A hydrate decomposition rate prediction method based on a fuzzy neural network

Abstract

This invention discloses a method for predicting the decomposition rate of hydrates based on a fuzzy neural network, comprising: S1, collecting multi-source parameter data affecting the decomposition rate and performing data preprocessing to obtain standardized input feature samples; S2, filtering parameters highly correlated with the decomposition rate through feature extraction and constructing training and test datasets; S3, establishing a fuzzy neural network model using an adaptive network fuzzy inference system and completing the model structure construction; S4, training the model using the training dataset, optimizing parameters and iteratively updating them using the backpropagation algorithm combined with gradient descent, and evaluating model performance through cross-validation; S5, inputting the test dataset or new operating condition data into the model for testing and outputting the decomposition rate prediction result; S6, visualizing and analyzing the prediction results and key parameters. This invention possesses the advantages of nonlinear modeling and strong adaptive capabilities, significantly improving prediction accuracy and engineering practicality.

Description

Technical Field

[0001] This invention relates to the field of industrial process modeling and data analysis technology, and in particular to a method for predicting the decomposition rate of hydrates based on fuzzy neural networks. This method is applicable to various scenarios including multi-source parameter data acquisition, intelligent feature extraction, decomposition rate modeling and prediction, and industrial production process optimization during hydrate decomposition. Background Technology Hydrates are an important energy resource, and accurate prediction of their decomposition rate is crucial for natural gas extraction, transportation safety, and improving energy efficiency. Traditional methods for predicting hydrate decomposition rates mainly rely on physicochemical models or empirical formulas, such as mechanistic models based on theories of mass conservation, energy conservation, or diffusion reaction kinetics. These methods typically assume uniform system parameter distribution and that the decomposition process is controlled by a single dominant factor, and they have limited consideration of multi-source disturbances, process variations, and nonlinear coupling effects in actual operating conditions. As the development of natural gas hydrate resources continues to deepen, the operating conditions at engineering sites are becoming increasingly complex, resulting in low prediction accuracy of traditional models in practical applications, making it difficult to adapt to complex decomposition processes involving multiple variables, strong coupling, and dynamic changes.

[0002] With the improvement of industrial informatization and automation, more and more process and environmental parameter data can be collected on-site, such as porosity, lithology, gas composition, and various environmental disturbance factors. This multi-source data contains rich process information, but traditional statistical regression, single neural networks, or algorithms based on linear assumptions are insufficient to fully explore and express the complex nonlinear relationships within it. Furthermore, in actual operating conditions, data often contains noise, missing values, outliers, and ambiguity and uncertainty between parameters, further increasing the difficulty of accurately predicting decomposition rates. In recent years, artificial intelligence and data-driven modeling methods have been gradually applied to the field of industrial process prediction. Although some studies have attempted to use machine learning algorithms, such as support vector machines, decision trees, and ordinary neural networks, to optimize the modeling and prediction of hydrate decomposition rates, these methods often lack the ability to handle data ambiguity and uncertainty, and their modeling effect on complex coupling relationships between input features is limited. Especially when facing high-dimensional, nonlinear, and multi-perturbation real-world application scenarios, the generalization ability and prediction accuracy of these models still have considerable room for improvement.

[0003] Existing methods for predicting hydrate decomposition rates suffer from insufficient levels of process intelligence and automation. Data preprocessing, feature extraction, model training, and optimization largely rely on human experience and trial and error, which can easily lead to feature redundancy, information loss, and unstable model performance. Furthermore, existing methods often lack data visualization, risk warning, and decision support functions closely integrated with actual engineering needs, hindering engineers from quickly understanding prediction results and optimizing processes.

[0004] Therefore, existing technologies for intelligent prediction of hydrate decomposition rates suffer from the following main shortcomings: First, they lack effective means for processing multi-source, multi-dimensional, nonlinear, and fuzzy uncertain information; second, their model generalization ability and prediction accuracy are limited, making it difficult to adapt to complex operating conditions and dynamic changes; and third, the level of intelligence and automation in data-driven processes and engineering applications is insufficient. To address these issues, there is an urgent need for an intelligent prediction method for hydrate decomposition rates that can integrate multi-source data, possess strong nonlinear modeling and adaptive capabilities, take into account data fuzziness and uncertainty, and achieve high-precision prediction and visualization analysis, in order to meet the practical needs of modern energy engineering for intelligent prediction and process optimization of hydrate decomposition rates. Summary of the Invention

[0005] One objective of this invention is to propose a hydrate decomposition rate prediction method based on a fuzzy neural network. This invention fully utilizes the Adaptive Network Fuzzy Inference System (ANFIS) and data analysis technology, and describes in detail the entire process from multi-source parameter data acquisition, feature extraction, model construction, training optimization to prediction analysis. It has the advantages of strong nonlinear modeling ability and adaptability, high prediction accuracy, and result visualization.

