Salt cavern gas storage site selection evaluation method based on multiple factors

Abstract

The invention discloses a salt cavern gas storage site selection evaluation method based on multiple factors, belongs to the technical direction of salt cavern gas storage site selection, can more accurately capture complex nonlinear coupling relations among geology, environment and engineering factors through an innovative fractional order neural network model, remarkably improves the accuracy of site selection evaluation, and improves the site selection evaluation accuracy. Particularly, when the data scale is increased, the classification accuracy is remarkably improved, multi-modal data can be effectively processed through dynamic weighted projection and a fractional order convolution kernel, the nonlinear relation and space-time dependency between features are reserved, the problem that feature relevance is lost due to simple splicing or standardized processing in a traditional method is solved, and the classification accuracy is improved. Feature extraction and modeling are carried out by adopting a fractional derivative model, so that the model can fully reflect historical dependence of geological parameters such as salt rock, the modeling capability of the model on a long-term geological process is enhanced, and dynamic structure evolution and adaptive weight adjustment are carried out.

Description

Technical Field

[0001] The present invention belongs to the technical direction of salt cavern gas storage site selection, and specifically relates to a salt cavern gas storage site selection evaluation method based on multiple factors. Background Art

[0002] With increasing energy demand and heightened environmental protection requirements, salt-cavern gas storage (NGS) as a key energy storage facility is attracting increasing attention. The site selection of a salt-cavern gas storage facility directly impacts its subsequent safety, economic viability, and sustainable development. Therefore, accurate and scientific assessment of site conditions is crucial for its construction. Traditional site selection evaluation methods, most of which rely on expert experience, historical data, and simplified computational models, fail to fully consider the complex interactions among geological, environmental, and engineering factors. These factors are inherently multimodal, dimensionally diverse, and nonlinearly coupled, making traditional methods often inadequate when faced with massive, multidimensional, and complex data. Furthermore, as data volumes continue to grow, the computational efficiency of traditional methods is insufficient to meet practical needs. This is particularly true when processing geological data involving long-term dependencies, where traditional methods lack adaptability.

[0003] The existing patent publication number CN119443618A discloses a supplementary site selection method for elderly care service institutions based on multi-source data and machine learning, which relates to the technical field of site selection for elderly care service institutions. The present invention uses machine learning to select sites for multi-source data on elderly care services, reducing labor costs and dependence on manual analysis, and improving the degree of automation in site selection. It optimizes multiple machine learning models and selects the optimal machine learning model for site selection. It has a wide range of applications and can dynamically respond to changing needs, ensuring that the coverage of elderly care facility services matches actual needs and adapts to dynamic changes in urban and population structures. However, it cannot effectively capture nonlinear coupling and dynamic correlation in the data, resulting in the inability to fully model the complex relationships between features, affecting the accuracy of the model. It cannot effectively handle long-term dependencies and non-local characteristics, especially in the modeling of long-term processes such as geological evolution. It cannot capture the memory effect of historical data, resulting in poor adaptability of the model to changes in geological parameters and inability to dynamically adjust according to changes in data distribution. As a result, the network cannot adapt to changes in feature distribution when processing data with non-steady-state characteristics, thereby affecting the accuracy and adaptability of the model. The lack of effective consideration of physical constraints (such as salt cavern stability, groundwater level, and engineering costs, etc.) may result in the model not meeting the requirements of engineering feasibility in practical applications, and some improvements are needed.

[0004] Therefore, it is necessary to design a practical salt cavern gas storage site selection evaluation method based on multiple factors. Summary of the Invention

[0005] The purpose of the present invention is to provide a salt cavern gas storage site selection evaluation method based on multiple factors for existing devices, so as to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solution: a multi-factor based salt cavern gas storage site selection evaluation method, the method comprising the following steps:

[0007] S1. Collection and labeling of training data

[0008] The training data comes from multi-source heterogeneous data of salt cavern gas storage site selection evaluation;

[0009] S2. Data Labeling

[0010] Data annotation is based on classification labels such as "suitable site" or "unsuitable site" based on actual site selection evaluation. The annotation method mainly relies on expert evaluation, combined with comprehensive analysis of various factors, and uses historical data and expert experience for annotation. The label of each sample is based on a comprehensive judgment of the site selection conditions based on multiple factors (geology, environment, engineering, etc.). The annotation categories include:

[0011] Category 1 (suitable site selection): indicates that the conditions in the area meet the requirements for gas storage construction;

[0012] Category 2 (unsuitable site selection): indicates that the conditions in the area do not meet the requirements for gas storage construction.

[0013] The labeled categories provide supervisory signals for the training objectives of the neural network, ensuring that the subsequent fractional-order neural network model can learn different site selection judgments based on the input features.

[0014] S3. Site selection evaluation model training

[0015] A fractional order neural network based on model structure evolution is used to evaluate the site selection of salt cavern gas storage. In one embodiment, the structure of the fractional order neural network is:

[0016] Input layer: receives data from geological, environmental and engineering factors, which are pre-processed and feature fused before being input into the network;

[0017] Feature extraction layer: fractional-order convolution kernel is used to extract features from data;

[0018] Fractional feature transformation layer: performs nonlinear transformation on the input features and uses fractional derivative models to model long-term dependencies of the data.

[0019] Hidden layer: contains several layers (such as 3 layers) of convolutional layers and fully connected layers, which perform nonlinear transformations through activation functions to extract higher-level features;

[0020] Output layer: Output classification results (through Softmax activation function), such as "suitable site selection" or "unsuitable site selection";

[0021] S4. Model reasoning.

[0022] The present invention further illustrates that the multi-source heterogeneous data includes three major categories of factors: geological, environmental, and engineering, as follows:

[0023] Geological factors (G): such as salt layer thickness, permeability, rock strength, etc. These data are obtained through geological exploration and drilling measurements;

[0024] Environmental factors (E): such as earthquake frequency, groundwater level, temperature, etc., usually collected through earthquake monitoring stations, meteorological stations, environmental monitoring stations, etc.

[0025] Engineering factors (C): such as construction costs, wellbore stability, infrastructure conditions, etc., mainly come from engineering design documents, historical construction projects and on-site engineering monitoring data.

[0026] The present invention further illustrates that the training process of the fractional-order neural network is as follows:

[0027] S301: Multi-source heterogeneous data fusion and adaptive weighted projection

[0028] Geological, environmental, and engineering factor data are multimodal, have large dimensional differences, and are characterized by nonlinear coupling. Conventional methods use simple splicing or standardization, which results in the loss of dynamic correlation between features.

