the temporal and spatial features of time series data, then performing high-dimensional embedding on the features through linear transformation and using an adaptive filter for denoising, performing non-linear dimensionality reduction on the feature sequence, using the inverse dynamic time window normalization method to restore to the original data scale, and finally splicing all the restored feature data into a final prediction vector and inputting it into a fully connected neural network to calculate the predicted value of the favorable geothermal region. This method can improve the adaptability to non-stationary data and fuse temporal and spatial features, thereby improving the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of geothermal favorable area prediction, and specifically relates to a method for predicting geothermal favorable areas based on machine learning. Background Art

[0002] Geothermal favorable areas refer to areas where underground geothermal energy resources are rich and geological conditions are suitable for development and utilization. These areas usually have high geothermal gradients, abundant hot water and steam resources. Although geothermal energy has the advantages of cleanness and renewability, unreasonable development of geothermal favorable areas may lead to geological disasters such as groundwater pollution and land subsidence. Therefore, predicting geothermal favorable areas is crucial for the scientific and reasonable development of geothermal resources and the reduction of environmental risks.

[0003] In the field of geothermal favorable area prediction, analysis methods based on statistical models and machine learning methods are mainly used. The analysis based on statistical models usually relies on expert experience, and its accuracy and applicability are limited in the face of complex non-linear relationships and dynamic changes. Machine learning methods can automatically learn trends and periodic features from data and have strong adaptability. However, existing machine learning methods generally cannot effectively handle the non-stationarity problem in time series, and separately modeling spatial and temporal features ignores the intrinsic coupling relationship between time and space, which limits the accuracy of machine learning methods in geothermal favorable area prediction.

[0004] Geothermal favorable area data often shows obvious non-stationarity, such as seasonal fluctuations and mutation events. In machine learning, adopting appropriate stationary transformation methods and methods of jointly modeling spatial and temporal features can improve the adaptability of the model to non-stationary data and avoid the information loss problem caused by separately modeling spatial and temporal features, thereby significantly improving the accuracy of geothermal favorable area prediction and providing more reliable decision-making support for resource development and environmental protection. Summary of the Invention

[0005] The present invention provides a method for predicting geothermal favorable areas based on machine learning, proposes a dynamic time window normalization method, adaptively adjusts the window size based on the volatility of the data, normalizes the time series data of geothermal favorable area factors, uses an adaptive Fourier convolution module to transform the normalized data into the Fourier space and perform convolution operations, fuses temporal and spatial features, further processes the data through high-dimensional embedding and noise filtering and performs dimensionality reduction, and finally uses the inverse dynamic time window normalization method to inverse-normalize the dimensionality-reduced feature data to obtain the feature data at the original scale, and inputs it into a fully connected neural network to complete the prediction of geothermal favorable areas.

[0006] The technical methods adopted by the present invention to achieve the above object specifically include the following steps:

[0007] S1. Collect the factor data affecting the geothermal favorable area and construct a prediction dataset for the geothermal favorable area;

[0008] S2. Since the factor data is time-series data, a dynamic time window normalization method is proposed. A dynamic time window is introduced, and the size of this window is automatically adjusted according to the volatility between the current time step and subsequent time steps. Based on the mean and standard deviation of the data within the window, the data at the current time step is normalized to obtain the time-series data of the normalized factors of the geothermal favorable area;

[0009] S3. Construct an adaptive Fourier convolution module, transform the normalized time-series data into the Fourier space, introduce an adaptive convolution kernel, and perform a convolution operation on the data to obtain a Fourier sequence that fuses time and spatial features;

[0010] S4. Perform high-dimensional embedding on the Fourier sequence of the geothermal favorable area factors through a linear transformation to obtain a high-dimensional embedding vector sequence, and use an adaptive filter to filter out the noise in the high-dimensional embedding vectors;

[0011] S5. Use a non-linear transformation to perform dimensionality reduction on the feature sequence after filtering out the noise;

[0012] S6. A reverse dynamic time window normalization method is proposed. Based on the mean and standard deviation of the data within the window obtained by the dynamic time window normalization method, the dimensionality-reduced feature sequence of the geothermal favorable area factors is reverse-normalized to obtain the feature data values at the original data scale;

[0013] S7. Concatenate all the feature data values at the original data scale to form a final prediction vector, input it into a fully connected neural network, and calculate to obtain the predicted value of the geothermal favorable area.

