Natural gas hydrate exploitation productivity prediction method and system

Abstract

The invention discloses a natural gas hydrate exploitation productivity prediction method and system. The method comprises the following steps: obtaining a target influence variable associated with an environmental factor in a target area; on the basis of the target influence variable, the exploitation productivity of the natural gas hydrate is predicted according to an exploitation productivity prediction model, and a natural gas hydrate exploitation productivity prediction result of the target area is obtained; wherein the mining productivity prediction model is obtained based on a loss function training transition model; the transition model is obtained by optimizing the initial model. According to the natural gas hydrate exploitation productivity prediction method and system provided by the invention, the data diversity is enhanced through positive and negative sample pairs, the model distinguishing capability is improved by designing a fine loss function, and the model performance is improved by combining a swarm intelligence optimization means, so that the prediction accuracy of the natural gas hydrate exploitation productivity is ensured; therefore, the problem that the prediction result of the exploitation productivity of the natural gas hydrate is inaccurate due to insufficient data volume or low data quality can be solved.

Description

Technical Field

[0001] The present invention relates to the field of natural gas hydrate exploitation, and in particular to a natural gas hydrate exploitation capacity prediction method and system. Background Art

[0002] Natural Gas Hydrate (NGH), commonly known as "combustible ice", is a non-stoichiometric cage-like crystalline substance formed by water and methane gas under low temperature and high pressure conditions. It mainly exists below the permafrost layer and on the seafloor slope, and has the characteristics of wide distribution, large resources, and high energy density. Against the background of growing energy demand, natural gas hydrate, as a new potential energy source, has attracted much attention in drilling projects for its mining capacity. Various regions and research institutions have stepped up their research on natural gas hydrates in order to achieve commercial mining. The existing method mainly uses existing natural gas hydrate mining data to establish an empirical formula between the capacity and various influencing factors (such as depressurization rate, hydrate layer thickness, gas composition, etc.) through statistical analysis to predict the mining capacity of natural gas hydrates. This prediction method is simple and direct.

[0003] However, for the existing technology, if the amount of data used for statistical analysis is insufficient, the established empirical formula may not be able to fully reflect the true relationship between various influencing factors and production capacity, which will lead to inaccurate prediction results; and if there are errors, omissions or inconsistencies in the data, the established empirical formula will also be affected, resulting in inaccurate prediction results, causing difficulties in the natural gas hydrate extraction process. Summary of the invention

[0004] The present invention provides a natural gas hydrate production capacity prediction method and system to solve the problem of inaccurate prediction results of natural gas hydrate production capacity due to insufficient data or low data quality.

[0005] Obtain target influencing variables associated with environmental factors in the target area; Based on the target influencing variables, the mining capacity of natural gas hydrates is predicted according to the mining capacity prediction model to obtain the prediction result of the mining capacity of natural gas hydrates in the target area; wherein, the mining capacity prediction model is obtained by training a transition model based on a loss function; the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization; the initial model is obtained by establishing a natural gas hydrate data sample set based on the influencing variables associated with environmental factors in the target area, and training the learning model according to the classification results of the natural gas hydrate data sample set; the loss function is established by increasing the distance between the positive sample pairs and the negative sample pairs in the influencing variables, and reducing the distance between the data in the positive sample pairs; the positive sample pairs and the negative samples are determined according to the similarity values ​​of environmental factors and mining capacity.

[0006] The present invention is based on the target influencing variable, directly predicts the mining capacity of natural gas hydrate according to the mining capacity prediction model, and can quickly and effectively obtain the prediction result of natural gas hydrate mining capacity. For the mining capacity prediction model, the learning model is trained based on the classification results of the natural gas hydrate data sample set, so that the model can initially learn the characteristics and laws in the data. The initial model is optimized by swarm intelligence optimization, which can make up for the problem of low data quality to a certain extent. Finally, the transition model is trained based on the loss function to obtain the final mining capacity prediction model. This method ensures that the model can fully learn the key information in the data during the training process, thereby improving the accuracy of the prediction; wherein, the positive sample pairs and negative sample pairs are determined according to the similarity values ​​of environmental factors and mining capacity, which helps the model learn the law of capacity change under different mining conditions, thereby making up for the problem of insufficient data volume; at the same time, the loss function also enhances the model's recognition ability of similar data by reducing the distance between the data in the positive sample pairs, which helps the model to make more consistent predictions when facing similar mining conditions, thereby improving the stability of the prediction and providing strong data support for drilling engineering.

[0007] Compared with the prior art, the present invention enhances data diversity through positive and negative sample pairs, designs a sophisticated loss function to improve the model's discrimination ability, and combines swarm intelligence optimization to improve model performance, thereby ensuring the accuracy of the prediction of natural gas hydrate production capacity. It can therefore solve the problem of inaccurate prediction results of natural gas hydrate production capacity due to insufficient data or low data quality.

[0008] As a preferred solution, the initial model is obtained by establishing a natural gas hydrate data sample set based on the influencing variables associated with environmental factors in the target area, and training the learning model according to the classification results of the natural gas hydrate data sample set, specifically: Converting the influencing variables associated with the environmental factors in the target area into a text sequence, and obtaining a global feature representing the entire text or sentence pair in the text sequence by adding a preset marker to the first position of a sentence in the text sequence; Generate an embedding vector according to the global feature, and form a processed text by the global feature and the embedding vector; Combining the processed text with the actual classification of the influencing variables to construct the natural gas hydrate data sample set; wherein the natural gas hydrate data sample set includes global features representing sentence pairs and separators for distinguishing different sentence pairs; Mapping the text features of the natural gas hydrate data sample set to the target classification of the influencing variable through a natural language processing model to obtain a classification result of the natural gas hydrate data sample set; The learning model is trained according to the classification results of the natural gas hydrate data sample set to obtain the initial model.

[0009] This preferred solution converts the influencing variables into text sequences, which can capture the semantic relationship and contextual information between variables. This representation is more flexible and richer than the traditional numerical representation, which helps the model understand how complex environmental factors affect the formation and distribution of natural gas hydrates. Adding a preset marker to the beginning of the sentence in the text sequence to generate a global feature representing the entire text or sentence pair helps the model maintain awareness of the overall context when processing the text. Constructing a natural gas hydrate data sample set containing global features and separators so that the data has a clear structure and hierarchy. This structured data representation method is conducive to the model learning effective feature combinations and classification rules.