[0029] The dynamic projection method based on maximum information entropy is used to map the three types of data into a unified feature space, thereby realizing nonlinear feature fusion and dynamic association preservation, which can be expressed as:

[0030] is the initial fusion feature vector of the i-th sample, i is a positive integer, i = 1, 2, ..., N, N is the total number of samples, and is used to characterize the joint distribution characteristics of the three types of factors (geology (G), environment (E), and engineering (C)) in the spatiotemporal dimension;

[0031] d∈{G,E,C} is the data category index, G represents geological factors (such as salt layer thickness and permeability), E represents environmental factors (such as seismic activity frequency and groundwater level), and C represents engineering factors (such as construction cost and wellbore stability);

[0032] is the dynamic weight coefficient of the t-th iteration, t is a positive integer, reflecting the importance of different factor categories to the site selection classification label Y. For example, geological factors (such as salt layer thickness) usually have a greater impact on site selection, but environmental factors (such as seismic activity) may have a sudden increase in weight in a specific area. The dynamic weight can adaptively adjust the dominant factors in different exploration stages. The calculation method is,

[0033] I(·) is the mutual information, X d is the data set of the d-th factor, X d′ is the data set of the d′th factor, d is a positive integer, d′ is a positive integer;

[0034] X i,d is the original data of the i-th sample under the d-type factors (G: geology, E: environment, C: engineering), such as X i,G represents the geological parameters of the i-th sample;

[0035] To normalize the data of the d-th factor, eliminate the dimension difference (such as the scale difference between permeability (0.1-10mD) and construction cost (million-level)), and ensure the comparability of the fusion features, min(X d ) and max(X d ) are the minimum and maximum values ​​of this type of data respectively;

[0036] Γ(·) is the Gamma function, Γ(X i,d ) represents the fractional order data transformation function defined by the Gamma function, expressed as Where α is the fractional strength coefficient and β∈(0,1) is the fractional derivative order, which is used to capture the memory effect of geological evolution processes such as salt rock creep. However, salt rock creep has time-dependent and non-local characteristics (such as historical stress accumulation), and integer derivatives cannot model such nonlinear behaviors. Represents X i,d Caputo fractional derivative of order β about iteration t.

[0037] The present invention further illustrates that the training process of the fractional-order neural network also includes:

[0038] S302: Constructing a fractional-order feature extraction layer

[0039] The samples after multi-source data fusion are used as the input of the fractional-order feature extraction layer to perform nonlinear transformation of the data features;

[0040] Conventional neural networks use integer-order derivatives to construct feature transformations, which cannot characterize the non-local dependence of salt cave geological parameters;

[0041] The fractional-order convolution kernel is designed and the non-local dependency modeling of features is achieved by adopting the Grünwald-Letnikov fractional-order derivative discretization method, which is expressed as:

[0042] is the output feature of the i-th sample after fractional-order convolution, and the superscript (1) represents the first hidden layer;

[0043] Sig(·) is the Sigmoid activation function, which maps the output to the (0,1) interval;

[0044] j=1,...,M is the input feature channel index, j is a positive integer, and M is the input feature dimension (i.e. dimensions);

[0045] is the fractional-order convolution kernel weight matrix, that is, the convolution kernel weight matrix of the i-th output channel and the j-th input channel. The superscript (frac) indicates the fractional-order characteristic. Its initialization adopts the improved Grünwald-Letnikov discretization method, which enables the fractional-order neural network to capture both short-term mutations (such as seismic activity) and long-term trends (such as salt rock creep). It is expressed as,

[0046]

[0047] in is the benchmark integer-order convolution kernel, Represents the kth weight parameter of the base integer-order convolution kernel;

[0048] is the kth weight parameter of the fractional-order convolution kernel;

[0049] (-1) k represents a sign-alternating term;

[0050] β is the order of the fractional derivative, β∈(0,1);

[0051] Characterize the Grünwald-Letnikov fractional discretization coefficients;

[0052] Γ(·) is the Gamma function;

[0053] * β is a fractional-order convolution operator. The calculation method introduces the weighted accumulation of historical eigenvalues ​​through the integral term. For example, the evolution of salt rock permeability has a time lag effect (such as the impact of historical water injection pressure on current permeability). The integral operation enhances the model's memory of historical states, which can be expressed as:

[0054]

[0055] in is the Caputo fractional derivative, which is used to model the long-term dependence of salt cavern parameters;

[0056] is the initial fusion feature vector of the jth sample;

[0057] τ represents the time delay variable, dτ represents the integral differential element;

[0058] Represents the value of the jth initial fusion feature at time t-τ (iteration);

[0059] Represents the integration of the historical time window (number of iterations) [0, t];

[0060] b i is the bias term of the i-th output channel.

[0061] The present invention further illustrates that the training process of the fractional-order neural network also includes:

[0062] S303, dynamic structural evolution of fractional-order neural networks

[0063] In view of the fact that data distribution in site selection evaluation changes with the progress of geological exploration, a topological evolution strategy based on complex network theory is adopted. The specific method is as follows:

[0064] a) Node growth criteria:

[0065] When the feature dimension D l KL divergence at layer l l Exceeding the threshold θ kl When the number of new nodes is By detecting feature distribution shifts (e.g., distribution changes caused by newly added exploration data) through KL divergence, the network width is dynamically expanded, so that geological exploration data gradually accumulates as the project progresses, and node growth adapts to the non-stationary characteristics of data distribution.

[0066] Where KL l is the KL divergence of the feature distribution of the lth layer, which measures the difference in feature distribution of adjacent layers. It is calculated as follows:

[0067] KL l =D KL (P(Z (l) )||P(Z (l-1) )),

[0068] η kl is the growth rate factor, which controls the node growth sensitivity (the smaller the value, the faster the growth);

[0069] Indicates rounding up;

[0070] D l is the feature dimension of the lth layer;

[0071] θ kl is the KL divergence threshold, preferably set to 0.1;

[0072] Δn is the number of newly added nodes;

[0073] D KL (·) represents the KL divergence function;

[0074] P(Z (l) ) represents the probability distribution of the feature vector of the lth layer;

[0075] P(Z (l-1) ) represents the probability distribution of the l-1th layer feature vector;

[0076] Z (l) is the output feature of the lth layer;

[0077] Z (l-1) is the output feature of the l-1th layer.

[0078] b) Weight evolution adjustment:

[0079] According to the mutual information and gradient significance between nodes, the connection weights are dynamically adjusted to achieve adaptive optimization of the network topology. The weight evolution equation is defined as:

[0080] Where, is the connection weight from the i-th node to the j-th node in the l-th layer at the t+1-th iteration;

[0081] is the connection weight from the i-th node to the j-th node in the l-th layer at the t-th iteration;

[0082] γ is the weight learning rate, which controls the weight update amplitude;

[0083] is the symbol of partial derivative;

[0084] L is the loss function;

[0085] Represents the partial derivative of the loss function with respect to the connection weight from the i-th node to the j-th node in the l-th layer, also known as the gradient;

[0086] As the weight decay term, an exponential term is used to prevent over-adjustment of salt cavern parameter-sensitive weights, suppress over-adjustment of sensitive weights (such as salt cavern stability-related parameters), and prevent the model from overfitting due to the hard boundaries of engineering constraints;

[0087] λ w is the attenuation coefficient, which suppresses over-adjustment of sensitive weights.