[0014] Preferably, in S1, when collecting the factor data affecting the geothermal favorable area, including crustal thickness, rock type, surface temperature, underground temperature, heat flow density, precipitation, for the data values missing at the th time step of the th factor data of the geothermal favorable area that has not been collected, linear interpolation is used to fill it. The specific mathematical model is:

[0015] ;

[0016] In the formula, is the data value obtained by linear interpolation at the th time step of the th factor data of the geothermal favorable area, , are respectively the th and , The data values collected at the time step are used to fill the missing values through this method to construct the geothermal favorable area prediction dataset.

[0017] Preferably, in S2, input Time series data of favorable regional factors for geothermal ,in is the size of the time series data lookback window, For the The data of favorable regional factors of geothermal The data value at the time step, , a dynamic time window normalization method is proposed, which specifically includes the following steps:

[0018] S21, introduce dynamic time window, calculate the The standard deviation between the current time step and the subsequent time step in the time series data of the geothermal favorable regional factors is used to quantify the volatility of the time series and determine the current time step. The dynamic time window size, the specific mathematical model is:

[0019] ;

[0020] ;

[0021] ;

[0022] Where, For time series data From Time step to The mean of the data within the time step, For the The number of data points between a time step and subsequent time steps, For time series data From Time step to The standard deviation of the data within a time step, For the The time series data of favorable regional factors for geothermal Dynamic time window size of the time step, A hyperparameter that controls the degree to which volatility affects the window size. , To ensure that the window size has a minimum bias, , To round up;

[0023] S22. Calculate time series data In dynamic time window The mean and standard deviation of the internal data, and the data at the time step are normalized according to the mean and standard deviation. The specific mathematical model is:

[0024] ;

[0025] ;

[0026] ;

[0027] In the formula, is the mean value of the data within the subsequence formed from the time step to the time step, is the standard deviation of the data within the subsequence formed from the time step to the time step, is the normalized data value. After normalizing all the time series data of the geothermal favorable area factors, the time series data is obtained. .

[0028] Preferably, by introducing a calculation mechanism for the dynamic time window size, the window at each time step can be adaptively adjusted according to the volatility of the current data. By calculating the mean and standard deviation of the data within the window, the change characteristics of the time series can be accurately captured. Especially when the data fluctuates greatly, flexibly adjusting the window size can obtain more accurate feature extraction. The dynamic time window normalization method eliminates the differences in data scales at different time steps, thereby improving the consistency and stability of the data.

[0029] Preferably, in S3, all the time series data of the geothermal favorable area factors after normalization are denoted as , where is the number of data types of the geothermal favorable area factors, is the time series data of the th geothermal favorable area factor after normalization. Regarding each data value in as a node, each node is connected to all other nodes to form a fully connected graph , where is the node feature, is the adjacency matrix of the fully connected graph. An adaptive Fourier convolution module is constructed by introducing an adaptive convolution kernel, which specifically includes the following steps:

[0030] S31. Embed each node using a lookup table, and calculate the similarity weights between their high-dimensional embedding vectors after obtaining them. The specific mathematical model is as follows:

[0031] ;

[0032] ;

[0033] ;

[0034] In the formula, is the high-dimensional embedding vector sequence obtained by the node in the graph through the lookup table function, where is the th data value of the geothermal favorable area factor at the th time step obtained by the lookup table function as the high-dimensional embedding vector, is the vector dimension obtained after the lookup table function, is the lookup table function, is the th data of the geothermal favorable area factor at the th time step and the similarity score of the data of the th geothermal favorable area factor at the th time step, is the transpose of , is the normalized similarity weight;