[0010] As a preferred solution, the text features of the natural gas hydrate data sample set are mapped to the target classification of the influencing variables through a natural language processing model to obtain the classification result of the natural gas hydrate data sample set, specifically: Based on the natural language processing model, the natural gas hydrate data sample set is used as an input of an embedding layer for semantic recognition to obtain a target embedding vector set of several labels in the natural gas hydrate data sample set; By encoding the target embedding vector set from both positive and negative directions, long-distance dependencies and contextual semantic knowledge in the text are captured to obtain a feature representation of the natural gas hydrate data sample set; Based on the classification layer of the natural language processing model, the feature representation is mapped to the classification space of the influencing variable, and the prediction probability of the natural gas hydrate data sample set on each influencing variable category is obtained by calculating the similarity or distance between the feature representation and a plurality of classification labels; According to the magnitude of the predicted probability, a classification label is assigned to each natural gas hydrate data sample in the natural gas hydrate data sample set to obtain a classification result of the natural gas hydrate data sample set.

[0011] This preferred solution captures long-distance dependencies and contextual semantic knowledge in the text by encoding, which helps to understand the implicit information and logical structure in complex texts; and the obtained feature representation helps the model learn more effective classification rules and improve classification performance. Based on the classification layer of the natural language processing model, the feature representation is mapped to the classification space of the influencing variables, and the prediction probability is obtained by calculating the similarity or distance. This classification method is not only flexible, but also can accurately reflect the relationship between text features and classification labels, thereby improving the accuracy of classification.

[0012] As a preferred solution, the learning model is trained according to the classification results of the natural gas hydrate data sample set to obtain the initial model, specifically: Balancing the category distribution of the classification results by resampling to obtain a first classification result; Training a plurality of learning models according to the first classification result to obtain a result set, and fusing the result set in a weighted manner to obtain a prediction result; The prediction results are used as new features to train the meta-learner to obtain the initial model.

[0013] This preferred solution balances the category distribution of classification results through resampling technology, which can effectively solve the problem of unbalanced data sets and help improve the model's ability to recognize minority class samples, thereby improving the overall classification performance. Using multiple learning models for training and fusing their results in a weighted manner can fully utilize the advantages of different models and improve the accuracy and robustness of predictions. Using the fused prediction results as new features to train the meta-learner allows the meta-learner to learn how to best combine the prediction results of several learning models to make a more accurate final decision.

[0014] As a preferred solution, the mining capacity prediction model is obtained by training the transition model based on the loss function, specifically: Introducing a minimum margin, and establishing a first loss function according to the distance loss between the negative sample pair and the positive sample pair; wherein the minimum margin is used to control the relative dispersion between samples within the positive sample pair; Introducing a maximum margin, and establishing a second loss function according to the distance loss between the positive sample pair and a preset representative point; wherein the maximum margin is used to control the relative compactness between the internal samples of the positive sample pair, and the preset representative point represents the cluster center of the positive sample pair and the negative sample pair; Establishing the loss function based on the first loss function and the second loss function; The transition model is trained according to the loss function to obtain the mining capacity prediction model.

[0015] The introduction of the minimum margin in this preferred solution can control the relative dispersion between the positive samples and the internal samples, which helps the model learn a more stable feature representation during the training process, making the model insensitive to small changes in the input data, thereby improving the robustness of the model. By controlling the relative compactness between the internal samples of the positive samples by the maximum margin, and introducing preset representative points as cluster centers, the model can pay more attention to the overall structure and distribution between samples during the training process, which helps the model maintain good performance on unseen data and improves the generalization ability of the model. In addition, the two loss functions established focus on the internal dispersion and compactness of the positive samples respectively. The design of this dual loss function can more comprehensively reflect the performance of the model during the training process and help optimize the model in multiple aspects at the same time.

[0016] As a preferred solution, the first loss function and the second loss function are specifically: ;

[0017] in, is the first loss function, is the second loss function, is the feature vector associated with the positive sample pair, is the feature vector associated with the negative sample pair, is the cluster center, and are the minimum margin and the maximum margin respectively.

[0018] As a preferred solution, the positive sample pair and the negative sample are determined based on similar values ​​of environmental factors and mining capacity, specifically: Based on the influencing variables associated with environmental factors in the target area, a sample pair whose similarity value of environmental factors is greater than a first standard and whose similarity value of mining capacity is greater than a second standard is defined as the positive sample pair; A sample pair whose similarity value of environmental factors is greater than the first standard but whose similarity value of mining capacity is less than or equal to the second standard is defined as the negative sample pair.

[0019] This preferred solution can clearly divide sample pairs into positive sample pairs and negative sample pairs by setting clear standards. This clear distinction helps in sample selection and weight allocation in the subsequent model training process, so as to improve the training efficiency and performance of the model. Since both positive sample pairs and negative sample pairs are defined based on environmental factors and mining capacity in actual data, the model can learn more realistic and complex feature representations during training, which helps the model maintain good performance on unseen data and improves the generalization ability of the model.

[0020] As a preferred solution, the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization, specifically: randomly generating a group of particles, and initializing a position vector of each particle in the group of particles to a random value of a model parameter in the initial model; According to the data iteration method, the personal best position of each particle is updated, and the global best position of the particle group to which each particle belongs is updated; If the evaluation result of the fitness function converges to a preset range, the data iteration is stopped and the current global optimal position of the particle is output; wherein the fitness function is used to evaluate the quality of each particle; The current global optimal position is used as a parameter configuration to establish the transition model.

[0021] Each particle in the preferred solution is an independent search entity, and the update process between them can be carried out in parallel, which enables the algorithm to fully utilize the multi-core processing power of modern computers and further improve the search efficiency and performance.

[0022] As a preferred solution, the fitness function is specifically:

[0023] in, is the objective function value, is the coefficient of the L2 regularization term, is the L2 norm of the parameter vector x, is the coefficient of the diversity promotion term, Represents particles The measurement result of the average distance to other particles in the particle group; is related to the number of iterations The coefficient of correlation, which changes as the iteration proceeds; is a preset function that varies with the number of iterations.