[0006] A method for predicting hydrate decomposition rate based on a fuzzy neural network according to an embodiment of the present invention includes the following steps: S1. Collect multi-source parameter data that affect the hydrate decomposition rate, fill in missing values, remove outliers and normalize the data, and perform quality checks. If the check passes, a standardized input feature sample is obtained. If the data is abnormal, it is fed back to the data acquisition stage for re-acquisition. S2. Extract features from the input samples, select input feature parameters that are highly correlated with the hydrate decomposition rate, and select appropriate input parameters and target outputs based on the extracted feature parameters and corresponding decomposition rate data to construct training and testing datasets. S3. Construct a fuzzy neural network model using an Adaptive Network Fuzzy Inference System (ANFIS), configure membership functions based on input features, establish fuzzy rules, set network parameters, and complete the model structure construction. S4. Train the fuzzy neural network model using the training dataset, optimize the parameters and iteratively update them using the backpropagation algorithm combined with the gradient descent method, evaluate the model performance through cross-validation, and obtain the training-completed prediction model if the model performance meets the standard; otherwise, feed back into the model training and parameter optimization. S5. Input the test dataset or new working condition data into the trained fuzzy neural network model, output the predicted result of hydrate decomposition rate, and establish a database corresponding to the input features and the predicted results; S6. Visualize and analyze the predicted results and their corresponding input features, output the main predicted results and changing trends, and use them to assist in engineering analysis and decision-making.

[0007] Optionally, step S1 includes: S11, The multi-source parameter data includes temperature (K) Pressure (MPa), porosity (Dimensionless), lithological parameter L, gas composition (mole fraction), environmental parameters Where L represents the classification code, Indicates the gas component number. Indicates the environmental parameter number; S12, Utilization Indicates the first The parameter in the first... The original observations under each sample, of which, Numbering of multi-source parameters, The sample number is... Indicates temperature In the The values ​​for each sample Indicates pressure in the first The values ​​for each sample are calculated, and so on.

[0008] S13. Remove any missing data. The mean imputation method is used for completion, and the completion formula is as follows: in, Indicates the first Each parameter actually represents the total number of samples of observed values. Indicates the first Each parameter has a sample number corresponding to an observed value; in practice, when it is necessary to... When imputing missing values, all samples with historical observations of that parameter are numbered. The missing data is filled in by iterating through the data and calculating the average. For example, suppose the third parameter (such as porosity) has observed values ​​in samples 1, 3, 4, and 6, but is missing in samples 2 and 5, then , , Take 1, 3, 4, 6, and accept missing values. and Need to use , , , The mean is used to fill in the gaps.

[0009] S14. Remove outliers from the original data using... In principle, if If the mean is an outlier, it is considered an outlier. and standard deviation They are respectively: S15. After removing outliers, the data is normalized using the min-max normalization method, with the following formula: in, For the first The minimum value of each parameter. For the first The maximum value of each parameter. These are the normalized, standardized input feature samples; S16. Perform a quality check on the above data processing results. If the quality check passes, a standardized input feature sample is obtained. If data anomalies are found or the data cannot meet the modeling requirements, the data is fed back to the data acquisition stage for re-acquisition and processing until the data quality meets the subsequent modeling requirements.