[0035] S32. Introduce an adaptive convolution kernel, and obtain the convolution kernel by transforming the adjacency matrix and the similarity weights. The specific mathematical model is as follows:

[0036] ;

[0037] In the formula, is the convolution kernel between the node and the node in , is the vector dimension, is the connection strength between the node and the node in , is the trainable weight matrix, is element-wise multiplication;

[0038] S33. After obtaining the adaptive convolution kernel, convert the convolution operation into multiplication in the Fourier space through Fourier transform. The specific mathematical model is as follows:

[0039] ;

[0040] Wherein, is the Fourier sequence of the geothermal favorable area factor data obtained after discrete Fourier transform and convolution operation, is the discrete Fourier transform, is the convolution operation.

[0041] Preferably, a fully connected graph is constructed to represent the time series data of each geothermal favorable area factor, which can effectively capture the local and global structural information of the data. Based on this graph, an adaptive convolution kernel is introduced, and the nodes are embedded in a high-dimensional space through a lookup table to calculate the similarity between nodes, so as to help generate an adaptive convolution kernel to capture the correlation between different time steps and regions, and obtain information containing both spatial and temporal features. The Fourier transform can transform the convolution operation into multiplication in the Fourier space, reducing the computational complexity and improving the operation efficiency.

[0042] Preferably, in S4, an adaptive filter is constructed to filter the Fourier sequence of the geothermal favorable area factor data, which specifically includes the following steps:

[0043] S41. Use a linear transformation to transform the Fourier sequence of the geothermal favorable area factor data to obtain a high-dimensional embedding vector sequence of the data. The specific mathematical model is:

[0044] ;

[0045] Wherein, is the high-dimensional embedding vector sequence obtained by linearly transforming the Fourier sequence of the geothermal favorable area factor data, is the vector dimension obtained after embedding, is the trainable weight matrix, is the bias term;

[0046] S42. Use a neural network to construct an adaptive filter, and apply the learned filter to the high-dimensional embedding vector sequence of the geothermal favorable area factor data to adjust the frequency information contained in the sequence. The specific mathematical model is:

[0047] ;

[0048] ;

[0049] Wherein, is the adaptive filter, is the th trainable weight matrix, is A trainable weight matrix, is the element-wise product along dimensions, is the sigmoid activation function, is the feature sequence after the high-dimensional embedding vector sequence of geothermal favorable area factor data is filtered by an adaptive filter to remove noise.

[0050] Preferably, the Fourier sequence of geothermal favorable area factors is mapped to a high-dimensional space through a linear transformation, so that the frequency domain information in the Fourier sequence can be embedded and represented. The adaptive filter constructed by the neural network optimizes these frequency information, automatically adjusts the weights of different frequency information, so as to strengthen the representation of key features and suppress the interference of irrelevant information in the subsequent processing process, and improve the accuracy of the model.

[0051] Preferably, in S5, a nonlinear transformation is used to reduce the dimension of the feature sequence after filtering out noise. The specific mathematical model is:

[0052] ;

[0053] In the formula, is the feature sequence after filtering out noise is the geothermal favorable area factor feature sequence obtained after dimensionality reduction, is the th eigenvalue obtained by dimensionality reduction of the feature sequence after filtering out noise, is a learnable weight matrix, is the bias term, is the ReLU activation function.

[0054] Preferably, by performing dimensionality reduction on the feature sequence after noise filtering through a nonlinear transformation, the dimension of the data can be reduced and the most representative features can be extracted. By applying the ReLU activation function, negative value features can be effectively removed, while the expression of positive value features can be strengthened, improving the learning ability and prediction accuracy of the subsequent model.

