A method for predicting the down-pressure mining potential of natural gas hydrates using deep neural networks

Abstract

A method for predicting the production capacity of natural gas hydrate decompression mining based on deep neural network is provided. [Solution] A method for building a neural network model and using existing multi-scale natural gas hydrate pressure drop mining test data to train, optimize, test and evaluate the model includes collecting and preprocessing multi-scale natural gas hydrate pressure drop mining test data, building a deep neural network model including one input layer, three hidden layers and one output layer, optimizing the network parameters through a backpropagation algorithm, and evaluating the calculation accuracy of the model using two parameters, namely the coefficient of determination and the average relative error, to obtain a production capacity prediction model with high accuracy and small error, and designing and using a production capacity prediction module to predict the production capacity of natural gas hydrate pressure drop mining tests at different scales and mining parameters.

Description

[Technical field]

[0001] The present invention relates to the field of gas hydrate mining technology, and in particular to an innovative deep neural network-based prediction method for predicting the production capacity of gas hydrate through decompression mining. [Background technology]

[0002] Natural gas hydrate is an ice-like solid compound formed from natural gas and water under high pressure and low temperature conditions. It has many advantages such as wide distribution, abundant reserves, shallow deposits, high energy efficiency, and low pollution, and is regarded as one of the most ideal alternative energy sources in the post-oil era. Among the many methods of extracting natural gas hydrate, the step-down mining method and its improved proposals are widely recognized as one of the most potential and commercializable routes. A deep understanding of the production capacity situation of step-down mining with different mining parameters can help adjust the mining strategy and fully exploit the potential of hydrate resources. Therefore, accurate production capacity prediction plays a vital role in optimizing the mining process and improving efficiency.

[0003] However, natural gas hydrate field mining projects usually require long preparation times, are costly and have large risks, so the field tests currently carried out are relatively limited, and most research work has to rely on laboratory tests. In addition, the diversity and complexity of indoor mining test data makes it difficult for traditional analysis methods to capture the laws and accurately predict production capacity. At the same time, the complex nonlinear relationship between mining parameters and step-down mining production capacity also limits the accuracy and applicability of existing prediction methods, and whether it is a simple linear regression based on empirical formulas or a complex calculation method based on physical models, there are certain limitations. Deep neural network, an advanced machine learning technology, has shown obvious advantages in dealing with complex data and nonlinear relationships. Its core advantage is that it can automatically learn and extract complex features in data by constructing a multi-layer neuron network. Therefore, there is currently an urgent need for a highly efficient and smart production capacity forecasting method for natural gas hydrate step-down mining. Compared with traditional forecasting methods, the present invention proposes a forecasting method based on deep neural network, which has a deeper understanding of the complex relationships between production capacity and various mining parameters and greatly improves the accuracy of forecasting. At the same time, with the ability to automatically learn features, the method can better adapt to a variety of different natural gas hydrate mining environments and does not need to rely on complex artificial parameter input and calculation. Summary of the Invention [Problem to be solved by the invention]

[0004] In order to address the lack of accuracy and adaptability of the traditional natural gas hydrate depressurization mining production capacity prediction method, the present invention proposes an innovative prediction method based on deep neural network, which can realize accurate prediction of production capacity for multi-scale natural gas hydrate depressurization mining tests under different conditions, which helps us adjust the mining strategy, optimize the mining process, and greatly improve the mining efficiency. [Means for solving the problem]