[0088] The present invention further illustrates that the training process of the fractional-order neural network also includes:

[0089] S304, perform fractional order parameter adaptive optimization

[0090] Conventional fractional-order neural networks have fixed fractional orders, which can easily lead to insufficient adaptability of the model to geological evolution processes;

[0091] A two-stage optimization is adopted to achieve dynamic adjustment of fractional order through gradient backpropagation and physical constraint joint optimization. The specific steps are as follows;

[0092] a) Coarse adjustment stage:

[0093] The creep rate of salt rock obeys the fractional-order dynamics law. Initialization needs to ensure that the model conforms to the geophysical prior. Based on the constraints of geophysical prior knowledge, fractional-order initialization is performed to achieve preliminary alignment of geological indicators and model parameters, which can be expressed as:

[0094] Where, φ k For key geological indicators, such as salt rock permeability;

[0095] β (0) is the initial fractional order, which is determined by minimizing the difference between the fractional derivative of the key geological indicator and the observed value;

[0096] is the observed value, representing the observed geological indicator value, such as the measured permeability;

[0097] φ k is the geological indicator value predicted by the model;

[0098] Represents φ k The β-order fractional derivative of ;

[0099] K is the number of key geological indicators (such as salt rock strength, creep rate, etc.);

[0100] argmin β To minimize the objective function to determine the optimal β;

[0101] b) Fine-tuning stage:

[0102] End-to-end optimization is achieved through differentiable programming. According to the gradient backpropagation and the chain rule of fractional derivatives, the dynamic adjustment of the fractional order is achieved, which can be expressed as:

[0103] Where, represents the gradient of the loss function with respect to the fractional order;

[0104] L layer is the total number of layers of the fractional-order neural network;

[0105] Represents the gradient of the loss function with respect to the output of layer l;

[0106] Represents the gradient of the output of the lth layer to β, representing the sensitivity of the output of the lth layer to β;

[0107] Γ(1-β) is the Gamma function, which is used for fractional-order integral normalization;

[0108] Z (l) Represents the output feature vector of the lth layer;

[0109] Z (l-1) Represents the output feature vector of the l-1th layer;

[0110] The fractional integral that represents historical characteristics reflects the long-term memory effect;

[0111] Furthermore, the fractional order of the next iteration is As the update gradient, the gradient descent method is used for updating.

[0112] The present invention further illustrates that the training process of the fractional-order neural network also includes:

[0113] S305. Construction of multi-objective loss function

[0114] Considering the balance between classification accuracy and engineering feasibility in site selection evaluation, a hybrid loss function is adopted to ensure classification accuracy while combining engineering feasibility constraints to balance different objectives, such as physical constraints including stability, water level and cost. The loss function is calculated as follows:

[0115] Where L is the loss function;

[0116] L ce is the cross entropy loss;

[0117] λ1 is the loss term weight, preferably set to 0.7;

[0118] ω m is the physical loss weight, satisfying

[0119] is the mth physical constraint item, m is a positive integer, and when m=1, 2, 3, they correspond to stability, water level, and cost constraints respectively. The calculation method is:

[0120] is the salt cavern stability coefficient With threshold The violation amount is to force the salt cavern stability coefficient to be no less than the threshold value (such as salt rock strength requirement).

[0121] The depth of groundwater level and the optimal value The absolute deviation of the optimized groundwater level is close to the optimal value of the project.

[0122] For project costs The indicator function of exceeding the limit, penalizing the cost exceeding the budget, directly constrains the feasibility of the project.

[0123] The upper limit of the project cost budget, for example, set it to 10000;

[0124] It is an indicator function that outputs 1 when the condition is met, otherwise it is 0.

[0125] The present invention further illustrates that the training process of the fractional-order neural network also includes:

[0126] S306, Memory-enhanced backpropagation

[0127] Salt cave site selection requires long-term monitoring data support. Fractional gradient can retain historical training state information and alleviate overfitting caused by insufficient data. In view of the long-range dependence characteristics of geological data, the present invention improves the back propagation algorithm and realizes the cumulative effect of historical gradient information by combining the fractional gradient term, which is expressed as:

[0128] In the formula, ΔW (l) is the update amount of the l-th layer weight;

[0129] η w is the basic learning rate, preferably set to 0.001;

[0130] μ is the fractional gradient gain coefficient, preferably set to 0.005;

[0131] is a fractional gradient term, which represents the fractional gradient of the loss function with respect to the weight. The Caputo definition is used for discretization calculation to enhance the memory of historical training states and improve the generalization performance under non-stationary data distribution. It is expressed as,

[0132]

[0133] N ce is the number of discretization steps;

[0134] k is a positive integer;

[0135] ∈ ce is a small perturbation step, such as 0.0001;

[0136] L(Wk∈ ce ) represents the weight W offset k∈ ce The loss value after

[0137] ∈ ce β represents the β-power of the perturbation step size;

[0138] S307, update the bias parameters of the fractional-order neural network

[0139] The fractional-order gradient descent algorithm is used to update the bias term of the fractional-order neural network, which is expressed as:

[0140] Where η b is the bias learning rate, preferably set to 0.001;

[0141] b i (t+1) is the value of the i-th bias term of the fractional-order neural network at the t+1th iteration;

[0142] b i (t) is the value of the i-th bias term of the fractional-order neural network at the t-th iteration;

[0143] is the gradient of the loss function with respect to the i-th bias term;

[0144] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0145] The present invention further illustrates that the model reasoning refers to using a trained fractional-order neural network to perform site selection evaluation classification, and the specific steps are:

[0146] S401. Input and preprocessing of new data

[0147] When new data is input into a trained fractional-order neural network, it will first undergo the same preprocessing steps as the training data. Specifically, through the dynamic projection method, the new data will be mapped to a unified feature space, the three types of data will be fused, and the dynamic correlation between different features will be maintained.

[0148] S402, Feature Extraction and Fractional Convolution

[0149] Through the fractional-order feature extraction layer, the new data is nonlinearly transformed to extract features with spatiotemporal dependence characteristics;

[0150] This layer processes the data using fractional-order convolution kernels to capture non-local dependencies in the data, especially changes over long time scales such as seismicity and salt creep.

[0151] S403. Network structure and dynamic weight adjustment

[0152] The network structure will be dynamically adjusted according to the characteristic distribution of new data. Based on the topological evolution strategy of the fractional-order neural network, the network structure and connection weights will be optimized in real time, so that the network can better adapt to different regions and different data distributions.

[0153] S404, Decision-making and Classification

[0154] After multiple layers of feature extraction and structural evolution, the final output is the site selection evaluation classification result;

[0155] Based on the characteristics of the new data, the network will calculate whether it belongs to the "suitable site selection" or "unsuitable site selection" category, and the output classification label will be used to support decision-making on the site selection of salt cavern gas storage.

[0156] Compared with the existing technology, the present invention has the following beneficial effects: the present invention, through the innovative fractional-order neural network model, can more accurately capture the complex nonlinear coupling relationship between geological, environmental and engineering factors, significantly improving the accuracy of site selection evaluation, especially when the data scale increases, the classification accuracy is significantly improved;

[0157] Through dynamic weighted projection and fractional-order convolution kernels, it can effectively process multimodal data, preserve the nonlinear relationship and spatiotemporal dependency between features, and solve the problem of feature correlation loss caused by simple splicing or standardization in traditional methods;

[0158] By using a fractional derivative model for feature extraction and modeling, the model can fully reflect the historical dependence of geological parameters such as salt rock, enhancing the model's ability to model long-term geological processes;

[0159] Through dynamic structural evolution and adaptive weight adjustment, the present invention avoids the redundant calculations of traditional fixed-structure networks when processing multi-scale geological features, improves computational efficiency, and as the amount of data increases, the increase in training time is significantly lower than that of traditional neural networks, highlighting the advantage of adaptive allocation of computing resources.