[0010] Optionally, step S2 includes: S21. Use Pearson correlation coefficient analysis to standardize input features. With hydrate decomposition rate The correlation between them, The correlation coefficient is calculated using the formula for the decomposition rate, which is obtained by measuring changes in gas production or mass at the experimental site. in, For the first The mean of each feature, This represents the average decomposition rate. S22. Based on the correlation analysis results, select characteristic parameters that are highly correlated with the decomposition rate. By removing parameters with low relevance or redundancy, an optimized feature subset is obtained. ; S23. Optimize the feature subset Principal component analysis is performed to reduce the dimensionality of the feature parameters to principal component features. The formula for calculating principal components is: in, This indicates the number of feature parameters after optimization. Represents the principal component coefficients. Indicates the principal component feature number and For example, after feature selection, there are ultimately 7 input parameters used for modeling. =7, during principal component analysis, each principal component All A linear combination of features; S24. Based on the extracted principal component features and their corresponding decomposition rates, select appropriate input parameters and target outputs to construct training and testing datasets.

[0011] Optionally, step S3 includes: S31. An adaptive network fuzzy inference system is used to construct a fuzzy neural network model, which includes an input layer, a fuzzification layer, a fuzzy rule layer, a defuzzification layer, and an output layer. S32, Principal component features for each input Configure Gaussian membership functions The expression is: in, Number the membership function and , For the first The number of membership functions of each principal component. For the first The first principal component Membership function center, For the corresponding width; S33. Construct a fuzzy rule layer based on the fuzzification results of the input principal components. The fuzzy rules are represented as follows: in, For the first Rule number 1 Fuzzy subsets of principal components For the rule consequent function, Output the rules; S34. In the defuzzification layer, a weighted average is performed on the outputs of all fuzzy rules to finally output the predicted value. The calculation formula is as follows: in, To represent the total number of fuzzy rules, For the first The activation strength of the rule; S35. Configure all network parameters, including , The model structure is built by defining fuzzy rule parameters and output layer weight parameters.

[0012] Optionally, step S4 includes: S41. Principal component features of the training dataset and corresponding decomposition rate The samples are input into the fuzzy neural network model. S42. Calculate the predicted output for each sample using the MSE loss function. With actual decomposition rate Mean squared error between, loss function The formula is: , in, This represents the total number of samples. The training is completed when the loss function converges or the training termination condition is met. S43. Combine the backpropagation algorithm with gradient descent to calculate the gradient of the loss function with respect to the network weights and membership function parameters, and iteratively update the parameters until the loss function converges or the training termination condition is met. S44. Evaluate the model's performance through cross-validation and calculate the coefficient of determination for the prediction accuracy index. The formula is: , in, This represents the average actual decomposition rate of the sample. S45. If the performance evaluation results meet the standards, the trained fuzzy neural network model is obtained; otherwise, the process is fed back to step S43 to continue model training and parameter optimization until the model achieves the required prediction accuracy and generalization ability.

[0013] Optionally, step S5 includes: S51, The new working condition data refers to multi-source parameter data that did not appear during the model training phase but was collected in actual application; S52, Principal component features of the test dataset or new working condition data. The input is fed into the trained fuzzy neural network model to obtain the decomposition rate prediction result. ,in, , To predict the total number of samples; S53. For each set of input features The predicted value is calculated using the weighted average formula of the deblurring layer. : S54. Output prediction results in batches for all input conditions. ; S55. Record the features of each set of inputs. With the corresponding predicted decomposition rate To form a mapping relationship database .

[0014] Optionally, step S6 includes: S61. Organize the forecast results and their corresponding input features ,in, And generate the corresponding relationship table; S62. Use visualization tools to analyze input features. With predicted decomposition rate Multidimensional visualization analysis of relationships is performed, including two-dimensional scatter plots, three-dimensional surface plots, and heat maps, to show the trend of the influence of features on the decomposition rate; S63. Plot the response curve of the decomposition rate as a function of key input features, and analyze the functional relationship between the prediction results and the feature changes. S64. Regarding the prediction results Perform statistical analysis and calculate the mean. and standard deviation The calculation formulas are as follows: , S65. Based on the visualization and statistical results, output the main prediction conclusions and trends, and form a comprehensive report to assist in engineering analysis and decision-making.

[0015] Optionally, the membership function type of the fuzzy neural network model can be Gaussian, trigonometric, or trapezoidal, with the specific type selected according to the modeling accuracy requirements.

[0016] Optionally, the decomposition rate prediction result database supports the retrieval, comparison, and batch export of historical prediction results, facilitating long-term tracking and engineering optimization analysis.

[0017] Beneficial effects This invention addresses the shortcomings of existing hydrate decomposition rate prediction methods in handling multi-source nonlinear data, model generalization ability, and intelligent processes, and proposes a hydrate decomposition rate prediction method based on fuzzy neural networks.