[0055] Preferably, in S6, an inverse dynamic time window normalization method is proposed. First, according to the geothermal favorable area factor data type, is segmented to obtain the feature sequence , where is the feature sequence of the th type of geothermal favorable area factor, where is the th feature value of the feature sequence of the th time step of the For each time step data in the characteristic sequence of geothermal favorable area factors, using the corresponding mean and standard deviation obtained in S2, restore it to the original data scale. The specific mathematical model is:

[0056] ;

[0057] In the formula, is the characteristic data value at the original data scale obtained by inverse normalization of the characteristic value at the th time step of the characteristic sequence of geothermal favorable area factors. and are the mean and standard deviation at the th time step of the characteristic sequence of geothermal favorable area factors obtained in S2. Combine all to obtain the inverse-normalized characteristic sequence , where is the inverse-normalized characteristic sequence of geothermal favorable area factors. , where is the th data value in the inverse-normalized characteristic sequence of geothermal favorable area factors.

[0058] Preferably, the normalized characteristic data is restored to its original data scale by the inverse dynamic time window normalization method. Since the normalization operation often makes the data lose the intuitive meaning of the original scale, this inverse normalization method can effectively map the normalized data back to the actual numerical range and ensure the interpretability of the data.

[0059] Preferably, in S7, the data values in the inverse-normalized characteristic sequence are concatenated in ascending order according to the data type and time step to obtain the unique feature vector for prediction. Pass it through a fully connected neural network to calculate the predicted value of the geothermal favorable area. The specific mathematical model is:

[0060] ;

[0061] ;

[0062] In the formula, is the unique feature vector for prediction, is the concatenation operation, is the final predicted value of the geothermal favorable area, is the forward propagation operation of the fully connected neural network.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention proposes a dynamic time window normalization method, which optimizes the normalization process according to data volatility by dynamically adjusting the size of the time window to ensure the stability of time series data. The normalized data is transferred to the Fourier space by an adaptive Fourier convolution module, and the time and space features are effectively fused through convolution operations to enhance the perception of complex patterns. Through high-dimensional embedding and adaptive filters, the feature representation is further optimized. After reducing the dimensionality of the feature data using dimensionality reduction techniques, the data is restored to the original scale by the inverse dynamic time window normalization method and input into a fully connected neural network for the final prediction of geothermal favorable areas, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flowchart of the steps of a method for predicting geothermal favorable areas based on machine learning;

[0065] Figure 2 is a diagram of dynamic time window normalization;

[0066] Figure 3 is a diagram of high-dimensional embedding and adaptive filters;

[0067] Figure 4 is a diagram of inverse dynamic time window normalization;

[0068] Figure 5 is a diagram of the prediction effect of geothermal favorable areas. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] The present invention proposes a method for predicting geothermal favorable areas based on machine learning. The method proposes dynamic time window normalization, which dynamically adjusts the window size according to the volatility of time steps to normalize time series data. An adaptive Fourier convolution module is used to transfer the normalized data to the Fourier space, and time and space features are fused through convolution. The sequence data in the Fourier space is subjected to high-dimensional embedding and noise is removed using an adaptive filter, followed by dimensionality reduction through a non-linear transformation. The inverse dynamic time window normalization method is used to restore the reduced-dimensional feature data to the original scale. Finally, all the data is concatenated and input into a fully connected neural network to calculate the predicted value of geothermal favorable areas. The technical solutions in the embodiments of the present invention will be described in detail and completely below, specifically including the following steps, as Figure 1 shown.

[0070] S1. Collect data on factors affecting geothermal favorable areas and construct a prediction dataset for geothermal favorable areas.

[0071] Further, in S1, data on crustal thickness, rock type, surface temperature, underground temperature, heat flow density, and precipitation for 95 days are collected, a total of For the factor data of the geothermal favorable area, preprocess the collected data, and use the linear interpolation method to fill in the data value of the time step where the factor data of the geothermal favorable area is not collected. The specific mathematical model is:

[0072] ;

[0073] In the formula, is the data value obtained by the linear interpolation method for the factor data of the geothermal favorable area at the th time step. , are respectively the data values collected for the factor data of the geothermal favorable area at the th , time steps. After supplementing the missing values, construct a prediction data set for the geothermal favorable area, and then divide the prediction data set for the geothermal favorable area into a training set and a validation set in chronological order, with a division ratio of 8:2. Use the data of the previous 7 days as the historical data input for model training.