[0024] The present application also provides a natural gas hydrate production capacity prediction system, including a data module and a prediction module; Wherein, the data module is used to obtain target influencing variables associated with environmental factors in the target area; The prediction module is used to predict the production capacity of natural gas hydrates based on the target influencing variables and the production capacity prediction model, and obtain the prediction result of the production capacity of natural gas hydrates in the target area; wherein the production capacity prediction model is obtained by training a transition model based on a loss function; the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization; the initial model is obtained by establishing a natural gas hydrate data sample set based on the influencing variables associated with environmental factors in the target area, and training the learning model according to the classification results of the natural gas hydrate data sample set; the loss function is established by increasing the distance between the positive sample pairs and the negative sample pairs in the influencing variables, and reducing the distance between the data in the positive sample pairs; the positive sample pairs and the negative samples are determined according to the similarity values ​​of environmental factors and production capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a method for predicting the production capacity of natural gas hydrate mining provided in an embodiment of the present application; Figure 2 It is a structural schematic diagram of a natural gas hydrate mining capacity prediction system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "several" means two or more.

[0028] Embodiment 1: See also Figure 1 The embodiment of the present application provides a method for predicting the production capacity of natural gas hydrate production, including S1-S2, and the specific implementation steps are as follows: S1. Obtain target influencing variables associated with environmental factors in the target area.

[0029] Step S1 of the embodiment of the present application is specifically as follows: Using sensor equipment installed in the target area or using satellite remote sensing technology to obtain target data associated with environmental factors in the target area in real time; the target data includes environmental data such as air quality, water quality and noise; The collected target data are cleaned, sorted and analyzed to remove outliers and missing values, and the target influencing variables associated with environmental factors in the target area are obtained.

[0030] S2. Based on the target influencing variables, the mining capacity of natural gas hydrates is predicted according to the mining capacity prediction model to obtain the prediction result of the mining capacity of natural gas hydrates in the target area; wherein, the mining capacity prediction model is obtained by training the transition model based on the loss function; the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization; the initial model is based on the influencing variables associated with environmental factors in the target area to establish a natural gas hydrate data sample set, and the learning model is trained according to the classification results of the natural gas hydrate data sample set; the loss function is established by increasing the distance between the positive sample pairs and the negative sample pairs in the influencing variables, and reducing the distance between the data in the positive sample pairs; the positive sample pairs and the negative samples are determined according to the similarity values ​​of the environmental factors and the mining capacity.

[0031] Step S2 of the embodiment of the present application includes S2.1 to S2.6, specifically: S2.1. In the target area, various data related to environmental factors are collected through geological exploration, geophysical measurement, chemical analysis and other means to obtain raw data. When collecting data, multiple different exploration points or sampling points should be selected as much as possible, and accurate and reliable data should be collected; The original data is preprocessed to obtain the influencing variables associated with environmental factors in the target area; wherein the preprocessing includes steps such as data cleaning, data conversion, and data normalization.

[0032] The influencing variables associated with environmental factors in the target area are converted into text sequences. By adding a preset mark (such as #, @, $, % etc.) to the first position of the sentence in the text sequence, a global feature representing the entire text or sentence pair in the text sequence is obtained; for example, using the special character $ at the beginning, the corresponding text can be expressed as "$environmental factor_text content"; Based on the global features, a pre-trained language model (such as BERT, GPT, etc.) is used to generate an embedding vector, and the processed text is composed of the global features and the embedding vector; The processed text is combined with the actual classification of the influencing variables to construct a natural gas hydrate data sample set; wherein the natural gas hydrate data sample set contains global features representing sentence pairs and separators for distinguishing different sentence pairs; and the actual classification of the influencing variables is obtained by dividing the influencing variables according to the defined classification criteria, and optimizing the division results of the influencing variables through methods such as manual judgment, machine learning algorithms or statistical analysis.

[0033] This embodiment S2.1 converts the influencing variables into text sequences, which can capture the semantic relationship and contextual information between the variables. This representation is more flexible and richer than the traditional numerical representation, which helps the model understand how complex environmental factors affect the formation and distribution of natural gas hydrates. Adding a preset tag to the first sentence of the text sequence to generate a global feature representing the entire text or sentence pair helps the model maintain awareness of the overall context when processing the text. Constructing a natural gas hydrate data sample set containing global features and separators so that the data has a clear structure and hierarchy, this structured data representation method is conducive to the model learning effective feature combinations and classification rules.

[0034] S2.2. Based on the natural language processing model, the natural gas hydrate data sample set is used as the input of the embedding layer for semantic recognition to obtain a target embedding vector set of several labels in the natural gas hydrate data sample set; Use the forward LSTM in the bidirectional LSTM (Long Short-Term Memory) layer to encode the target embedding vector set from left to right, thereby capturing the long-distance dependencies and contextual semantic knowledge in the text from left to right; The reverse LSTM in the bidirectional LSTM layer is used to encode the target embedding vector set from right to left, thereby capturing the long-range dependencies and contextual semantic knowledge in the text from right to left. The outputs of the forward LSTM and the backward LSTM are merged to obtain the complete context representation of each time step, that is, the feature representation of the natural gas hydrate data sample set. When merging the outputs of the forward and backward LSTMs, the attention mechanism is introduced to allow the model to dynamically adjust the degree of attention to information at different time steps, so that the model can focus more on key information when processing complex information.

[0035] Based on the classification layer of the natural language processing model, the feature representation is mapped to the classification space of the influencing variables, and the prediction probability of the natural gas hydrate data sample set on each influencing variable category is obtained by calculating the similarity or distance between the feature representation and several classification labels; According to the size of the predicted probability, a classification label is assigned to each gas hydrate data sample in the gas hydrate data sample set to obtain the classification result of the gas hydrate data sample set; when processing gas hydrate data samples, a clear probability threshold needs to be set as the basis for judging whether the sample belongs to a certain category. When the predicted probability of a sample for a certain category reaches or exceeds this threshold, the sample is considered to have the characteristics of the category and is classified into the corresponding category. In addition, it should be noted that under this classification strategy, a sample is no longer simply divided into a single category, but may be assigned multiple classification labels at the same time. This means that according to the characteristics and predicted probability of the sample, it may belong to multiple different categories at the same time. For example, for the classification of sample A, the predicted probability is 0.3 for category 1, which is lower than the set threshold of 0.5, so it is determined that sample A does not belong to category 1; for category 2, the predicted probability is 0.7, which is higher than the threshold of 0.5, so it is determined that sample A belongs to category 2; for category 3, the predicted probability is 0.8, which is higher than the threshold of 0.5, so it is determined that sample A belongs to category 3; in summary, sample A is assigned classification labels of category 2 and category 3.