[0005] In order to achieve the above-mentioned goals, the technical solution adopted by the present invention is as follows: A method for predicting natural gas hydrate step-down mining capacity by deep neural network is proposed. The core of this method is to build a five-layer deep neural network model, and use existing multi-scale natural gas hydrate step-down mining test data to train, optimize, test and evaluate the model. In addition, a production capacity prediction module is developed based on this model, and accurate prediction of production capacity of multi-scale natural gas hydrate step-down mining test under various different conditions is realized. This process is not only helpful for adjusting the mining strategy and optimizing the mining process, but also can effectively improve the mining efficiency. The specific implementation steps include: S1: Collect multi-scale natural gas hydrate step-down mining test data, and these test data cover discrete features and continuous features, of which the discrete features include sample scale information and mining well information, and the continuous features include mining parameters and production capacity data. S2: Preprocessing of test data. S3: Build a deep neural network model that includes one input layer, three hidden layers, and one output layer. S4: Using step-down mining test data with different mining parameters at multiple scales, a deep neural network model is trained to accurately capture the complex modes and nonlinear relationships in the test data, and the network parameters are optimized through the backpropagation algorithm. S5: The calculation accuracy of the model is evaluated using two parameters, the coefficient of determination and the average relative error, thereby obtaining a production capacity forecasting model with high accuracy and small error. S6: Equipped with a production capacity prediction model, a production capacity prediction module is designed, and the production capacity prediction module is used to predict the production capacity of natural gas hydrate step-down mining tests at different scales and mining parameters. Furthermore, in step S1, the sample scale information includes three-scale sample scale information of one-dimensional, two-dimensional and three-dimensional; the drilling well information includes drilling well information of horizontal wells, vertical wells and horizontal wells + vertical wells; the drilling parameters include sample volume, porosity, gas hydrate saturation, pre-mining pressure, post-mining pressure, initial mining temperature and pressure drop rate; and the production capacity data includes cumulative gas production volume and average gas production rate. Furthermore, the method of data pre-processing in step S2 is as follows: S2.1: Clean the collected data, remove potential outliers and noise, and obtain normative test data. S2.2: Sampling is performed using stratified random sampling with a floating range of 5% for each mining parameter and expanding the test sample size, and the total number of samples is N. S2.3: Normalization is performed on the test data to obtain nine types of features, and the normalized dataset is divided into a training set, a validation set, and a test set. S2.4: Split the normalized dataset into a 70% training set, a 15% validation set, and a 15% test set according to the split ratio. Further: the test data normalization processing method in step 2.3 is as follows: S2.3.1: For discrete features, formally normalize sample scale information and well information using a unique thermal code: Sample scale information: 1st dimension = (1,0,0), 2nd dimension = (0,1,0), 3rd dimension = (0,0,1); Mining well information: horizontal well = (1,0,0), vertical well = (0,1,0), horizontal well + vertical well = (0,0,1); S2.3.2: For continuous features, a normalization process is performed by scaling the mining parameter data to the range [0,1]. The formula is:

[0006]

number

[0007] Here, x is the seven mining parameters in step S1, min(x) is the minimum value of the corresponding mining parameters in all test samples extracted in step S2.2, max(x) is the maximum value of the corresponding mining parameters in all test samples extracted in step S2.2, and xn is the normalized value. Furthermore, step S3 specifically includes the following steps: S3.1: Build a 5-layer deep neural network model, including one input layer, three hidden layers, and one output layer. S3.2: The number of input layer nodes is nine, which fully correspond to a total of nine types of features, namely, sample scale information, mining well information, and seven mining parameters. S3.3: There are two nodes in the output layer, each corresponding to the two types of production capacity data in step S1. S3.4: The number of nodes per hidden layer in the deep neural network model is 60. Furthermore, step S4 specifically includes the following steps: S4.1: Initialize the model deviations to zero and initialize the model weights randomly using a normal distribution, with the formula: W~N(0,0.012) In the formula, W is the weight matrix and N(0,0.012) is the normal distribution with mean 0 and standard deviation 0.01. S4.2: Select Leaky ReLU as the activation function. The formula is:

[0008]

number

[0009]

number

[0010]

number

[0011]

number

[0012]

number

[0013] Further, the detailed procedure of step S6 is as follows: In step S5, a production capacity prediction module is designed by installing a production capacity prediction model with high fitting accuracy and small error, and gas hydrate step-down mining test data is obtained in real time. The test data is input into the production capacity prediction module according to the normalized data format in S2.3 to obtain a production capacity prediction value, thereby providing accurate multi-scale gas hydrate step-down mining test production capacity information in real time, and further providing a data reference for field application. Effect of the Invention

[0014] In relation to all the above technical aspects, the beneficial effects of the present invention include: (1) The production capacity prediction method for natural gas hydrate step-down mining based on deep neural network provided by the present invention significantly reduces the large amount of human and material resources required for traditional natural gas hydrate step-down mining tests. In addition, the automatic learning and pattern recognition capabilities of the deep learning model effectively reduce the burden of laboratory work and data analysis, providing researchers with a more efficient experimental means. (2) The production capacity prediction method for decompression mining based on deep neural network for natural gas hydrate decompression mining can provide assistance for the optimal mining strategy for decompression mining projects, and the generalization ability and real-time nature of the model enable production capacity prediction information to be obtained in an instant, so as to adjust and optimize decompression mining operations, improve production efficiency, and reduce costs. (3) The production capacity prediction method for natural gas hydrate step-down mining based on deep neural network provided by the present invention has higher prediction accuracy than traditional methods, and the model can better capture the complex relationships in natural gas hydrate step-down mining tests, providing more reliable production capacity prediction support for step-down mining projects, and being helpful for scientific decision-making and resource planning. [Brief description of the drawings]