[0074] S2. Since the factor data is time series data, a dynamic time window normalization method is proposed. Introduce a dynamic time window, and the size of this window is automatically adjusted according to the volatility between the current time step and subsequent time steps. Normalize the data of the current time step based on the mean and standard deviation within the window to obtain the time series data of the normalized geothermal favorable area factors.

[0075] Furthermore, in S2, the size of the input data backtracking window is for the time series data of the geothermal favorable area factors. Among them, is the data value of the geothermal favorable area factor at the th time step. , a dynamic time window normalization method is proposed. As Figure 2 shown, it specifically includes the following steps:

[0076] S21. Introduce a dynamic time window, and calculate the standard deviation between the time step and the subsequent time steps in the time series data of the geothermal favorable area factors. , where is the number of data between the th time step and the subsequent time steps. Quantify the volatility of the time series according to the standard deviation, and determine the size of the dynamic time window for the current time step . The specific mathematical model is:

[0077] ;

[0078] ;

[0079] ;

[0080] wherein, is the time series data from the th time step to the th time step, the mean value of the data within the time steps, is the time series data from the th time step to the th time step, the standard deviation of the data within the time steps, is the th dynamic time window size in the time series data of the favorable geothermal area factors, is the hyperparameter that controls the influence degree of volatility on the window size, is set to 1, is the bias term, is set to 2 to ensure that the window size is not less than , is the ceiling function;

[0081] S22. According to the dynamic time window size calculated in S21 , calculate the mean value and standard deviation of the data within the window, and normalize the data at the th time step. The specific mathematical model is:

[0082] ;

[0083] ;

[0084] ;

[0085] wherein, , are respectively from the th time step to the th time step, the mean value and standard deviation of the data within the subsequence formed, is the normalized data value. After normalizing all the time series data of the favorable geothermal area factors, the time series data is obtained .

[0086] S3. Construct an adaptive Fourier convolution module to transform the normalized time series data into the Fourier space, introduce an adaptive convolution kernel, perform a convolution operation on the data, and obtain a Fourier sequence that fuses temporal and spatial features.

[0087] Further, in S3, combine the time series data of all types of geothermal favorable area factors after normalization to obtain , and regard each data value in as a node to obtain node features . Fully connect all nodes through the adjacency matrix to form a graph , where is the fully connected adjacency matrix. Propose an adaptive Fourier convolution module, which specifically includes the following steps:

[0088] S31. For each node in the graph , use a lookup table for high-dimensional embedding. After obtaining the high-dimensional embedding vector of the node, calculate the similarity between nodes through the dot product calculation method. The specific mathematical model is:

[0089] ;

[0090] ;

[0091] ;

[0092] In the formula, is the high-dimensional embedding vector sequence obtained by the node in the graph through the lookup table function, where is the high-dimensional embedding vector obtained by the data value of the -th geothermal favorable area factor at the -th time step through the lookup table function, is the data value of the is the vector dimension obtained after passing through the lookup table function, , is the lookup table function, is 's transpose, is the similarity score between the data of the -th geothermal favorable area factor at the -th time step and the data of the -th geothermal favorable area factor at the -th time step, is 's normalized similarity weight;

[0093] S32. Introduce an adaptive convolution kernel, and through the similarity weight and the adjacency matrix Perform a transformation to obtain a convolution kernel. The specific mathematical model is as follows:

[0094] ;

[0095] In the formula, is the convolution kernel between node and node in is the connection strength between node and node in is a trainable weight matrix, is element-wise multiplication;

[0096] S33. After obtaining the adaptive convolution kernel, convert the convolution operation into multiplication in the Fourier space through Fourier transform. The specific mathematical model is as follows:

[0097] ;

[0098] In the formula, is the Fourier sequence of the geothermal favorable area factor data obtained after discrete Fourier transform and convolution operation, is the discrete Fourier transform operation, is the convolution operation.

[0099] S4. Perform high-dimensional embedding on the Fourier sequence of the geothermal favorable area factor through linear transformation to obtain a high-dimensional embedding vector sequence, and use an adaptive filter to filter out the noise in the high-dimensional embedding vector.