[0036] The formula for similarity is:

[0037] The formula for distance is:

[0038] Among them, in the formula of similarity, and are feature vectors and classification labels respectively, · represents the dot product, and Respectively The modulus and The module length, Represents the dimension of the vector; In the distance formula, and Respectively represent the coordinate values ​​of vector A and vector B in the i-th dimension.

[0039] This embodiment S2.2 captures the long-distance dependencies and contextual semantic knowledge in the text by encoding, which helps to understand the implicit information and logical structure in complex texts; and the obtained feature representation helps the model learn more effective classification rules and improve classification performance. Based on the classification layer of the natural language processing model, the feature representation is mapped to the classification space of the influencing variables, and the prediction probability is obtained by calculating the similarity or distance. This classification method is not only flexible, but also can accurately reflect the relationship between text features and classification labels, thereby improving the accuracy of classification.

[0040] S2.3. Balance the category distribution of the classification results of the natural gas hydrate data sample set by resampling to obtain a first classification result; specifically: if the number of the first category in the category distribution of the classification result is less than a preset value, copy the first category or generate synthetic samples to increase its number; if the number of the second category in the category distribution of the classification result is greater than or equal to a preset value, randomly delete some samples from the second category to reduce its number to make it close to the number of the first category, thereby obtaining the first classification result.

[0041] According to the first classification result, several learning models are trained to obtain a result set, and the result set is fused in a weighted manner according to the weight of each model to obtain a prediction result; wherein the weight of each model is determined by cross-validation, grid search or other methods, and the weight is used to reflect the importance of each model in the fusion result; With the goal of minimizing the preset first function, the prediction results are used as new features to perform final training on the meta-learner to obtain the initial model.

[0042] Among them, the first function is:

[0043] in, are the parameters of the meta-learner, is the number of training samples; is the loss function, which is used to calculate the prediction error of a single sample; and They are A new feature set and target variable for each training sample.

[0044] This embodiment S2.3 balances the category distribution of the classification results through resampling technology, which can effectively solve the problem of unbalanced data sets and help improve the model's recognition ability for minority class samples, thereby improving the overall classification performance. Using multiple learning models for training and fusing their results in a weighted manner can fully utilize the advantages of different models and improve the accuracy and robustness of predictions. Using the fused prediction results as new features to train the meta-learner can allow the meta-learner to learn how to best combine the prediction results of several learning models to make a more accurate final decision.

[0045] S2.4, randomly generating a group of particles, and initializing the position vector of each particle in the group of particles to a random value of a model parameter in the initial model; According to the data iteration method, the personal best position of each particle is updated, and the global best position of the particle group to which each particle belongs is updated; wherein, in each iteration, the first particle set with the best performance is retained, and in the next iteration, the position or position information of the first particle set is directly used to guide other particles to search, so as to accelerate the convergence process; If the fitness function When the evaluation result converges to the preset range, the data iteration is stopped and the current global optimal position of the particle is output; among them, the fitness function It is used to evaluate the quality of each particle, that is, the function represents the performance of the model parameter configuration; The current global optimal position is used as parameter configuration to establish the transition model.

[0046] Among them, the fitness function is:

[0047] in, is the objective function value, which is used to measure the particle The quality of the location; is the coefficient of the L2 regularization term, which is used to penalize the model parameters to avoid overfitting; is the L2 norm of the parameter vector x, is the coefficient of the diversity promotion term, Represents particles The measurement result of the average distance to other particles in the particle group; is related to the number of iterations The coefficient of correlation, which changes as the iteration proceeds; is a preset function that varies with the number of iterations.

[0048] Each particle in S2.4 of this embodiment is an independent search entity, and the update process between them can be carried out in parallel, which enables the algorithm to fully utilize the multi-core processing power of modern computers and further improve the search efficiency and performance.

[0049] S2.5. Based on the influencing variables associated with environmental factors in the target area, sample pairs whose similarity values ​​of environmental factors are greater than the first standard and whose similarity values ​​of mining capacity are greater than the second standard are defined as positive sample pairs; The sample pairs whose similarity values ​​of environmental factors are greater than the first standard but whose similarity values ​​of mining capacity are less than or equal to the second standard are defined as negative sample pairs.

[0050] Introducing minimum margins , and establish the first loss function based on the distance loss between the negative sample pair and the positive sample pair ; Among them, the minimum margin It is the lower limit of the distance between the positive sample pair and the cluster center (or a representative point), which is used to control the relative dispersion between the internal samples of the positive sample pair; Introducing maximum margins , and establish the second loss function based on the distance loss between the positive sample pair and the preset representative point ; Among them, the maximum margin is the upper limit of the distance between the negative sample pair and the positive sample pair, which is used to control the relative compactness between the internal samples of the positive sample pair; the preset representative points represent the cluster centers of the positive sample pair and the negative sample pair; Based on the first loss function and the second loss function Building a loss function ; According to the loss function The transition model is trained to obtain the mining capacity prediction model.

[0051] Among them, the first loss function and the second loss function are: ;

[0052] The loss function is:

[0053] in, is the first loss function, is the second loss function, is the loss function, is the feature vector associated with the positive sample pair, is the feature vector associated with the negative sample pair, is the cluster center, and are the minimum margin and maximum margin respectively; is a preset adjustment parameter used to control and The difference between sensitivity to impact; is the preset weight parameter used to adjust the > When, yes the degree of additional punishment; is a preset exponential parameter that controls the degree of nonlinearity of the additional penalty.