[0015] [Figure 1]1 is a flowchart of a method for predicting production capacity of natural gas hydrate decompression mining based on a deep neural network. [Diagram 2] FIG. 10 is a comparison diagram between predicted and actual cumulative gas production values ​​in a partial gas hydrate step-down mining test based on a deep neural network provided by an embodiment of the present invention. [Diagram 3] 1 is a graph comparing partial average gas generation rate prediction values ​​and actual values ​​in a natural gas hydrate step-down mining test based on a deep neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] In the following, the construction process of the embodiments of the present invention will be described in detail in conjunction with the accompanying drawings, so as to make the advantages and features of the present invention easier to understand for those skilled in the art, and to define the protection scope of the present invention more clearly and unambiguously. As shown in FIG. 1, the present invention is an embodiment of a method for predicting the production capacity of natural gas hydrate step-down mining based on deep neural network, which involves collecting and pre-processing engineered multi-scale natural gas hydrate step-down mining test data, constructing a deep neural network model including one input layer, three hidden layers, and one output layer; training the deep neural network model through the step-down mining test data of three different scales and different mining parameters to accurately capture the complex modes and nonlinear relationships in the test data, and optimizing the network parameters through the backpropagation algorithm; evaluating the calculation accuracy of the model using two parameters, the coefficient of determination and the average relative error; designing a production capacity prediction module to predict the production capacity of natural gas hydrate step-down mining test at different scales and mining parameters. The main steps of this method are as follows: TIFF0007629242000012.tif42166S2: Data preprocessing S2.1: Clean the collected data, remove potential outliers and noise, obtain standard test data, and merge the data within the error tolerance range for data fluctuations due to test operations, etc., as shown in Table 1.

[0017] [Table 1] S2.2: 5% of each mining parameter is set as the floating range, the number of test samples is expanded, a total of 104976 sets of test samples are obtained, and stratified random sampling method is adopted to perform sampling, the total number of samples is 3200; S2.3: Data normalization S2.3.1: For discrete features, formally normalize sample scale information and well information using a unique thermal code: Sample scale information: 1st dimension = (1,0,0), 2nd dimension = (0,1,0), 3rd dimension = (0,0,1); Mining well information: horizontal well = (1,0,0), vertical well = (0,1,0), horizontal well + vertical well = (0,0,1); S2.3.2: For continuous features, a normalization process is performed by scaling the mining parameter data to the range [0,1]. The formula is:

[0018]

number

[0019]

number

[0020]

number

[0021]

number

[0022]

number

[0023]

number

[0024]

number

[0025] Of course, the above description does not limit the present invention, and the scope of application of the present invention is not limited to the above specific examples. Any modifications, variations, additions or replacements made by experts in the technical field within the scope of the basic principles and essence of the invention shall be considered as part of the protection scope of the invention.

Claims

1. A method for predicting production capacity of decompression mining of natural gas hydrate based on a deep neural network, comprising: Steps below: S1: Collect multi-scale natural gas hydrate decompression mining experimental data, covering discrete and continuous features, the discrete features include sample scale information and mining well information, and the continuous features include mining parameters and production capacity data; In the deep neural network-based natural gas hydrate decompression mining production capacity prediction method described above, in step S1, the sample scale information includes one-dimensional, two-dimensional and three-dimensional three-scale sample scale information, the mining well information includes horizontal well, vertical well, and horizontal well + vertical well mining well information, the mining parameters include sample volume, porosity, gas hydrate saturation, pre-mining pressure, post-mining pressure, initial mining temperature, and depressurization rate, and the production capacity data includes cumulative gas production volume and average gas production rate; S2: Pre-processing of test data; the next step is: S2.1: Clean the collected data, remove potential outliers and noise, and obtain normative test data; S2.2: Sampling is carried out using stratified random sampling with 5% of each mining parameter as the floating range and expanding the number of test samples, and the total number of samples is N; S2.3: Normalization is performed on the test data to obtain nine features, and the normalized dataset is divided into a training set, a validation set, and a test set: S2.3.1: For discrete features, sample scale information and well information are formally normalized using one-hot coding; Sample scale information: 1st dimension = (1,0,0), 2nd dimension = (0,1,0), 3rd dimension = (0,0,1); Mining well information: horizontal well = (1,0,0), vertical well = (0,1,0), horizontal well + vertical well = (0,0,1); S2.3.2: For continuous features, a normalization process is performed by scaling the mining parameter data to the range [0,1]; the formula is: [0010] Here, x is the seven mining parameters in step S1, and min(x) is the step The minimum value of the corresponding mining parameter in all test samples extracted in S2.2, m ax(x) is the corresponding mining pattern for all test samples extracted in step S2.