[0018] First, this invention fully integrates multi-source process and environmental data, including temperature, pressure, porosity, lithology, gas composition, and various environmental parameters. Through automated data preprocessing, feature extraction, and principal component analysis, it effectively reduces the impact of redundant information and noise on model prediction performance, improving the quality and representativeness of the input data. Second, the fuzzy neural network model established using the Adaptive Network Fuzzy Inference System (ANFIS) can fully capture the complex nonlinear relationships and fuzzy uncertainties between process parameters, achieving high-precision modeling and prediction of hydrate decomposition rates. This significantly outperforms traditional physical models and ordinary machine learning methods in terms of adaptability and generalization under complex operating conditions. Furthermore, this invention effectively improves the model's stability and practical application value through cross-validation, parameter optimization, and batch prediction. Combined with visualization analysis tools, it intuitively displays the relationship between prediction results and key process parameters, providing engineers with intuitive data support and decision-making basis, enabling intelligent monitoring and optimization of the hydrate decomposition process.

[0019] In summary, this invention not only improves the accuracy and reliability of hydrate decomposition rate prediction, but also promotes the improvement of intelligent, automated and refined management of industrial processes. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a hydrate decomposition rate prediction method based on a fuzzy neural network according to the present invention; Figure 2 This is a schematic diagram of data preprocessing and feature optimization in this invention; Figure 3 This is a schematic diagram illustrating the application and feedback optimization of the decomposition rate prediction results of this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-3 A method for predicting hydrate decomposition rate based on fuzzy neural network, comprising: S1. Collect multi-source parameter data that affect the hydrate decomposition rate, and perform missing value imputation, outlier removal and normalization on the data to obtain standardized input samples; S2. Extract features from the input samples, select input feature parameters that are highly correlated with the hydrate decomposition rate, and select appropriate input parameters and target outputs based on the extracted feature parameters and corresponding decomposition rate data to construct training and testing datasets. S3. Construct a fuzzy neural network model using an Adaptive Network Fuzzy Inference System (ANFIS), configure membership functions based on input features, establish fuzzy rules, set network parameters, and complete the model structure construction. S4. Train the fuzzy neural network model using the training dataset, optimize the parameters and iteratively update them using the backpropagation algorithm combined with the gradient descent method, evaluate the model performance through cross-validation, and obtain the trained prediction model if the model performance meets the standard; otherwise, feed it back into the model training and parameter optimization. S5. Input the test dataset or new working condition data into the trained fuzzy neural network model, output the prediction result of the hydrate decomposition rate, and establish a database corresponding to the input features and prediction results. S6. Visualize and analyze the prediction results and their corresponding input features, and output the main prediction results and trends to assist engineering analysis and decision-making.

[0023] In this embodiment, step S1 includes: S11, The multi-source parameter data includes temperature (K) Pressure (MPa), porosity (Dimensionless), lithological parameter L, gas composition (mole fraction), environmental parameters Where L represents the classification code, Indicates the gas component number. Indicates the environmental parameter number; S12, Utilization Indicates the first The parameter in the first... The original observations under each sample, of which, Numbering of multi-source parameters, The sample number is... Indicates temperature In the The values ​​for each sample Indicates pressure in the first The values ​​for each sample are calculated, and so on.

[0024] S13. Remove any missing data. The mean imputation method is used for completion, and the completion formula is as follows: in, Indicates the first Each parameter actually represents the total number of samples of observed values. Indicates the first Each parameter has a sample number corresponding to an observed value; in practice, when it is necessary to... When imputing missing values, all samples with historical observations of that parameter are numbered. The missing data is filled in by iterating through the data and calculating the average. For example, suppose the third parameter (such as porosity) has observed values ​​in samples 1, 3, 4, and 6, but is missing in samples 2 and 5, then , , Take 1, 3, 4, 6, and accept missing values. and Need to use , , , The mean is used to fill in the gaps.