[0160] By introducing a multi-objective loss function and engineering feasibility constraints, the present invention ensures that the site selection of salt cavern gas storage not only has advantages in classification accuracy, but also has better guarantees in engineering feasibility, enabling the model to provide comprehensive support for actual site selection decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0161] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0162] Figure 1 It is a schematic diagram of the dynamic weighted fusion effect of the technology of the present invention;

[0163] Figure 2 It is a schematic diagram of the traditional average fusion effect of the present invention;

[0164] Figure 3 2. It is a schematic diagram of comparison of classification accuracy of the present invention;

[0165] Figure 4 It is a schematic diagram of the two-stage optimization path analysis of the present invention;

[0166] Figure 5 It is a schematic diagram of the fractional order of the present invention;

[0167] Figure 6 It is a schematic diagram of dynamic monitoring of network structure evolution of the present invention;

[0168] Figure 7 2 is a schematic diagram comparing the relationship between training time and data volume of the present invention. DETAILED DESCRIPTION

[0169] The following is a non-limiting detailed description of the technical solutions of the present invention in conjunction with preferred embodiments and the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0170] See also Figure 1-7 The present invention provides a technical solution: a multi-factor based salt cavern gas storage site selection evaluation method, the main steps of which are as follows:

[0171] S1. Collection and labeling of training data

[0172] The training data of the present invention comes from multi-source heterogeneous data of salt cavern gas storage site selection evaluation, which includes three categories of factors: geological, environmental, and engineering. The specific data are as follows:

[0173] Geological factors (G): such as salt layer thickness, permeability, rock strength, etc. These data are obtained through geological exploration and drilling measurements;

[0174] Environmental factors (E): such as earthquake frequency, groundwater level, temperature, etc., usually collected through earthquake monitoring stations, meteorological stations, environmental monitoring stations, etc.

[0175] Engineering factors (C): such as construction costs, wellbore stability, infrastructure conditions, etc., mainly come from engineering design documents, historical construction projects and on-site engineering monitoring data.

[0176] S2. Data Labeling

[0177] Data annotation is based on classification labels based on actual site selection evaluations, including labels such as "suitable site selection" or "unsuitable site selection." The annotation method relies primarily on expert evaluation, combined with a comprehensive analysis of various factors, using historical data and expert experience. Each sample's label is based on a comprehensive assessment of site selection conditions based on multiple factors (geology, environment, engineering, etc.). In one embodiment, the annotation categories include:

[0178] Category 1 (suitable site selection): indicates that the conditions in the area meet the requirements for gas storage construction;

[0179] Category 2 (unsuitable site selection): indicates that the conditions in the area do not meet the requirements for gas storage construction.

[0180] The labeled categories provide supervisory signals for the training objectives of the neural network, ensuring that the subsequent fractional-order neural network model can learn different site selection judgments based on the input features.

[0181] S3. Site selection evaluation model training

[0182] The present invention uses a fractional-order neural network based on model structure evolution to evaluate the site selection of salt cavern gas storage. In one embodiment, the structure of the fractional-order neural network is:

[0183] 1) Input layer: Receives data from geological, environmental, and engineering factors, which are pre-processed and feature-fused before being input into the network;

[0184] 2) Feature extraction layer: Fractional-order convolution kernels are used to extract features from the data. Unlike traditional convolutional neural networks, the present invention uses fractional-order convolution, which can better capture the non-local dependencies in the data.

[0185] 3) Fractional feature transformation layer: performs nonlinear transformation on the input features and uses fractional derivative models to model long-term dependencies of the data;

[0186] 4) Hidden layer: Contains several layers (e.g., 3 layers) of convolutional layers and fully connected layers, which perform nonlinear transformations through activation functions to extract higher-level features;

[0187] 5) Output layer: Outputs classification results (through the Softmax activation function), such as "suitable site selection" or "unsuitable site selection".

[0188] Specifically, the training process of the fractional-order neural network is as follows:

[0189] S301: Multi-source heterogeneous data fusion and adaptive weighted projection

[0190] Geological, environmental, and engineering factor data are multimodal, have large dimensional differences, and are characterized by nonlinear coupling. Conventional methods use simple splicing or standardization, which results in the loss of dynamic correlation between features.

[0191] The present invention adopts a dynamic projection method based on maximum information entropy to map the three types of data into a unified feature space, thereby realizing nonlinear feature fusion and dynamic association preservation, which can be expressed as:

[0192] is the initial fusion feature vector of the i-th sample, i is a positive integer, i = 1, 2, ..., N, N is the total number of samples, and is used to characterize the joint distribution characteristics of the three types of factors (geology (G), environment (E), and engineering (C)) in the spatiotemporal dimension;

[0193] d∈{G,E,C} is the data category index, G represents geological factors (such as salt layer thickness and permeability), E represents environmental factors (such as seismic activity frequency and groundwater level), and C represents engineering factors (such as construction cost and wellbore stability);

[0194] is the dynamic weight coefficient of the t-th iteration, t is a positive integer, reflecting the importance of different factor categories to the site selection classification label Y. For example, geological factors (such as salt layer thickness) usually have a greater impact on site selection, but environmental factors (such as seismic activity) may have a sudden increase in weight in a specific area. The dynamic weight can adaptively adjust the dominant factors in different exploration stages. The calculation method is,

[0195] I(·) is the mutual information, X d is the data set of the d-th factor, X d′ is the data set of the d′th factor, d is a positive integer, d′ is a positive integer;

[0196] X i,d is the original data of the i-th sample under the d-type factors (G: geology, E: environment, C: engineering), such as X i,G represents the geological parameters of the i-th sample;

[0197] To normalize the data of the d-th factor, eliminate the dimension difference (such as the scale difference between permeability (0.1-10mD) and construction cost (million-level)), and ensure the comparability of the fusion features, min(X d ) and max(X d ) are the minimum and maximum values ​​of this type of data respectively;

[0198] Γ(·) is the Gamma function, Γ(Xi,d ) represents the fractional order data transformation function defined by the Gamma function, expressed as Where α is the fractional strength coefficient and β∈(0,1) is the fractional derivative order, which is used to capture the memory effect of geological evolution processes such as salt rock creep. However, salt rock creep has time-dependent and non-local characteristics (such as historical stress accumulation), and integer derivatives cannot model such nonlinear behaviors. Represents X i,d Caputo fractional derivative of order β about iteration t.

[0199] S302: Constructing a fractional-order feature extraction layer

[0200] The samples after multi-source data fusion are used as the input of the fractional-order feature extraction layer to perform nonlinear transformation of the data features;

[0201] Conventional neural networks use integer-order derivatives to construct feature transformations, which cannot characterize the non-local dependence of salt cave geological parameters;

[0202] The present invention designs a fractional-order convolution kernel and realizes the non-local dependency modeling of features by adopting the Grünwald-Letnikov fractional-order derivative discretization method, which is expressed as:

[0203] is the output feature of the i-th sample after fractional-order convolution, and the superscript (1) represents the first hidden layer;

[0204] Sig(·) is the Sigmoid activation function, which maps the output to the (0,1) interval;

[0205] j=1,...,M is the input feature channel index, j is a positive integer, and M is the input feature dimension (i.e. dimensions);

[0206] is the fractional-order convolution kernel weight matrix, that is, the convolution kernel weight matrix of the i-th output channel and the j-th input channel. The superscript (frac) indicates the fractional-order characteristic. Its initialization adopts the improved Grünwald-Letnikov discretization method, which enables the fractional-order neural network to capture both short-term mutations (such as seismic activity) and long-term trends (such as salt rock creep). It is expressed as,

[0207]

[0208] in is the benchmark integer-order convolution kernel, Represents the kth weight parameter of the base integer-order convolution kernel;

[0209] is the kth weight parameter of the fractional-order convolution kernel;

[0210] (-1) k represents a sign-alternating term;

[0211] β is the order of the fractional derivative, β∈(0,1);

[0212] Characterize the Grünwald-Letnikov fractional discretization coefficients;

[0213] Γ(·) is the Gamma function;

[0214] * β is a fractional-order convolution operator. The calculation method introduces the weighted accumulation of historical eigenvalues ​​through the integral term. For example, the evolution of salt rock permeability has a time lag effect (such as the impact of historical water injection pressure on current permeability). The integral operation enhances the model's memory of historical states, which can be expressed as:

[0215]

[0216] in is the Caputo fractional derivative, which is used to model the long-term dependence of salt cavern parameters;

[0217] is the initial fusion feature vector of the jth sample;

[0218] τ represents the time delay variable, dτ represents the integral differential element;

[0219] Represents the value of the jth initial fusion feature at time t-τ (iteration);

[0220] Represents the integration of the historical time window (number of iterations) [0, t];

[0221] b i is the bias term of the i-th output channel.