2. The maximum value of the parameter, xn is the normalized value; S2.4: Split the normalized dataset into 70% training set, 15% validation set, and 15% test set according to the split ratio; S3: Build a deep neural network model that includes one input layer, three hidden layers, and one output layer; as follows: S3.1: Construct a five-layer deep neural network model, including one input layer, three hidden layers, and one output layer; S3.2: The number of input layer nodes is nine, which fully correspond to nine kinds of features, including sample scale information, mining well information, and seven mining parameters; S3.3: There are two nodes in the output layer, which correspond to the two types of production capacity data in step S1; S3.4: The number of nodes per hidden layer in the deep neural network model is 60; S4: Using step-down mining test data of different mining parameters at multiple scales, train a deep neural network model to accurately capture the complex modes and nonlinear relationships in the test data, and optimize the network parameters through the backpropagation algorithm; S4.1: Initialize the model deviations to zero and initialize the model weights randomly using a normal distribution, the formula is: [0025] In the formula, W is the weighting matrix and N(0,0.012) has a mean of 0 and a standard deviation of 0. 01 is a normal distribution; S4.2: Select Leaky ReLU as the activation function, the formula is: [0030] In the formula, α is 0.01, and Xi are the nine features in step S2.3; S4.3: Use MSE as the loss function, measuring the squared difference between the production capacity forecast and the actual value obtained using the training set of step S2.4; the formula is: [0045] 【number】 S4.4: Using the chain rule, starting from the output layer, inversely calculate the gradient of the loss against the weights and deviations of the model parameters, and then use the batch gradient descent optimization algorithm to realize the parameter update, as follows: [0050] In the formula, ▽ θ J(θ) is the gradient calculation value for the model parameters of the loss function, and θ is the depth The parameters of the neural network model, β, are the learning rate, and the initial value is 0.

01. is; S4.5: An L2 regularization technique is used to penalize the model weights to prevent them from being too large and causing the model to over-fit the training data; the formula is: [006] In the formula, JL2 is the loss function after regularization by adding L2, and λ is the strength of regularization. , the initial value is 0.1, and Σi∥ωi∥2 is the sum of squares of the ownership weights of the model; S4.6: Evaluate the performance of the model on the validation set and adjust the learning rate and regularization strength based on the representation on the validation set; S4.7: Repeat steps S4.3 to S4.6 as many times as necessary until the model performs satisfactorily; S5: Evaluate the calculation accuracy of the model using two parameters: the coefficient of determination and the average relative error; S6: Equipping the production capacity prediction model, designing a production capacity prediction module, and using the production capacity prediction module to predict the production capacity of the natural gas hydrate step-down mining test at different scales and mining parameters; have A method for predicting production capacity of decompression mining of natural gas hydrate based on a deep neural network, comprising:

2. In step S5, the following procedure is performed: The test set was used to evaluate the generalization performance of the trained and optimized model to unknown data, and the coefficient of determination R ２ The calculation accuracy of the model is evaluated using the criteria that the mean relative error E is at least 0.8 and the mean relative error E is less than 10%, and a model with high fitting accuracy and low error is obtained; the coefficient of determination R ２ and the expression for the mean relative error E is: [0070] 【number】 R ２ The closer to 1, the higher the accuracy of the model's fit, and the closer to 0, the smaller the error between the model's predicted production capacity value and the actual production capacity value. have The method for predicting production capacity of natural gas hydrate under reduced pressure based on a deep neural network as claimed in claim 1.

3. In step S6, the following procedure is performed: In step S5, a production capacity prediction module is designed by installing a model with high fitting accuracy and small error, and the gas hydrate drop-pressure mining test data is obtained in real time, and input into the production capacity prediction module according to the data format after normalization in S2.3 to obtain a production capacity prediction value, thereby providing accurate multi-scale gas hydrate drop-pressure mining test production capacity information in real time, and further providing data reference for field application. have The method for predicting production capacity of natural gas hydrate under reduced pressure based on a deep neural network as claimed in claim 2.

Images