[0054] The minimum margin introduced in S2.5 of this embodiment can control the relative dispersion between internal samples of the positive sample pair, which helps the model learn a more stable feature representation during training, making the model insensitive to small changes in the input data, thereby improving the robustness of the model. By controlling the relative compactness between internal samples of the positive sample pair by the maximum margin, and introducing preset representative points as cluster centers, the model can pay more attention to the overall structure and distribution between samples during training, which helps the model maintain good performance on unseen data and improves the generalization ability of the model. In addition, the two loss functions established focus on the dispersion and compactness of the internal positive sample pair respectively. The design of this dual loss function can more comprehensively reflect the performance of the model during training and help optimize the model in multiple aspects at the same time; Moreover, by setting clear standards, sample pairs can be clearly divided into positive sample pairs and negative sample pairs. This clear distinction helps in sample selection and weight allocation in the subsequent model training process, so as to improve the training efficiency and performance of the model. Since both positive and negative sample pairs are defined based on environmental factors and mining capacity in actual data, the model can learn more realistic and complex feature representations during training, which helps the model maintain good performance on unseen data and improves the generalization ability of the model.

[0055] S2.6. Based on the target influencing variables, the production capacity of natural gas hydrate is predicted according to the production capacity prediction model to obtain the prediction result of the production capacity of natural gas hydrate in the target area; The prediction results of natural gas hydrate production capacity are presented in various forms, such as specific production capacity values, production capacity distribution diagrams, and production capacity trend diagrams over time. Specific examples are as follows: directly provide the estimated production volume of natural gas hydrate under different conditions, and show the production potential in precise numerical form; use maps or charts to intuitively depict the spatial distribution pattern of production capacity in the target area, help identify high-yield and low-yield areas, and provide a strong basis for resource planning; by drawing time series charts, deeply analyze the dynamic change trend of natural gas hydrate production capacity in the future, predict the increase or decrease in production capacity, and provide a reference for the formulation of long-term production strategies.

[0056] Overall, this embodiment has the following beneficial effects: This application predicts the mining capacity of natural gas hydrates directly based on the mining capacity prediction model based on the target influencing variables, and can quickly and effectively obtain the prediction results of natural gas hydrate mining capacity. For the mining capacity prediction model, training the learning model based on the classification results of the natural gas hydrate data sample set can enable the model to initially learn the characteristics and laws in the data. Optimizing the initial model by means of group intelligence optimization can make up for the problem of low data quality to a certain extent. Finally, the transition model is trained based on the loss function to obtain the final mining capacity prediction model. This method ensures that the model can fully learn the key information in the data during the training process, thereby improving the accuracy of the prediction; among them, determining the positive sample pairs and negative sample pairs according to the similarity values ​​of environmental factors and mining capacity helps the model learn the law of capacity changes under different mining conditions, thereby making up for the problem of insufficient data volume; at the same time, the loss function also enhances the model's ability to recognize similar data by reducing the distance between the data in the positive sample pairs, which helps the model to make more consistent predictions when facing similar mining conditions, thereby improving the stability of the prediction and providing strong data support for drilling engineering.

[0057] Embodiment 2: See also Figure 2 , the embodiment of the present application provides a natural gas hydrate production capacity prediction system, including a data module 10 and a prediction module 20; Wherein, the data module 10 is used to obtain the target influencing variables associated with the environmental factors in the target area; The prediction module 20 is used to predict the mining capacity of natural gas hydrates based on the target influencing variables and the mining capacity prediction model, and obtain the prediction result of the mining capacity of natural gas hydrates in the target area; wherein the mining capacity prediction model is obtained by training the transition model based on the loss function; the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization; the initial model is based on the influencing variables associated with environmental factors in the target area to establish a natural gas hydrate data sample set, and the learning model is trained according to the classification results of the natural gas hydrate data sample set; the loss function is established by increasing the distance between the positive sample pairs and the negative sample pairs in the influencing variables, and reducing the distance between the data in the positive sample pairs; the positive sample pairs and the negative samples are determined according to the similarity values ​​of the environmental factors and the mining capacity.

[0058] In one embodiment, the data module 10 is specifically: Using sensor equipment installed in the target area or using satellite remote sensing technology to obtain target data associated with environmental factors in the target area in real time; the target data includes environmental data such as air quality, water quality and noise; The collected target data are cleaned, sorted and analyzed to remove outliers and missing values, and the target influencing variables associated with environmental factors in the target area are obtained.

[0059] In one embodiment, the prediction module 20 includes a sample unit, a classification unit, an initial unit, a transition unit, a model unit, and a prediction unit, specifically: The sample unit is used to collect various data related to environmental factors in the target area through geological exploration, geophysical measurement, chemical analysis and other means to obtain raw data; when collecting data, multiple different exploration points or sampling points should be selected as much as possible, and accurate and reliable data should be collected; The sample unit is also used to preprocess the original data to obtain the influencing variables associated with environmental factors in the target area; wherein the preprocessing includes steps such as data cleaning, data conversion, and data normalization.

[0060] The sample unit is also used to convert the influencing variables associated with environmental factors in the target area into a text sequence. By adding a preset mark (such as #, @, $, % etc.) at the beginning of the sentence in the text sequence, the global feature representing the entire text or sentence pair in the text sequence is obtained; for example, using the special character $ at the beginning, the corresponding text can be expressed as "$environmental factor_text content"; The sample unit is also used to generate an embedding vector based on the global features using a pre-trained language model (such as BERT, GPT, etc.), and the processed text is composed of the global features and the embedding vector; The sample unit is also used to combine the processed text with the actual classification of the influencing variables to construct a natural gas hydrate data sample set; wherein the natural gas hydrate data sample set contains global features representing sentence pairs and separators for distinguishing different sentence pairs; and the actual classification of the influencing variables is obtained by dividing the influencing variables according to the defined classification criteria, and optimizing the division results of the influencing variables through methods such as manual judgment, machine learning algorithms or statistical analysis.

[0061] The sample unit of this embodiment converts the influencing variables into text sequences, which can capture the semantic relationship and contextual information between variables. This representation is more flexible and richer than the traditional numerical representation, which helps the model understand how complex environmental factors affect the formation and distribution of natural gas hydrates. Adding a preset marker to the beginning of the sentence in the text sequence to generate a global feature representing the entire text or sentence pair helps the model maintain awareness of the overall context when processing the text. Constructing a natural gas hydrate data sample set containing global features and separators gives the data a clear structure and hierarchy. This structured data representation method is conducive to the model learning effective feature combinations and classification rules.