[0025] S14. Remove outliers from the original data using... In principle, if If the mean is an outlier, it is considered an outlier. and standard deviation They are respectively: S15. After removing outliers, the data is normalized using the min-max normalization method, with the following formula: in, For the first The minimum value of each parameter. For the first The maximum value of each parameter. These are the normalized, standardized input feature samples; S16. Perform a quality check on the above data processing results. If the quality check passes, a standardized input feature sample is obtained. If data anomalies are found or the data cannot meet the modeling requirements, the data is fed back to the data acquisition stage for re-acquisition and processing until the data quality meets the subsequent modeling requirements.

[0026] In this embodiment, step S2 includes: S21. Use Pearson correlation coefficient analysis to standardize input features. With hydrate decomposition rate The correlation between them, The correlation coefficient is calculated using the formula for the decomposition rate, which is obtained by measuring changes in gas production or mass at the experimental site. in, For the first The mean of each feature, This represents the average decomposition rate. S22. Based on the correlation analysis results, select characteristic parameters that are highly correlated with the decomposition rate. By removing parameters with low relevance or redundancy, an optimized feature subset is obtained. ; S23. Optimize the feature subset Principal component analysis is performed to reduce the dimensionality of the feature parameters to principal component features. The formula for calculating principal components is: in, This indicates the number of feature parameters after optimization. Represents the principal component coefficients. Indicates the principal component feature number and For example, after feature selection, there are ultimately 7 input parameters used for modeling. =7, during principal component analysis, each principal component All A linear combination of features; S24. Based on the extracted principal component features and their corresponding decomposition rates, select appropriate input parameters and target outputs to construct training and testing datasets.

[0027] In this embodiment, step S3 includes: S31. An adaptive network fuzzy inference system is used to construct a fuzzy neural network model, which includes an input layer, a fuzzification layer, a fuzzy rule layer, a defuzzification layer, and an output layer. S32, Principal component features for each input Configure Gaussian membership functions The expression is: in, Number the membership function and , For the first The number of membership functions of each principal component. For the first The first principal component Membership function center, For the corresponding width; S33. Construct a fuzzy rule layer based on the fuzzification results of the input principal components. The fuzzy rules are represented as follows: in, For the first Rule number 1 Fuzzy subsets of principal components For the rule consequent function, Output the rules; S34. In the defuzzification layer, a weighted average is performed on the outputs of all fuzzy rules to finally output the predicted value. The calculation formula is as follows: in, To represent the total number of fuzzy rules, For the first The activation strength of the rule; S35. Configure all network parameters, including , The model structure is built by defining fuzzy rule parameters and output layer weight parameters.

[0028] In this embodiment, step S4 includes: S41. Principal component features of the training dataset and corresponding decomposition rate The samples are input into the fuzzy neural network model. S42. Calculate the predicted output for each sample using the MSE loss function. With actual decomposition rate Mean squared error between, loss function The formula is: , in, This represents the total number of samples. The training is completed when the loss function converges or the training termination condition is met. S43. Combine the backpropagation algorithm with gradient descent to calculate the gradient of the loss function with respect to the network weights and membership function parameters, and iteratively update the parameters until the loss function converges or the training termination condition is met. S44. Evaluate the model's performance through cross-validation and calculate the coefficient of determination for the prediction accuracy index. The formula is: , in, This represents the average actual decomposition rate of the sample. S45. If the performance evaluation results meet the standards, the trained fuzzy neural network model is obtained; otherwise, the process is fed back to step S43 to continue model training and parameter optimization until the model achieves the required prediction accuracy and generalization ability.

[0029] In this embodiment, step S5 includes: S51, The new working condition data refers to multi-source parameter data that did not appear during the model training phase but was collected in actual application; S52, Principal component features of the test dataset or new working condition data. The input is fed into the trained fuzzy neural network model to obtain the decomposition rate prediction result. ,in, , To predict the total number of samples; S53. For each set of input features The predicted value is calculated using the weighted average formula of the deblurring layer. : S54. Output prediction results in batches for all input conditions. ; S55. Record the features of each set of inputs. With the corresponding predicted decomposition rate To form a mapping relationship database .

[0030] In this embodiment, step S6 includes: S61. Organize the forecast results and their corresponding input features ,in, And generate the corresponding relationship table; S62. Use visualization tools to analyze input features. With predicted decomposition rate Multidimensional visualization analysis of relationships is performed, including two-dimensional scatter plots, three-dimensional surface plots, and heat maps, to show the trend of the influence of features on the decomposition rate; S63. Plot the response curve of the decomposition rate as a function of key input features, and analyze the functional relationship between the prediction results and the feature changes. S64. Regarding the prediction results Perform statistical analysis and calculate the mean. and standard deviation The calculation formulas are as follows: , S65. Based on the visualization and statistical results, output the main prediction conclusions and trends, and form a comprehensive report to assist in engineering analysis and decision-making.