[0222] S303, dynamic structural evolution of fractional-order neural networks

[0223] In view of the fact that data distribution in site selection evaluation changes with the progress of geological exploration, the present invention adopts a topological evolution strategy based on complex network theory. The specific method is as follows:

[0224] a) Node growth criteria:

[0225] When the feature dimension D l KL divergence at layer l l Exceeding the threshold θ kl When the number of new nodes is By detecting feature distribution shifts (e.g., distribution changes caused by newly added exploration data) through KL divergence, the network width is dynamically expanded, so that geological exploration data gradually accumulates as the project progresses, and node growth adapts to the non-stationary characteristics of data distribution.

[0226] Where KL l is the KL divergence of the feature distribution of the lth layer, which measures the difference in feature distribution of adjacent layers. It is calculated as follows:

[0227] KL l =D KL (P(Z (l) )||P(Z (l-1) )),

[0228] η kl is the growth rate factor, which controls the node growth sensitivity (the smaller the value, the faster the growth);

[0229] Indicates rounding up;

[0230] D l is the feature dimension of the lth layer;

[0231] θ kl is the KL divergence threshold, preferably set to 0.1;

[0232] Δn is the number of newly added nodes;

[0233] D KL (·) represents the KL divergence function;

[0234] P(Z (l) ) represents the probability distribution of the feature vector of the lth layer;

[0235] P(Z (l-1) ) represents the probability distribution of the l-1th layer feature vector;

[0236] Z (l) is the output feature of the lth layer;

[0237] Z (l-1) is the output feature of the l-1th layer.

[0238] b) Weight evolution adjustment:

[0239] According to the mutual information and gradient significance between nodes, the connection weights are dynamically adjusted to achieve adaptive optimization of the network topology. The weight evolution equation is defined as:

[0240] Where, is the connection weight from the i-th node to the j-th node in the l-th layer at the t+1-th iteration;

[0241] is the connection weight from the i-th node to the j-th node in the l-th layer at the t-th iteration;

[0242] γ is the weight learning rate, which controls the weight update amplitude;

[0243] is the symbol of partial derivative;

[0244] L is the loss function;

[0245] Represents the partial derivative of the loss function with respect to the connection weight from the i-th node to the j-th node in the l-th layer, also known as the gradient;

[0246] As the weight decay term, an exponential term is used to prevent over-adjustment of salt cavern parameter-sensitive weights, suppress over-adjustment of sensitive weights (such as salt cavern stability-related parameters), and prevent the model from overfitting due to the hard boundaries of engineering constraints;

[0247] λ w is the attenuation coefficient, which suppresses over-adjustment of sensitive weights.

[0248] S304, perform fractional order parameter adaptive optimization

[0249] Conventional fractional-order neural networks have fixed fractional orders, which can easily lead to insufficient adaptability of the model to geological evolution processes;

[0250] The present invention adopts a two-stage optimization, which realizes the dynamic adjustment of fractional order through gradient back propagation and physical constraint joint optimization. The specific steps are as follows:

[0251] a) Coarse adjustment stage:

[0252] The creep rate of salt rock obeys the fractional-order dynamics law. Initialization needs to ensure that the model conforms to the geophysical prior. Based on the constraints of geophysical prior knowledge, fractional-order initialization is performed to achieve preliminary alignment of geological indicators and model parameters, which can be expressed as:

[0253] Where, φ k For key geological indicators, such as salt rock permeability;

[0254] β (0) is the initial fractional order, which is determined by minimizing the difference between the fractional derivative of the key geological indicator and the observed value;

[0255] is the observed value, representing the observed geological indicator value, such as the measured permeability;

[0256] φ k is the geological indicator value predicted by the model;

[0257] Represents φ k The β-order fractional derivative of ;

[0258] K is the number of key geological indicators (such as salt rock strength, creep rate, etc.);

[0259] argmin β To minimize the objective function, we need to determine the optimal β.

[0260] b) Fine-tuning stage:

[0261] End-to-end optimization is achieved through differentiable programming. According to the gradient backpropagation and the chain rule of fractional derivatives, the dynamic adjustment of the fractional order is achieved, which can be expressed as:

[0262] Where, represents the gradient of the loss function with respect to the fractional order;

[0263] L layer is the total number of layers of the fractional-order neural network;

[0264] Represents the gradient of the loss function with respect to the output of layer l;

[0265] Represents the gradient of the output of the lth layer to β, representing the sensitivity of the output of the lth layer to β;

[0266] Γ(1-β) is the Gamma function, which is used for fractional-order integral normalization;

[0267] Z (l) Represents the output feature vector of the lth layer;

[0268] Z (l-1) Represents the output feature vector of the l-1th layer;

[0269] The fractional-order integral that characterizes historical characteristics reflects the long-term memory effect.

[0270] Furthermore, the fractional order of the next iteration is As the update gradient, the gradient descent method is used for updating.

[0271] S305. Construction of multi-objective loss function

[0272] Considering the balance between classification accuracy and engineering feasibility in site selection evaluation, a hybrid loss function is adopted to ensure classification accuracy while combining engineering feasibility constraints to balance different objectives, such as physical constraints including stability, water level and cost. The loss function is calculated as follows:

[0273] Where L is the loss function;

[0274] L ce is the cross entropy loss;

[0275] λ1 is the loss term weight, preferably set to 0.7;

[0276] ω m is the physical loss weight, satisfying

[0277] is the mth physical constraint item, m is a positive integer, and when m=1, 2, 3, they correspond to stability, water level, and cost constraints respectively. The calculation method is:

[0278] is the salt cavern stability coefficient With threshold The violation amount is to force the salt cavern stability coefficient to be no less than the threshold value (such as salt rock strength requirement).

[0279] The depth of groundwater level and the optimal value The absolute deviation of the optimized groundwater level is close to the optimal value of the project.

[0280] For project costs The indicator function of exceeding the limit, penalizing the cost exceeding the budget, directly constrains the feasibility of the project.

[0281] The upper limit of the project cost budget, for example, set it to 10000;

[0282] It is an indicator function that outputs 1 when the condition is met, otherwise it is 0.