[0062] A classification unit is used to perform semantic recognition based on a natural language processing model by taking the natural gas hydrate data sample set as an input of the embedding layer to obtain a target embedding vector set of several labels in the natural gas hydrate data sample set; The classification unit is also used to use the forward LSTM in the bidirectional LSTM (Long Short-Term Memory) layer to encode the target embedding vector set from left to right, thereby capturing the long-distance dependencies and contextual semantic knowledge in the text from left to right; The classification unit is also used to encode the target embedding vector set from right to left using the reverse LSTM in the bidirectional LSTM layer, thereby capturing the long-range dependencies and contextual semantic knowledge in the text from right to left; The classification unit is also used to merge the outputs of the forward LSTM and the backward LSTM to obtain a complete context representation of each time step, that is, to obtain a feature representation of the natural gas hydrate data sample set; when merging the outputs of the forward and backward LSTMs, the attention mechanism is introduced to allow the model to dynamically adjust the degree of attention to information at different time steps, so that the model can focus more on key information when processing complex information.

[0063] The classification unit is also used for mapping the feature representation to the classification space of the influencing variables based on the classification layer of the natural language processing model, and obtaining the predicted probability of the natural gas hydrate data sample set on each influencing variable category by calculating the similarity or distance between the feature representation and several classification labels; The classification unit is also used to assign a classification label to each gas hydrate data sample in the gas hydrate data sample set according to the size of the predicted probability, and obtain the classification result of the gas hydrate data sample set; wherein, when processing the gas hydrate data sample, it is necessary to set a clear probability threshold as the basis for judging whether the sample belongs to a certain category. When the predicted probability of a sample for a certain category reaches or exceeds this threshold, the sample is considered to have the characteristics of the category and is classified into the corresponding category. In addition, it should be noted that under this classification strategy, a sample is no longer simply divided into a single category, but may be assigned multiple classification labels at the same time. This means that, based on the characteristics and predicted probability of the sample, it may belong to multiple different categories at the same time. For example, for the classification of sample A, the predicted probability is 0.3 for category 1, which is lower than the set threshold of 0.5, so it is determined that sample A does not belong to category 1; for category 2, the predicted probability is 0.7, which is higher than the threshold of 0.5, so it is determined that sample A belongs to category 2; for category 3, the predicted probability is 0.8, which is higher than the threshold of 0.5, so it is determined that sample A belongs to category 3; in summary, sample A is assigned classification labels of category 2 and category 3.

[0064] The formula for similarity is:

[0065] The formula for distance is:

[0066] Among them, in the formula of similarity, and are feature vectors and classification labels respectively, · represents the dot product, and Respectively The modulus and The module length, Represents the dimension of the vector; In the distance formula, and Respectively represent the coordinate values ​​of vector A and vector B in the i-th dimension.

[0067] The classification unit of this embodiment captures the long-distance dependencies and contextual semantic knowledge in the text by encoding, which helps to understand the implicit information and logical structure in complex texts; and the obtained feature representation helps the model learn more effective classification rules and improve classification performance. Based on the classification layer of the natural language processing model, the feature representation is mapped to the classification space of the influencing variables, and the prediction probability is obtained by calculating the similarity or distance. This classification method is not only flexible, but also can accurately reflect the relationship between text features and classification labels, thereby improving the accuracy of classification.

[0068] The initialization unit is used to balance the category distribution of the classification results of the natural gas hydrate data sample set by resampling to obtain a first classification result; specifically: if the number of the first category in the category distribution of the classification result is less than a preset value, the first category is copied or synthetic samples are generated to increase its number; if the number of the second category in the category distribution of the classification result is greater than or equal to a preset value, some samples are randomly deleted from the second category to reduce its number to make it close to the number of the first category, thereby obtaining the first classification result.

[0069] The initialization unit is further used to train several learning models according to the first classification result to obtain a result set, and fuse the result set in a weighted manner according to the weight of each model to obtain a prediction result; wherein the weight of each model is determined by cross-validation, grid search or other methods, and the weight is used to reflect the importance of each model in the fusion result; The initial unit is also used to minimize the preset first function and use the prediction results as new features to perform final training on the meta-learner to obtain the initial model.

[0070] Among them, the first function is:

[0071] in, are the parameters of the meta-learner, is the number of training samples; is the loss function, which is used to calculate the prediction error of a single sample; and They are A new feature set and target variable for each training sample.

[0072] The initial unit of this embodiment balances the category distribution of the classification results through resampling technology, which can effectively solve the problem of unbalanced data sets and help improve the model's recognition ability for minority class samples, thereby improving the overall classification performance. Using multiple learning models for training and fusing their results in a weighted manner can fully utilize the advantages of different models and improve the accuracy and robustness of predictions. Using the fused prediction results as new features to train the meta-learner allows the meta-learner to learn how to best combine the prediction results of several learning models to make a more accurate final decision.

[0073] A transition unit, used for randomly generating a group of particles, and initializing the position vector of each particle in the group of particles to a random value of a model parameter in the initial model; The transition unit is also used to update the personal best position of each particle and the global best position of the particle group to which each particle belongs in accordance with the data iteration method; wherein, in each iteration, the first particle set with the best performance is retained, and in the next iteration, the position or position information of the first particle set is directly used to guide other particles to search, so as to accelerate the convergence process; The transition unit is also used if the fitness function When the evaluation result converges to the preset range, the data iteration is stopped and the current global optimal position of the particle is output; among them, the fitness function It is used to evaluate the quality of each particle, that is, the function represents the performance of the model parameter configuration; The transition unit is also used to configure the current global optimal position as a parameter to establish a transition model.

[0074] Among them, the fitness function is:

[0075] in, is the objective function value, which is used to measure the particle The quality of the location; is the coefficient of the L2 regularization term, which is used to penalize the model parameters to avoid overfitting; is the L2 norm of the parameter vector x, is the coefficient of the diversity promotion term, Represents particles The measurement result of the average distance to other particles in the particle group; is related to the number of iterations The coefficient of correlation, which changes as the iteration proceeds; is a preset function that varies with the number of iterations.