[0031] In this embodiment, the membership function type of the fuzzy neural network model is Gaussian, trigonometric, or trapezoidal, and the specific type is selected according to the modeling accuracy requirements.

[0032] In this embodiment, the decomposition rate prediction result database supports the retrieval, comparison, and batch export of historical prediction results, which facilitates long-term tracking and engineering optimization analysis.

[0033] Example 1: To verify the feasibility of this invention in practical engineering, it was applied to a natural gas hydrate decomposition experiment and pilot production scenario. The experimental site had complex geological conditions; the temperature of the hydrate-bearing layer fluctuated between 275K and 285K, the pressure ranged from 10MPa to 15MPa, the porosity was between 0.18 and 0.34, the lithology included sandstone and mudstone, and the gas components were mainly methane, ethane, and a small amount of carbon dioxide. Multiple types of high-precision sensors and automated sampling equipment were deployed on-site, enabling real-time online acquisition and data storage of multiple parameters such as temperature, pressure, porosity, lithology, and gas composition.

[0034] In this scenario, traditional decomposition rate prediction methods mainly rely on empirical formulas or multiple linear regression models. However, due to the influence of multiple factors such as temperature fluctuations, pressure changes, lithological heterogeneity, and gas composition variations in the field, traditional methods exhibit significant prediction errors when encountering sudden parameter changes or complex coupling. For some samples, the predicted values ​​deviate from the actual measured values ​​by as much as 30%, making it difficult to provide stable and reliable support for refined production management and process optimization.

[0035] The application process of the method of this invention in this scenario is as follows: First, the automated monitoring system continuously collects temperature, pressure, porosity, lithology, gas composition, and other environmental parameters during the hydrate decomposition process. All raw data are preprocessed using the method of this invention, including mean imputation of missing values, outlier removal, and min-max normalization to ensure that the input parameters are complete and dimensionless. Subsequently, Pearson correlation coefficient analysis and principal component analysis are used to automatically select feature variables highly correlated with the decomposition rate from the multi-source parameters, such as temperature, pressure, and methane mole fraction, effectively reducing the impact of redundant features on the model's generalization ability.

[0036] In the modeling phase, standardized multi-source parameters and historical experimental measured decomposition rate data were selected, and a fuzzy neural network model was constructed based on the Adaptive Network Fuzzy Inference System (ANFIS). Taking 350 sets of field experimental data collected in the spring and summer of a certain year as an example, they were divided into training and test sets at a ratio of 70% and 30%, respectively. Each sample included temperature, pressure, porosity, lithology, molar fractions of methane, ethane, and carbon dioxide, as well as the measured decomposition rate. During training, backpropagation and cross-validation methods were used to optimize the model parameters, resulting in a decomposition rate prediction model with strong generalization ability and high prediction accuracy.

[0037] The decomposition rate predicted by the method of this invention, combined with data visualization tools, can automatically generate response curves between the decomposition rate and key parameters such as temperature, pressure, and composition, assisting engineers in quickly identifying sensitive factors affecting the decomposition process. In actual production, the method of this invention has been repeatedly applied for intelligent prediction and real-time monitoring of the decomposition rate, promptly detecting and issuing warnings of decomposition rate anomalies caused by sudden changes in gas composition, lithological fluctuations, etc. Through dynamic optimization of process parameters, no production anomalies or equipment risk events caused by prediction errors have occurred on-site, significantly improving the safety and efficiency of the hydrate decomposition process.