[0283] S306, Memory-enhanced backpropagation

[0284] Salt cave site selection requires long-term monitoring data support. Fractional gradient can retain historical training state information and alleviate overfitting caused by insufficient data. In view of the long-range dependence characteristics of geological data, the present invention improves the back propagation algorithm and realizes the cumulative effect of historical gradient information by combining the fractional gradient term, which is expressed as:

[0285] In the formula, ΔW (l) is the update amount of the l-th layer weight;

[0286] η w is the basic learning rate, preferably set to 0.001;

[0287] μ is the fractional gradient gain coefficient, preferably set to 0.005;

[0288] is a fractional gradient term, which represents the fractional gradient of the loss function with respect to the weight. The Caputo definition is used for discretization calculation to enhance the memory of historical training states and improve the generalization performance under non-stationary data distribution. It is expressed as,

[0289]

[0290] N ce is the number of discretization steps;

[0291] k is a positive integer;

[0292] ∈ ce is a small perturbation step, such as 0.0001;

[0293] L(Wk∈ ce ) represents the weight W offset k∈ ce The loss value after

[0294] ∈ ce β represents the perturbation step size to the power of β.

[0295] S307, update the bias parameters of the fractional-order neural network

[0296] The fractional-order gradient descent algorithm is used to update the bias term of the fractional-order neural network, which is expressed as:

[0297] Where η b is the bias learning rate, preferably set to 0.001;

[0298] b i (t+1) is the value of the i-th bias term of the fractional-order neural network at the t+1th iteration;

[0299] b i (t) is the value of the i-th bias term of the fractional-order neural network at the t-th iteration;

[0300] is the gradient of the loss function with respect to the i-th bias term.

[0301] 8. Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0302] S4. Model Reasoning

[0303] Model inference refers to the use of trained fractional-order neural networks to perform site selection evaluation and classification. The specific steps are as follows:

[0304] S401. Input and preprocessing of new data

[0305] When new data is input into a trained fractional-order neural network, it will first undergo the same preprocessing steps as the training data. Specifically, through the dynamic projection method, the new data will be mapped to a unified feature space, the three types of data will be fused, and the dynamic correlation between different features will be maintained.

[0306] S402, Feature Extraction and Fractional Convolution

[0307] Through the fractional-order feature extraction layer, the new data is nonlinearly transformed to extract features with spatiotemporal dependence characteristics;

[0308] This layer processes the data using fractional-order convolution kernels to capture non-local dependencies in the data, especially changes over long time scales such as seismicity and salt creep.

[0309] S403. Network structure and dynamic weight adjustment

[0310] The network structure will be dynamically adjusted according to the characteristic distribution of new data. Based on the topological evolution strategy of the fractional-order neural network, the network structure and connection weights will be optimized in real time, so that the network can better adapt to different regions and different data distributions.

[0311] S404, Decision-making and Classification

[0312] After multiple layers of feature extraction and structural evolution, the final output is the site selection evaluation classification result;

[0313] Based on the characteristics of the new data, the network will calculate whether it belongs to the "suitable site selection" or "unsuitable site selection" category, and the output classification label will be used to support decision-making on the site selection of salt cavern gas storage.

[0314] In one embodiment, to verify the advantages of the technology of the present invention, the following experiments were performed:

[0315] The effects of different data fusion methods were compared through three-dimensional surface visualization, verifying the advantages of the dynamic weighted projection mechanism in preserving the correlation of multi-source features. The experiment compared the method of combining dynamic weight allocation with fractional-order transformation proposed in this invention with the traditional mean fusion method. The results showed that the traditional method exhibited a single linear growth pattern in the feature space and had difficulty capturing the nonlinear coupling relationship between geological, environmental, and engineering factors. The multi-modal distribution surface formed by this technology revealed the accurate characterization of the difference in feature importance due to dynamic weights. The surface fluctuation characteristics directly reflected the ability of fractional-order transformation to model the historical state memory effect during salt rock creep, proving that the method of this invention can effectively maintain the dynamic correlation characteristics of multimodal data.

[0316] By comparing the classification accuracy of different models in the salt cavern site selection task, the technical advantages of fractional-order neural networks in multi-source heterogeneous data fusion were verified. The experiment compared the present invention with support vector machines, random forests, and traditional integer-order convolutional neural networks. The results showed that the accuracy of the present invention was significantly higher than that of other methods, especially when the data scale increased. This shows that the fractional-order operator effectively solves the problem of feature correlation loss caused by simple splicing or standardization in traditional methods through dynamic weight adjustment and nonlinear feature fusion. It can more accurately depict the complex coupling relationship between geological, environmental and engineering factors, and verify the model's dynamic modeling ability for multimodal data.

[0317] The adaptive adjustment process of the fractional order is analyzed through the parameter optimization trajectory diagram. It shows that this technology makes the fractional derivative order converge along the theoretical optimal path through physical law matching in the coarse adjustment stage and data-driven optimization in the fine adjustment stage. The oscillatory convergence characteristics presented in the optimization trajectory reflect the model's adaptive ability in balancing the long-term memory effect of geological parameters with real-time engineering constraints, achieving a deep integration of physical mechanisms and data characteristics.

[0318] By analyzing the classification accuracy trends of the model under different order values, it was found that the model reached its peak performance within a specific range, and the adaptive mechanism was able to dynamically adjust the order to the optimal range based on the evolution characteristics of geological parameters. This proved that the proposed strategy of combining physical constraints with gradient backpropagation optimization overcomes the lack of adaptability caused by the fixed order of traditional fractional-order networks, achieves accurate modeling of long-term geological processes such as salt rock creep, and significantly improves the model's generalization ability for non-steady-state data distributions.

[0319] The network structure evolution process is presented using a heat map combined with topological growth point markers. The adaptability of the dynamically evolving network is compared with that of the fixed-structure network. By monitoring the feature distribution difference index of each network layer, the experiment found that the traditional fixed-structure network experienced feature space mismatch in the middle and late stages of training. However, the trigger mechanism of this technology based on the divergence threshold can autonomously identify the distribution changes of geological exploration data and achieve dynamic optimization of the topological structure through node growth and weight reorganization. The structural growth markers that appear at specific rounds in the heat map intuitively demonstrate the network's ability to capture the spatiotemporal evolution characteristics of salt cavern parameters. The stable low-divergence regions of the deep network verify the effectiveness of the evolutionary strategy in ensuring model convergence.

[0320] By comparing the relationship between training time and data volume of different models, the advantage of the dynamic structural evolution strategy in computational efficiency is verified. Experimental results show that as the amount of data increases, the increase in training time of the present invention is significantly lower than that of the traditional integer-order convolutional neural network, indicating that the network topology structure dynamically adjusts the node mechanism according to the difference in feature distribution, avoiding redundant calculations of fixed structure networks when processing multi-scale geological features, and achieving adaptive allocation of computing resources while ensuring accuracy, highlighting the engineering practicality of the model in complex geological scenarios.

[0321] In the description of the present invention, it should be understood that the terms "up", "down", "front", "back", "left", "right", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as a limitation on the present invention.