[0076] In this embodiment, each particle in the transition unit is an independent search entity, and the update process between them can be performed in parallel, which enables the algorithm to fully utilize the multi-core processing capabilities of modern computers and further improve the search efficiency and performance.

[0077] A model unit is used for defining a sample pair whose similarity value of the environmental factor is greater than a first standard and whose similarity value of the mining capacity is greater than a second standard as a positive sample pair based on the influencing variables associated with the environmental factor in the target area; The model unit is also used to define a sample pair whose similarity value of environmental factors is greater than a first standard but whose similarity value of mining capacity is less than or equal to a second standard as a negative sample pair.

[0078] Model unit, also used to introduce minimum margins , and establish the first loss function based on the distance loss between the negative sample pair and the positive sample pair ; Among them, the minimum margin It is the lower limit of the distance between the positive sample pair and the cluster center (or a representative point), which is used to control the relative dispersion between the internal samples of the positive sample pair; Model unit, also used to introduce maximum margins , and establish the second loss function based on the distance loss between the positive sample pair and the preset representative point ; Among them, the maximum margin is the upper limit of the distance between the negative sample pair and the positive sample pair, which is used to control the relative compactness between the internal samples of the positive sample pair; the preset representative points represent the cluster centers of the positive sample pair and the negative sample pair; Model unit, also used based on the first loss function and the second loss function Building a loss function ; Model unit, also used to calculate the loss function The transition model is trained to obtain the mining capacity prediction model.

[0079] Among them, the first loss function and the second loss function are: ;

[0080] The loss function is:

[0081] in, is the first loss function, is the second loss function, is the loss function, is the feature vector associated with the positive sample pair, is the feature vector associated with the negative sample pair, is the cluster center, and are the minimum margin and maximum margin respectively; is a preset adjustment parameter used to control and The difference between sensitivity to impact; is the preset weight parameter used to adjust the > When, yes the degree of additional punishment; is a preset exponential parameter that controls the degree of nonlinearity of the additional penalty.

[0082] The model unit of this embodiment introduces a minimum margin to control the relative dispersion between internal samples of the positive sample pair, which helps the model learn a more stable feature representation during the training process, making the model insensitive to small changes in the input data, thereby improving the robustness of the model. By controlling the relative compactness between internal samples of the positive sample pair by the maximum margin, and introducing preset representative points as cluster centers, the model can pay more attention to the overall structure and distribution between samples during the training process, which helps the model maintain good performance on unseen data and improves the generalization ability of the model. In addition, the two loss functions established focus on the internal dispersion and compactness of the positive sample pair respectively. The design of this dual loss function can more comprehensively reflect the performance of the model during the training process and help the model to optimize multiple aspects simultaneously; Moreover, by setting clear standards, sample pairs can be clearly divided into positive sample pairs and negative sample pairs. This clear distinction helps in sample selection and weight allocation in the subsequent model training process, so as to improve the training efficiency and performance of the model. Since both positive and negative sample pairs are defined based on environmental factors and mining capacity in actual data, the model can learn more realistic and complex feature representations during training, which helps the model maintain good performance on unseen data and improves the generalization ability of the model.

[0083] A prediction unit, used to predict the production capacity of natural gas hydrate based on the target influencing variables and the production capacity prediction model, and obtain the prediction result of the production capacity of natural gas hydrate in the target area; The prediction unit is also used to present the prediction results of natural gas hydrate production capacity in various forms, such as specific production capacity values, production capacity distribution diagrams, and production capacity trend diagrams over time. Specific examples are as follows: directly provide the estimated production volume of natural gas hydrate under different conditions, and show the production potential in precise digital form; use maps or charts to intuitively depict the spatial distribution pattern of production capacity in the target area, help identify high-yield and low-yield areas, and provide a strong basis for resource planning; by drawing time series charts, deeply analyze the dynamic change trend of natural gas hydrate production capacity in the future, predict the increase or decrease in production capacity, and provide a reference for the formulation of long-term production strategies.

[0084] Overall, this embodiment has the following beneficial effects: This application predicts the mining capacity of natural gas hydrates directly based on the mining capacity prediction model based on the target influencing variables, and can quickly and effectively obtain the prediction results of natural gas hydrate mining capacity. For the mining capacity prediction model, training the learning model based on the classification results of the natural gas hydrate data sample set can enable the model to initially learn the characteristics and laws in the data. Optimizing the initial model by means of group intelligence optimization can make up for the problem of low data quality to a certain extent. Finally, the transition model is trained based on the loss function to obtain the final mining capacity prediction model. This method ensures that the model can fully learn the key information in the data during the training process, thereby improving the accuracy of the prediction; among them, determining the positive sample pairs and negative sample pairs according to the similarity values ​​of environmental factors and mining capacity helps the model learn the law of capacity changes under different mining conditions, thereby making up for the problem of insufficient data volume; at the same time, the loss function also enhances the model's ability to recognize similar data by reducing the distance between the data in the positive sample pairs, which helps the model to make more consistent predictions when facing similar mining conditions, thereby improving the stability of the prediction and providing strong data support for drilling engineering.

[0085] Embodiment three: The embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for predicting the production capacity of natural gas hydrate mining; Wherein, the method for predicting the production capacity of natural gas hydrate exploitation, if implemented in the form of a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0086] The above are preferred embodiments of the present invention. It should be noted that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting the production capacity of natural gas hydrate production, characterized in that: include: Obtain target influencing variables associated with environmental factors in the target area; Based on the target influencing variables, the mining capacity of natural gas hydrates is predicted according to the mining capacity prediction model to obtain the prediction result of the mining capacity of natural gas hydrates in the target area; wherein, the mining capacity prediction model is obtained by training a transition model based on a loss function; the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization; the initial model is obtained by establishing a natural gas hydrate data sample set based on the influencing variables associated with environmental factors in the target area, and training the learning model according to the classification results of the natural gas hydrate data sample set; the loss function is established by increasing the distance between the positive sample pairs and the negative sample pairs in the influencing variables, and reducing the distance between the data in the positive sample pairs; the positive sample pairs and the negative samples are determined according to the similarity values ​​of environmental factors and mining capacity.