[0100] Furthermore, in S4, as Figure 3 shown, pass the Fourier sequence of the geothermal favorable area factor data through a linear transformation and an adaptive filter, which specifically includes the following steps:

[0101] S41. Perform a linear transformation on the Fourier sequence of the geothermal favorable area factor data to obtain a high-dimensional embedding vector sequence of the data. The specific mathematical model is as follows:

[0102] ;

[0103] In the formula, is the high-dimensional embedding vector sequence obtained by performing a linear transformation on the Fourier sequence of the geothermal favorable area factor data, is a trainable weight matrix responsible for mapping the Fourier sequence to a high-dimensional space, is the bias term;

[0104] S42. Construct an adaptive filter using a neural network. The high-dimensional embedded vector sequence of geothermal favorable area factor data contains frequency information. Use trainable weight matrices to adjust the frequency information of this sequence. The specific mathematical model is:

[0105] ;

[0106] ;

[0107] In the formula, is the adaptive filter, is the th trainable weight matrix, is th trainable weight matrix, is the element-wise product along the last dimension of the vector, is the sigmoid activation function, is the feature sequence of geothermal favorable area factors after being filtered by the adaptive filter.

[0108] S5. Use non-linear transformation to reduce the dimension of the feature sequence after filtering out noise.

[0109] Furthermore, in S5, through non-linear transformation, the feature sequence after being processed by the adaptive filter is reduced to a lower dimension. The specific mathematical model is:

[0110] ;

[0111] In the formula, is the feature sequence after being processed by the adaptive filter is the feature sequence of geothermal favorable area factors obtained after dimensionality reduction, is the th eigenvalue obtained by reducing the dimension of the feature after filtering out noise, is the learnable weight matrix, is the bias term, is the ReLU activation function.

[0112] S6. Propose an inverse dynamic time window normalization method. Based on the mean and standard deviation of the data within the window obtained by the dynamic time window normalization method, inverse normalize the dimensionality-reduced feature sequence of geothermal favorable area factors to obtain the feature data values at the original data scale.

[0113] Furthermore, in S6, propose an inverse dynamic time window normalization method. As Figure 4 shown, restore the normalized feature sequence of geothermal favorable area factors to the original scale. First, The data in is segmented according to the type of geothermal favorable area factor data to obtain a new feature sequence , where is the feature sequence of the th type of geothermal favorable area factor, is the feature value of the th time step of the feature sequence of the th type of geothermal favorable area factor. Using the corresponding mean and standard deviation obtained in S2,

[0114] is restored to the original data scale. The specific mathematical model is:

[0115] In the formula, is the feature data value of the original data scale restored after inverse normalization of the feature value of the th type of geothermal favorable area factor at the th time step, is the mean of the th type of geothermal favorable area factor at the th time step obtained in S2, is the standard deviation of the th type of geothermal favorable area factor at the th time step obtained in S2. After inverse normalization of all data values in , the feature sequence is obtained, where is the feature sequence of the

[0116] S7. Concatenate all feature data values of the original data scale to form a final prediction vector, and input it into a fully connected neural network to calculate the predicted value of the geothermal favorable area.

[0117] Further, in S7, the data values in the inverse-normalized feature sequence are concatenated in ascending order according to the data type and time step to form a unique feature vector for prediction. Use the fully connected neural network to calculate the unique feature vector to obtain the predicted value of the geothermal favorable area. The specific mathematical model is:

[0118] ;

[0119] ;

[0120] In the formula, is the concatenated feature vector, which contains the feature data corresponding to each time step of the It is a splicing operation, is the predicted value of the final geothermal favorable area, is the forward propagation operation of the fully connected neural network.

[0121] Furthermore, this method uses the Python 3.8 programming language and is implemented based on the PyTorch framework. The training process is carried out in the CUDA 12 environment and the NVIDIA RTX 3090 GPU is used for calculation. During the training process, the learning rate of this method is set to , the batch size is set to , the optimizer is optimizer, and the loss function is mean squared error . In order to stabilize the training process, the weight decay is adopted as .