[0038] In this embodiment, ten sets of typical hydrate decomposition experimental sample data were collected, covering various complex working conditions with changes in temperature, pressure, porosity, lithology, and major gas components. Each set of data includes the decomposition rate measured in the field and the prediction results obtained using the method of this invention and the traditional multiple linear regression method, respectively. Specific comparative data on hydrate decomposition rate predictions are shown in Table 1: Table 1: Comparison of Predicted Decomposition Rates for Selected Samples in the Hydrate Decomposition Experiment The data in the table reflect the variation of decomposition rates under different temperatures, pressures, porosities, lithologies, and gas compositions. It can be seen that the prediction method based on fuzzy neural networks in this invention closely matches the measured values ​​on most samples, with an average absolute error of only 0.047 mol / (kg·h), significantly better than the average absolute error of 0.128 mol / (kg·h) of the traditional method. For example, under conditions of 278.6 K, 12.3 MPa, a porosity of 0.21, and a methane content of 94.2%, the measured decomposition rate is 0.512 mol / (kg·h), the predicted value of this invention is 0.505 mol / (kg·h), while the predicted value of the traditional method is only 0.392 mol / (kg·h). Under extreme conditions of high temperature, high pressure, and complex compositions, the method of this invention also exhibits high reliability and stability, effectively improving the accuracy and engineering applicability of decomposition rate prediction. The above data fully demonstrate that the method of this invention can significantly improve the intelligent prediction capability of hydrate decomposition rates under complex conditions.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the decomposition rate of hydrates based on a fuzzy neural network, characterized in that, Includes the following steps: S1. Collect multi-source parameter data that affect the hydrate decomposition rate, fill in missing values, remove outliers and normalize the data, and perform quality checks. If the check passes, generate standardized input feature samples. If the data is abnormal, it is fed back to the data acquisition stage for re-acquisition. S2. Extract features from the input samples, select input feature parameters that are highly correlated with the hydrate decomposition rate, and select appropriate input parameters and target outputs based on the extracted feature parameters and corresponding decomposition rate data to construct training and testing datasets. S3. Construct a fuzzy neural network model using an Adaptive Network Fuzzy Inference System (ANFIS), configure membership functions based on input features, establish fuzzy rules, set network parameters, and complete the model structure construction. S4. Train the fuzzy neural network model using the training dataset, optimize the parameters and iteratively update them using the backpropagation algorithm combined with the gradient descent method, evaluate the model performance through cross-validation, and obtain the training-completed prediction model if the model performance meets the standard; otherwise, feed back into the model training and parameter optimization. S5. Input the test dataset or new working condition data into the trained fuzzy neural network model, output the predicted result of hydrate decomposition rate, and establish a database corresponding to the input features and the predicted results. S6. Visualize and analyze the prediction results and their corresponding input features, and output the main prediction results and trends.

2. The hydrate decomposition rate prediction method based on fuzzy neural network according to claim 1, characterized in that, Step S1 includes: S11, The multi-source parameter data includes temperature (K) Pressure (MPa), porosity (Dimensionless), lithological parameter L, gas composition (mole fraction), environmental parameters Where L represents the classification code, Indicates the gas component number. Indicates the environmental parameter number; S12, Utilization Indicates the first The parameter in the first... The original observations under each sample, of which, Numbering of multi-source parameters, The sample number is... Indicates temperature In the The values ​​for each sample Indicates pressure in the first The values ​​for each sample are calculated, and so on. S13. Remove any missing data. The mean imputation method is used for completion, and the completion formula is as follows: in, Indicates the first Each parameter actually represents the total number of samples of observed values. Indicates the first Each parameter has a sample number with observed values; S14. Remove outliers from the original data using... In principle, if If the mean is an outlier, it is considered an outlier. and standard deviation They are respectively: S15. After removing outliers, the data is normalized using the min-max normalization method, with the following formula: in, For the first The minimum value of each parameter. For the first The maximum value of each parameter. These are the normalized, standardized input feature samples. S16. Perform a quality check on the above data processing results. If the quality check passes, a standardized input feature sample is obtained. If data anomalies are found or the data cannot meet the modeling requirements, the data is fed back to the data acquisition stage for re-acquisition and processing until the data quality meets the subsequent modeling requirements.

3. The hydrate decomposition rate prediction method based on fuzzy neural network according to claim 1, characterized in that, Step S2 includes: S21. Use Pearson correlation coefficient analysis to standardize input features. With hydrate decomposition rate The correlation between them, The correlation coefficient is calculated using the formula for the decomposition rate, which is obtained by measuring changes in gas production or mass at the experimental site. in, For the first The mean of each feature, This represents the average decomposition rate. S22. Based on the correlation analysis results, select characteristic parameters that are highly correlated with the decomposition rate. By removing parameters with low relevance or redundancy, an optimized feature subset is obtained. ; S23. Optimize the feature subset Principal component analysis is performed to reduce the dimensionality of the feature parameters to principal component features. The formula for calculating principal components is: in, This indicates the number of feature parameters after optimization. Represents the principal component coefficients. Indicates the principal component feature number and ; S24. Based on the extracted principal component features and their corresponding decomposition rates, select appropriate input parameters and target outputs to construct training and testing datasets.