2. A method for predicting the production capacity of natural gas hydrate production according to claim 1, characterized in that: The initial model is obtained by establishing a natural gas hydrate data sample set based on the influencing variables associated with environmental factors in the target area, and training the learning model according to the classification results of the natural gas hydrate data sample set, specifically: Converting the influencing variables associated with the environmental factors in the target area into a text sequence, and obtaining a global feature representing the entire text or sentence pair in the text sequence by adding a preset marker to the first position of a sentence in the text sequence; Generate an embedding vector according to the global feature, and form a processed text by the global feature and the embedding vector; Combining the processed text with the actual classification of the influencing variables to construct the natural gas hydrate data sample set; wherein the natural gas hydrate data sample set includes global features representing sentence pairs and separators for distinguishing different sentence pairs; Mapping the text features of the natural gas hydrate data sample set to the target classification of the influencing variable through a natural language processing model to obtain a classification result of the natural gas hydrate data sample set; The learning model is trained according to the classification results of the natural gas hydrate data sample set to obtain the initial model.

3. A method for predicting the production capacity of natural gas hydrate production according to claim 2, characterized in that: The text features of the natural gas hydrate data sample set are mapped to the target classification of the influencing variables through a natural language processing model to obtain the classification result of the natural gas hydrate data sample set, which is specifically: Based on the natural language processing model, the natural gas hydrate data sample set is used as an input of an embedding layer for semantic recognition to obtain a target embedding vector set of several labels in the natural gas hydrate data sample set; By encoding the target embedding vector set from both positive and negative directions, long-distance dependencies and contextual semantic knowledge in the text are captured to obtain a feature representation of the natural gas hydrate data sample set; Based on the classification layer of the natural language processing model, the feature representation is mapped to the classification space of the influencing variable, and the prediction probability of the natural gas hydrate data sample set on each influencing variable category is obtained by calculating the similarity or distance between the feature representation and a plurality of classification labels; According to the magnitude of the predicted probability, a classification label is assigned to each natural gas hydrate data sample in the natural gas hydrate data sample set to obtain a classification result of the natural gas hydrate data sample set.

4. A method for predicting the production capacity of natural gas hydrate production according to claim 2, characterized in that: The learning model is trained according to the classification results of the natural gas hydrate data sample set to obtain the initial model, which is specifically: Balancing the category distribution of the classification results by resampling to obtain a first classification result; Training a plurality of learning models according to the first classification result to obtain a result set, and fusing the result set in a weighted manner to obtain a prediction result; The prediction results are used as new features to train the meta-learner to obtain the initial model.

5. A method for predicting the production capacity of natural gas hydrate production according to claim 1, characterized in that: The mining capacity prediction model is obtained based on the loss function training transition model, specifically: Introducing a minimum margin, and establishing a first loss function according to the distance loss between the negative sample pair and the positive sample pair; wherein the minimum margin is used to control the relative dispersion between samples within the positive sample pair; Introducing a maximum margin, and establishing a second loss function according to the distance loss between the positive sample pair and a preset representative point; wherein the maximum margin is used to control the relative compactness between the internal samples of the positive sample pair, and the preset representative point represents the cluster center of the positive sample pair and the negative sample pair; Establishing the loss function based on the first loss function and the second loss function; The transition model is trained according to the loss function to obtain the mining capacity prediction model.

6. A method for predicting the production capacity of natural gas hydrate production according to claim 5, characterized in that: The first loss function and the second loss function are specifically: ； ； in, is the first loss function, is the second loss function, is the feature vector associated with the positive sample pair, is the feature vector associated with the negative sample pair, is the cluster center, and are the minimum margin and the maximum margin respectively.

7. A method for predicting the production capacity of natural gas hydrate production according to claim 1, characterized in that: The positive sample pair and the negative sample are determined based on similar values ​​of environmental factors and mining capacity, specifically: Based on the influencing variables associated with environmental factors in the target area, a sample pair whose similarity value of environmental factors is greater than a first standard and whose similarity value of mining capacity is greater than a second standard is defined as the positive sample pair; A sample pair whose similarity value of environmental factors is greater than the first standard but whose similarity value of mining capacity is less than or equal to the second standard is defined as the negative sample pair.

8. A method for predicting the production capacity of natural gas hydrate production according to claim 1, characterized in that: The transition model is obtained by optimizing the initial model by means of swarm intelligence optimization, specifically: randomly generating a group of particles, and initializing a position vector of each particle in the group of particles to a random value of a model parameter in the initial model; According to the data iteration method, the personal best position of each particle is updated, and the global best position of the particle group to which each particle belongs is updated; If the evaluation result of the fitness function converges to a preset range, the data iteration is stopped and the current global optimal position of the particle is output; wherein the fitness function is used to evaluate the quality of each particle; The current global optimal position is used as a parameter configuration to establish the transition model.

9. A method for predicting the production capacity of natural gas hydrate production according to claim 8, characterized in that: The fitness function is specifically: ； in, is the objective function value, is the coefficient of the L2 regularization term, is the L2 norm of the parameter vector x, is the coefficient of the diversity promotion term, Represents particles The measurement result of the average distance to other particles in the particle group; is related to the number of iterations The coefficient of correlation, which changes as the iteration proceeds; is a preset function that varies with the number of iterations.

10. A natural gas hydrate production capacity prediction system, characterized in that: Includes data module and prediction module; Wherein, the data module is used to obtain target influencing variables associated with environmental factors in the target area; The prediction module is used to predict the production capacity of natural gas hydrates based on the target influencing variables and the production capacity prediction model, and obtain the prediction result of the production capacity of natural gas hydrates in the target area; wherein the production capacity prediction model is obtained by training a transition model based on a loss function; the transition model is obtained by optimizing the initial model by means of swarm intelligence optimization; the initial model is obtained by establishing a natural gas hydrate data sample set based on the influencing variables associated with environmental factors in the target area, and training the learning model according to the classification results of the natural gas hydrate data sample set; the loss function is established by increasing the distance between the positive sample pairs and the negative sample pairs in the influencing variables, and reducing the distance between the data in the positive sample pairs; the positive sample pairs and the negative samples are determined according to the similarity values ​​of environmental factors and production capacity.

Images