[0122] Furthermore, the prediction effect of this method is as Figure 5 shown, where the ordinate is the possibility (%) of the geothermal favorable area, and the abscissa is the time (days). The gray dotted line in the figure represents the actual possibility value of the geothermal favorable area, and the black solid line is the predicted possibility value of the geothermal favorable area obtained by this method. Through the analysis of Figure 5 , it can be seen that there is a high degree of fit between the predicted value and the actual possibility value, indicating that this method has strong prediction ability in the prediction of the geothermal favorable area, and verifies the effectiveness and feasibility of this method in this field.

[0123] The above is only the preferred implementation manner of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the creative concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for predicting favorable geothermal areas based on machine learning, characterized in that, Including the following steps: S1. Collect the factor data affecting the geothermal favorable area and construct a prediction data set for the geothermal favorable area; S2. The factor data is time series data. A dynamic time window normalization method is proposed, and a dynamic time window is introduced. The size of this window is automatically adjusted according to the volatility between the current time step and subsequent time steps. The data at the current time step is normalized based on the mean and standard deviation within the window to obtain the time series data of the normalized geothermal favorable area factors; S3. Construct an adaptive Fourier convolution module, transform the normalized time series data into the Fourier space, introduce an adaptive convolution kernel, and perform a convolution operation on the data to obtain a Fourier sequence integrating time and spatial features; S4. Perform high-dimensional embedding on the Fourier sequence of the geothermal favorable area factors through a linear transformation to obtain a high-dimensional embedding vector sequence, and use an adaptive filter to filter out the noise in the high-dimensional embedding vectors; S5. Use a non-linear transformation to perform dimensionality reduction on the feature sequence after filtering out the noise; S6. A reverse dynamic time window normalization method is proposed. Based on the mean and standard deviation of the data within the window obtained by the dynamic time window normalization method, the dimensionality-reduced feature sequence of the geothermal favorable area factors is reverse-normalized to obtain the feature data values at the original data scale; S7. Concatenate all the feature data values at the original data scale to form a final prediction vector, input it into a fully connected neural network, and calculate the predicted value of the geothermal favorable area.

2. The geothermal favorable area prediction method based on machine learning according to claim 1, wherein In S1, data on factors affecting favorable geothermal areas are collected, including crustal thickness, rock type, surface temperature, underground temperature, heat flow density, and precipitation. The collected data is preprocessed, and the linear interpolation method is used to fill in the data values of the time steps when the data of the factors of favorable geothermal areas are not collected. The specific mathematical model is as follows: ； wherein, is the data value obtained by linear interpolation for the nth geothermal favorable area factor data at the nth time step, , are respectively the data values collected for the nth geothermal favorable area factor data at the , nth time step. After filling in all missing values, a prediction data set for geothermal favorable areas is constructed.

3. The geothermal favorable area prediction method based on machine learning according to claim 2, characterized in that In S2, a dynamic time window normalization method is proposed. The time series data of the th geothermal favorable area factor is denoted as , where is the size of the time series backtracking window, and is the th geothermal favorable area factor data at the th time step. For , dynamic time window normalization is performed on , which specifically includes the following steps: S21. Introduce a dynamic time window, and calculate the standard deviation between the time series data of the favorable geothermal area factors for farming in the time step and subsequent time steps, and obtain the dynamic time window size of the time step. The specific mathematical model is as follows:​ ； ； ； Wherein, and are the mean and standard deviation of the data within the time series data from the th time step to the th time step, is the number of data between the th time step and the subsequent time steps, is the dynamic time window size of the th time step in the time series data of the favorable geothermal area factor, is the hyperparameter that controls the influence degree of volatility on the window size, is the bias term, is the ceiling function; is rounding up; S22. Calculate the dynamic time window of the data within the time step, and normalize the data of the time step according to the calculated mean and standard deviation. The specific mathematical model is as follows: ； ； ； In the formula, and are respectively the mean value and standard deviation of the data within the subsequence formed from the th time step to the th time step. is the normalized data value. After normalizing all the time series data of the th geothermal favorable area factor, the time series data is obtained.