[0322] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will appreciate that modifications may be made to the technical solutions described in the aforementioned embodiments, or that some of the technical features may be replaced with equivalents. Such modifications or replacements do not deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-factor based salt cavern gas storage site selection evaluation method, characterized in that: The method comprises the following steps: S1. Collection and labeling of training data The training data comes from multi-source heterogeneous data of salt cavern gas storage site selection evaluation; S2. Data Labeling Data annotation is based on classification labels such as "suitable site" or "unsuitable site" based on actual site selection evaluations. This labeling relies primarily on expert evaluation, combined with a comprehensive analysis of various factors, using historical data and expert experience. Each sample's label is based on a comprehensive assessment of site selection conditions based on multiple factors (geology, environment, engineering, etc.). The labeling categories include: Category 1 (suitable site selection): indicates that the conditions in the area meet the requirements for gas storage construction; Category 2 (unsuitable site selection): indicates that the conditions in the area do not meet the requirements for gas storage construction. The labeled categories provide supervisory signals for the training objectives of the neural network, ensuring that the subsequent fractional-order neural network model can learn different site selection judgments based on the input features. S3. Site selection evaluation model training A fractional order neural network based on model structure evolution is used to evaluate the site selection of salt cavern gas storage. In one embodiment, the structure of the fractional order neural network is: Input layer: receives data from geological, environmental and engineering factors, which are pre-processed and feature fused before being input into the network; Feature extraction layer: fractional-order convolution kernel is used to extract features from data; Fractional feature transformation layer: performs nonlinear transformation on the input features and uses fractional derivative models to model long-term dependencies of the data. Hidden layer: contains several layers (such as 3 layers) of convolutional layers and fully connected layers, which perform nonlinear transformations through activation functions to extract higher-level features; Output layer: Output classification results (through Softmax activation function), such as "suitable site selection" or "unsuitable site selection"; S4. Model reasoning.

2. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 1, characterized in that: The multi-source heterogeneous data includes three categories of factors: geological, environmental, and engineering, as follows: Geological factors (G): such as salt layer thickness, permeability, rock strength, etc. These data are obtained through geological exploration and drilling measurements; Environmental factors (E): such as earthquake frequency, groundwater level, temperature, etc., usually collected through earthquake monitoring stations, meteorological stations, environmental monitoring stations, etc. Engineering factors (C): such as construction costs, wellbore stability, infrastructure conditions, etc., mainly come from engineering design documents, historical construction projects and on-site engineering monitoring data.

3. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 2, characterized in that: The training process of the fractional-order neural network is as follows: S301: Multi-source heterogeneous data fusion and adaptive weighted projection Geological, environmental, and engineering factor data are multimodal, have large dimensional differences, and are characterized by nonlinear coupling. Conventional methods use simple splicing or standardization, which results in the loss of dynamic correlation between features. The dynamic projection method based on maximum information entropy is used to map the three types of data into a unified feature space, thereby realizing nonlinear feature fusion and dynamic association preservation, which can be expressed as: is the initial fusion feature vector of the i-th sample, i is a positive integer, i = 1, 2, ..., N, N is the total number of samples, and is used to characterize the joint distribution characteristics of the three types of factors (geology (G), environment (E), and engineering (C)) in the spatiotemporal dimension; d∈{G,E,C} is the data category index, G represents geological factors (such as salt layer thickness and permeability), E represents environmental factors (such as seismic activity frequency and groundwater level), and C represents engineering factors (such as construction cost and wellbore stability); is the dynamic weight coefficient of the t-th iteration, t is a positive integer, reflecting the importance of different factor categories to the site selection classification label Y. For example, geological factors (such as salt layer thickness) usually have a greater impact on site selection, but environmental factors (such as seismic activity) may have a sudden increase in weight in a specific area. The dynamic weight can adaptively adjust the dominant factors in different exploration stages. The calculation method is, I(·) is the mutual information, X d is the data set of the d-th factor, X d′ is the data set of the d′th factor, d is a positive integer, d′ is a positive integer; X i,d is the original data of the i-th sample under the d-type factors (G: geology, E: environment, C: engineering), such as X i,G represents the geological parameters of the i-th sample; To normalize the data of the d-th factor, eliminate the dimension difference (such as the scale difference between permeability (0.1-10mD) and construction cost (million-level)), and ensure the comparability of the fusion features, min(X d ) and max(X d ) are the minimum and maximum values ​​of this type of data respectively; Γ(·) is the Gamma function, Γ(X i,d ) represents the fractional order data transformation function defined by the Gamma function, expressed as Where α is the fractional strength coefficient and β∈(0,1) is the fractional derivative order, which is used to capture the memory effect of geological evolution processes such as salt rock creep. However, salt rock creep has time-dependent and non-local characteristics (such as historical stress accumulation), and integer derivatives cannot model such nonlinear behaviors. Represents X i,d Caputo fractional derivative of order β about iteration t.

4. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 3 is characterized in that: The training process of the fractional-order neural network also includes: S302: Constructing a fractional-order feature extraction layer The samples after multi-source data fusion are used as the input of the fractional-order feature extraction layer to perform nonlinear transformation of the data features; Conventional neural networks use integer-order derivatives to construct feature transformations, which cannot characterize the non-local dependence of salt cave geological parameters; The fractional-order convolution kernel is designed and the non-local dependency modeling of features is achieved by adopting the Grünwald-Letnikov fractional-order derivative discretization method, which is expressed as: is the output feature of the i-th sample after fractional-order convolution, and the superscript (1) represents the first hidden layer; Sig(·) is the Sigmoid activation function, which maps the output to the (0,1) interval; j=1,...,M is the input feature channel index, j is a positive integer, and M is the input feature dimension (i.e. dimensions); is the fractional-order convolution kernel weight matrix, that is, the convolution kernel weight matrix of the i-th output channel and the j-th input channel. The superscript (frac) indicates the fractional-order characteristic. Its initialization adopts the improved Grünwald-Letnikov discretization method, which enables the fractional-order neural network to capture both short-term mutations (such as seismic activity) and long-term trends (such as salt rock creep). It is expressed as, in is the benchmark integer-order convolution kernel, Represents the kth weight parameter of the base integer-order convolution kernel; is the kth weight parameter of the fractional-order convolution kernel; (-1) k represents a sign-alternating term; β is the order of the fractional derivative, β∈(0,1); Characterize the Grünwald-Letnikov fractional discretization coefficients; Γ(·) is the Gamma function; * β is a fractional-order convolution operator. The calculation method introduces the weighted accumulation of historical eigenvalues ​​through the integral term. For example, the evolution of salt rock permeability has a time lag effect (such as the impact of historical water injection pressure on current permeability). The integral operation enhances the model's memory of historical states, which can be expressed as: in is the Caputo fractional derivative, which is used to model the long-term dependence of salt cavern parameters; is the initial fusion feature vector of the jth sample; τ represents the time delay variable, dτ represents the integral differential element; Represents the value of the jth initial fusion feature at time t-τ (iteration); Represents the integration of the historical time window (number of iterations) [0, t]; b i is the bias term of the i-th output channel.

5. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 4 is characterized in that: The training process of the fractional-order neural network also includes: S303, dynamic structural evolution of fractional-order neural networks In view of the fact that data distribution in site selection evaluation changes with the progress of geological exploration, a topological evolution strategy based on complex network theory is adopted. The specific method is as follows: a) Node growth criteria: When the feature dimension D l KL divergence at layer l l Exceeding the threshold θ kl When the number of new nodes is By detecting feature distribution shifts (e.g., distribution changes caused by newly added exploration data) through KL divergence, the network width is dynamically expanded, so that geological exploration data gradually accumulates as the project progresses, and node growth adapts to the non-stationary characteristics of data distribution. Where KL l is the KL divergence of the feature distribution of the lth layer, which measures the difference in feature distribution of adjacent layers. It is calculated as follows: CL l =D KL (P(Z (l) )||P(Z (l-1) )), η kl is the growth rate factor, which controls the node growth sensitivity (the smaller the value, the faster the growth); Indicates rounding up; D l is the feature dimension of the lth layer; θ kl is the KL divergence threshold, preferably set to 0.1; Δn is the number of newly added nodes; D KL (·) represents the KL divergence function; P(Z (l) ) represents the probability distribution of the feature vector of the lth layer; P(Z (l-1) ) represents the probability distribution of the l-1th layer feature vector; Z (l) is the output feature of the lth layer; Z (l-1) is the output feature of the l-1th layer. b) Weight evolution adjustment: According to the mutual information and gradient significance between nodes, the connection weights are dynamically adjusted to achieve adaptive optimization of the network topology. The weight evolution equation is defined as: Where, is the connection weight from the i-th node to the j-th node in the l-th layer at the t+1-th iteration; is the connection weight from the i-th node to the j-th node in the l-th layer at the t-th iteration; γ is the weight learning rate, which controls the weight update amplitude; is the symbol of partial derivative; L is the loss function; Represents the partial derivative of the loss function with respect to the connection weight from the i-th node to the j-th node in the l-th layer, also known as the gradient; As the weight decay term, an exponential term is used to prevent over-adjustment of salt cavern parameter-sensitive weights, suppress over-adjustment of sensitive weights (such as salt cavern stability-related parameters), and prevent the model from overfitting due to the hard boundaries of engineering constraints; λ w is the attenuation coefficient, which suppresses over-adjustment of sensitive weights.

6. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 5, characterized in that: The training process of the fractional-order neural network also includes: S304, perform fractional order parameter adaptive optimization Conventional fractional-order neural networks have fixed fractional orders, which can easily lead to insufficient adaptability of the model to geological evolution processes; A two-stage optimization is adopted to achieve dynamic adjustment of fractional order through gradient backpropagation and physical constraint joint optimization. The specific steps are as follows; a) Coarse adjustment stage: The creep rate of salt rock obeys the fractional-order dynamics law. Initialization needs to ensure that the model conforms to the geophysical prior. Based on the constraints of geophysical prior knowledge, fractional-order initialization is performed to achieve preliminary alignment of geological indicators and model parameters, which can be expressed as: Where, φ k For key geological indicators, such as salt rock permeability; β (0) is the initial fractional order, which is determined by minimizing the difference between the fractional derivative of the key geological indicator and the observed value; is the observed value, representing the observed geological indicator value, such as the measured permeability; φ k is the geological indicator value predicted by the model; Represents φ k The β-order fractional derivative of ; K is the number of key geological indicators (such as salt rock strength, creep rate, etc.); argmin β To minimize the objective function to determine the optimal β; b) Fine-tuning stage: End-to-end optimization is achieved through differentiable programming. According to the gradient backpropagation and the chain rule of fractional derivatives, the dynamic adjustment of the fractional order is achieved, which can be expressed as: Where, represents the gradient of the loss function with respect to the fractional order; L layer is the total number of layers of the fractional-order neural network; Represents the gradient of the loss function with respect to the output of layer l; Represents the gradient of the output of the lth layer to β, representing the sensitivity of the output of the lth layer to β; Γ(1-β) is the Gamma function, which is used for fractional-order integral normalization; Z (l) Represents the output feature vector of the lth layer; Z (l-1) Represents the output feature vector of the l-1th layer; The fractional integral that represents historical characteristics reflects the long-term memory effect; Furthermore, the fractional order of the next iteration is As the update gradient, the gradient descent method is used for updating.

7. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 6, characterized in that: The training process of the fractional-order neural network also includes: S305. Construction of multi-objective loss function Considering the balance between classification accuracy and engineering feasibility in site selection evaluation, a hybrid loss function is adopted to ensure classification accuracy while combining engineering feasibility constraints to balance different objectives, such as physical constraints including stability, water level and cost. The loss function is calculated as follows: Where L is the loss function; L ce is the cross entropy loss; λ1 is the loss term weight, preferably set to 0.7; ω m is the physical loss weight, satisfying is the mth physical constraint item, m is a positive integer, and when m=1, 2, 3, they correspond to stability, water level, and cost constraints respectively. The calculation method is: is the salt cavern stability coefficient With threshold The violation amount is to force the salt cavern stability coefficient to be no less than the threshold value (such as salt rock strength requirement). The depth of groundwater level and the optimal value The absolute deviation of the optimized groundwater level is close to the optimal value of the project. For project costs The indicator function of exceeding the limit, penalizing the cost exceeding the budget, directly constrains the feasibility of the project. The upper limit of the project cost budget, for example, set it to 10000; It is an indicator function that outputs 1 when the condition is met, otherwise it is 0.

8. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 7, characterized in that: The training process of the fractional-order neural network also includes: S306, Memory-enhanced backpropagation Salt cave site selection requires long-term monitoring data support. Fractional gradient can retain historical training state information and alleviate overfitting caused by insufficient data. In view of the long-range dependence characteristics of geological data, the present invention improves the back propagation algorithm and realizes the cumulative effect of historical gradient information by combining the fractional gradient term, which is expressed as: In the formula, ΔW (l) is the update amount of the l-th layer weight; η w is the basic learning rate, preferably set to 0.001; μ is the fractional gradient gain coefficient, preferably set to 0.005; is a fractional gradient term, which represents the fractional gradient of the loss function with respect to the weight. The Caputo definition is used for discretization calculation to enhance the memory of historical training states and improve the generalization performance under non-stationary data distribution. It is expressed as, N ce is the number of discretization steps; k is a positive integer; ∈ ce is a small perturbation step, such as 0.0001; L(Wk∈ ce ) represents the weight W offset k∈ ce The loss value after ∈ ce β represents the β-power of the perturbation step size; S307, update the bias parameters of the fractional-order neural network The fractional-order gradient descent algorithm is used to update the bias term of the fractional-order neural network, which is expressed as: Where η b is the bias learning rate, preferably set to 0.001; b i (t+1) is the value of the i-th bias term of the fractional-order neural network at the t+1th iteration; b i (t) is the value of the i-th bias term of the fractional-order neural network at the t-th iteration; is the gradient of the loss function with respect to the i-th bias term; Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

9. The multi-factor based salt cavern gas storage site selection evaluation method according to claim 8, characterized in that: The model reasoning refers to using the trained fractional-order neural network to perform site selection evaluation classification. The specific steps are as follows: S401. Input and preprocessing of new data When new data is input into a trained fractional-order neural network, it will first undergo the same preprocessing steps as the training data. Specifically, through the dynamic projection method, the new data will be mapped to a unified feature space, the three types of data will be fused, and the dynamic correlation between different features will be maintained. S402, Feature Extraction and Fractional Convolution Through the fractional-order feature extraction layer, the new data is nonlinearly transformed to extract features with spatiotemporal dependence characteristics; This layer processes the data using fractional-order convolution kernels to capture non-local dependencies in the data, especially changes over long time scales such as seismicity and salt creep. S403. Network structure and dynamic weight adjustment The network structure will be dynamically adjusted according to the characteristic distribution of new data. Based on the topological evolution strategy of the fractional-order neural network, the network structure and connection weights will be optimized in real time, so that the network can better adapt to different regions and different data distributions. S404, Decision-making and Classification After multiple layers of feature extraction and structural evolution, the final output is the site selection evaluation classification result; Based on the characteristics of the new data, the network calculates whether it belongs to the "suitable site" or "unsuitable site" category. The output classification label will be used to support decision-making on the site selection of salt cavern gas storage.

Images