4. The hydrate decomposition rate prediction method based on fuzzy neural network according to claim 1, characterized in that, Step S3 includes: S31. An adaptive network fuzzy inference system is used to construct a fuzzy neural network model, which includes an input layer, a fuzzification layer, a fuzzy rule layer, a defuzzification layer, and an output layer. S32, Principal component features for each input Configure Gaussian membership functions The expression is: in, Number the membership function and , For the first The number of membership functions of each principal component. For the first The first principal component Membership function center, For the corresponding width; S33. Construct a fuzzy rule layer based on the fuzzification results of the input principal components. The fuzzy rules are represented as follows: in, For the first Rule number 1 Fuzzy subsets of principal components For the rule consequent function, Output the rules; S34. In the defuzzification layer, a weighted average is performed on the outputs of all fuzzy rules to finally output the predicted value. The calculation formula is as follows: in, The total number of fuzzy rules, For the first The activation strength of the rule; S35. Configure all network parameters, including , The model structure is built by determining the fuzzy rule parameters and output layer weight parameters.

5. The hydrate decomposition rate prediction method based on fuzzy neural network according to claim 1, characterized in that, Step S4 includes: S41. Principal component features of the training dataset and corresponding decomposition rate The samples are input into the fuzzy neural network model. S42. Calculate the predicted output for each sample using the MSE loss function. With actual decomposition rate Mean squared error between, loss function The formula is: ， in, This represents the total number of samples. The training is completed when the loss function converges or the training termination condition is met. S43. Combine the backpropagation algorithm with gradient descent to calculate the gradient of the loss function with respect to the network weights and membership function parameters, and iteratively update the parameters until the loss function converges or the training termination condition is met. S44. Evaluate the model's performance through cross-validation and calculate the coefficient of determination for the prediction accuracy index. The formula is: ， in, This represents the average actual decomposition rate of the sample. S45. If the performance evaluation results meet the standards, the trained fuzzy neural network model is obtained; otherwise, the process is fed back to step S43 to continue model training and parameter optimization until the model achieves the required prediction accuracy and generalization ability.

6. The hydrate decomposition rate prediction method based on fuzzy neural network according to claim 1, characterized in that, Step S5 includes: S51, The new working condition data refers to multi-source parameter data that did not appear during the model training phase but was collected in actual application; S52. Principal component features of the test dataset or new working condition data. The input is fed into the trained fuzzy neural network model to obtain the decomposition rate prediction result. ,in, , To predict the total number of samples; S53. For each set of input features The predicted value is calculated using the weighted average formula of the deblurring layer. : S54. Output prediction results in batches for all input conditions. ; S55. Record the features of each set of inputs. Corresponding predicted decomposition rate To form a mapping relationship database .

7. The hydrate decomposition rate prediction method based on fuzzy neural network according to claim 1, characterized in that, Step S6 includes: S61. Organize the forecast results and their corresponding input features ,in, And generate the corresponding relationship table; S62. Use visualization tools to analyze input features. With predicted decomposition rate Multidimensional visualization analysis of relationships is performed, including two-dimensional scatter plots, three-dimensional surface plots, and heat maps; S63. Plot the response curve of the decomposition rate as a function of key input features, and analyze the functional relationship between the prediction results and the feature changes. S64. Regarding the prediction results Perform statistical analysis and calculate the mean. and standard deviation The calculation formulas are as follows: ， S65. Based on the visualized statistical results, output the main prediction conclusions and trends, and form a comprehensive report.

8. The method according to any one of claims 1 to 7, characterized in that, The membership function type of the fuzzy neural network model is Gaussian, trigonometric, or trapezoidal, and the specific type is selected according to the modeling accuracy requirements.

9. The method according to any one of claims 1 to 7, characterized in that, The decomposition rate prediction result database supports the retrieval, comparison, and batch export of historical prediction results.

Images