4. A method for predicting geothermal favorable areas based on machine learning according to claim 3, characterized in that, In S3, the time series data combinations of all geothermal favorable area factors collected are obtained , where is the number of geothermal favorable area factor data types, is the time series data of the th normalized geothermal favorable area factor. Regarding each data value in as a node, the node feature is obtained. All nodes are pairwise connected to form a fully connected graph , where is the adjacency matrix of the fully connected graph. An adaptive Fourier convolution module is proposed, which specifically includes the following steps: S31. Use a lookup table function to obtain the high-dimensional embedding vectors of each node, and calculate the similarity weights between the high-dimensional embedding vector nodes. The specific mathematical model is: ； ； ； In the formula, is the high-dimensional embedding vector sequence obtained by the node in Figure through the look-up table function, where is the th geothermal favorable area factor at the th time step, is the transpose of , is the vector dimension obtained after passing through the look-up table function, is the look-up table function, is the th geothermal favorable area factor at the th time step and the th geothermal favorable area factor at the th time step similarity score, is the normalized similarity weight; S32. Introduce an adaptive convolution kernel, and calculate the convolution kernel of the node pair through the similarity weight and the adjacency matrix The specific mathematical model is as follows: ； In the formula, and are respectively the convolution kernel and the connection strength between nodes and node , is a trainable weight matrix, is element-wise multiplication; S33. Convert the convolution operation on the graph to multiplication in the Fourier space through Fourier transform. The specific mathematical model is as follows: ； In the formula, is the Fourier sequence obtained after discrete Fourier transform and convolution operation, is the discrete Fourier transform, is the convolution operation.

5. The geothermal favorable area prediction method based on machine learning according to claim 4, characterized in that In step S4, an adaptive filter is constructed, which specifically includes the following steps: S41. Embed the Fourier sequence of the geothermal favorable area factor data into a high-dimensional space through a linear transformation to obtain a high-dimensional embedding vector sequence of the data. The specific mathematical model is: ； In the formula, is a high-dimensional embedding vector sequence obtained by performing a linear transformation on the Fourier sequence of the geothermal favorable area factor data, is a trainable weight matrix, is a bias term, is the vector dimension obtained after embedding through a linear transformation; S42. Use a neural network to construct an adaptive filter, and adjust the frequency information contained in this sequence through the adaptive filter. The specific mathematical model is: ； ； In the formula, is an adaptive filter, is the th trainable weight matrix, is trainable weight matrices, is the element-wise product along the dimension, is the sigmoid activation function, is the feature sequence after the high-dimensional embedding vector sequence of geothermal favorable area factors is filtered by the adaptive filter to remove noise.

6. The geothermal favorable area prediction method based on machine learning according to claim 5, wherein, In the step S5, the feature sequence after filtering out noise by the adaptive filter is dimension-reduced by a non-linear transformation. The specific mathematical model is as follows: ； In the formula, is the feature sequence after filtering out noise is the feature sequence of geothermal favorable area factors obtained after dimensionality reduction is the eigenvalue obtained by dimensionality reduction of the th feature in is the learnable weight matrix is the bias term is the ReLU activation function 7. A method for predicting geothermal favorable areas based on machine learning according to claim 6, characterized in that, In S6, an inverse dynamic time window normalization method is proposed to normalize the feature sequence of geothermal favorable area factors which is segmented according to factor types to obtain a new feature sequence , where is the feature sequence of the th geothermal favorable area factor, is the feature value of the th geothermal favorable area factor at the th time step. Using the corresponding mean and standard deviation obtained in S2, the data at each time step in the new feature sequence is restored to the original data scale. The specific mathematical model is as follows: ； In the formula, is the characteristic data value of the original data scale restored after inverse normalization of the characteristic sequence of the nth geothermal favorable area factor at the nth time step, and are the mean and standard deviation of the nth geothermal favorable area factor at the nth time step obtained from S2, and after the inverse dynamic time window normalization method, the .

Images