A dynamic control and safety assessment system and method for hydrogen blending ratio in natural gas

Abstract

This invention discloses a dynamic control and safety assessment system and method for the hydrogen blending ratio of natural gas, belonging to the field of clean energy transmission and distribution and safety control technology. It includes: collecting constraint data and real-time multimodal sensor data; fusing them to generate a fused initial dataset; performing domain adaptation initialization and parameter optimization to generate an initial symbiotic decision framework; performing state synchronization and risk path exploration to generate a predictive state space; executing hierarchical reinforcement learning self-evolutionary training to generate an evolutionary control strategy group; calculating the reality gap residual as a feedback signal to generate a self-healing optimization model; performing multi-dimensional risk quantification and dynamic ratio control; and outputting a safety assessment report. This invention employs a self-evolutionary symbiotic architecture of multimodal digital twins and hierarchical reinforcement learning, enabling dynamic optimization of the hydrogen blending ratio and real-time risk assessment, significantly improving the control stability and safety reliability of natural gas pipeline networks under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of clean energy transmission and distribution and safety control technology, and in particular to a system and method for dynamic control and safety assessment of the hydrogen blending ratio in natural gas. Background Technology

[0002] With the advancement of the clean energy transition, natural gas blending with hydrogen technology, as an emerging energy utilization method, has been widely applied in urban pipeline transportation, industrial fuel supply, and distributed energy systems. By digitally processing and optimizing the hydrogen blending ratio through computer systems, digital management and process control of energy transportation can be achieved. This technology relies on multimodal data fusion, model building, and computational frameworks to support real-time data processing and strategy generation for pipeline network operation, playing a crucial role in the energy infrastructure sector.

[0003] However, existing natural gas hydrogen blending control systems generally suffer from technical deficiencies. Under complex operating conditions, the data processing process struggles to adapt to multi-source heterogeneous inputs, leading to unstable control strategy generation. Traditional methods often rely on fixed-parameter models, which cannot effectively cope with real-time changing operating environments, affecting the overall system response capability. Furthermore, the safety assessment phase often neglects the calculation and integration of multi-dimensional risks, limiting the flexibility and reliability of proportional control. These problems constrain the system's performance in practical applications. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a dynamic control and safety assessment system and method for the hydrogen blending ratio of natural gas. Employing a self-evolving symbiotic architecture combining multimodal digital twins and hierarchical reinforcement learning, it enables dynamic optimization of the hydrogen blending ratio and real-time risk assessment, significantly improving the control stability and safety reliability of natural gas pipeline networks under complex operating conditions.

[0005] The above objectives can be achieved through the following approach:

[0006] A dynamic control and safety assessment system for hydrogen blending ratio in natural gas includes a data acquisition and fusion module for acquiring constraint data of the physical characteristics of the natural gas pipeline network and multimodal sensing data of the real-time operating status of the natural gas pipeline network, and fusing the constraint data and the multimodal sensing data to generate a fused initial dataset; a model building module for performing domain adaptation initialization and parameter optimization based on the fused initial dataset, constructing a multimodal digital twin model and a hierarchical reinforcement learning network, and generating an initial symbiotic decision framework; and a state synchronization exploration module for performing state synchronization and risk path exploration based on the initial symbiotic decision framework. The system explores and generates a predictive state space; a policy evolution module is used to perform hierarchical reinforcement learning self-evolutionary training based on the predictive state space, evolve the optimal policy through a virtual trial-and-error mechanism, and generate an evolutionary control policy group; a self-healing optimization module is used to calculate the real-world gap residual as a meta-level feedback signal based on the evolutionary control policy group, and adjust the multimodal digital twin model and the hierarchical reinforcement learning network in reverse to generate a self-healing optimization model; a risk regulation output module is used to perform multi-dimensional risk quantification and dynamic ratio regulation based on the self-healing optimization model, output a safety assessment report, and perform adaptive control of the hydrogen doping ratio.

[0007] Optionally, the data acquisition and fusion module includes: a data acquisition unit, used to acquire the constraint data of the physical characteristics of the natural gas pipeline network and the multimodal sensing data of the real-time operating status of the natural gas pipeline network, and generate an initial data set; an adaptive synchronization unit, used to perform edge computing adaptive data synchronization based on the initial data set, dynamically adjust the synchronization frequency to match data heterogeneity, and generate a synchronized data set; a fusion optimization unit, used to perform variational autoencoder fusion optimization based on the synchronized data set, quantify uncertainty and refine features, and generate a fusion feature set; and an output generation unit, used to perform standardization processing based on the fusion feature set, and generate a fusion initial dataset.

[0008] Optionally, the model building module includes: a data input unit for the fused initial dataset, performing data preprocessing and feature mapping to generate a preprocessed dataset; a domain adaptation initialization unit for the preprocessed dataset, performing variational inference domain adaptation initialization, quantifying data uncertainty and adapting to changes in operating conditions, and generating an initial model parameter set; a parameter optimization unit for the initial model parameter set, performing multi-objective search to optimize parameters, constructing the multimodal digital twin model and the hierarchical reinforcement learning network, and generating an optimized network structure; and a framework output unit for the optimized network structure, performing integration verification and framework assembly, and generating an initial symbiotic decision framework.

[0009] Optionally, the state synchronization exploration module includes: a state synchronization unit, used to perform adaptive state alignment optimization based on the initial symbiotic decision framework, dynamically calibrate multi-source operating state matching risk scenarios, and generate a synchronized state group; and a risk exploration unit, used in the synchronized state group to perform probabilistic risk path mining, simulate evolution trajectory and quantify boundary uncertainty, and generate a predictive state space.

[0010] Optionally, the system further includes: performing heterogeneous data state coupling analysis based on the synchronized data group and the synchronized state group, fusing multimodal data characteristics with operational state correlations to generate a preliminary calibration parameter group; performing dynamic calibration optimization based on the preliminary calibration parameter group, iteratively adjusting data state matching weights and verifying calibration stability to generate an optimized calibration parameter group; and performing comprehensive index integration based on the optimized calibration parameter group, quantifying the impact of multi-source consistency, and generating a comprehensive dynamic calibration index.

[0011] Optionally, the strategy evolution module includes: a state input unit, used to perform initial state mapping and calibration injection on the predictive state space and the comprehensive dynamic calibration index to generate a calibration state space; a cross-layer distillation unit, used in the calibration state space to perform cross-layer knowledge distillation, transferring the macroscopic safety objective to hydrogen doping ratio actions, and generating a distillation strategy prototype; a feedback coupling unit, used to perform dynamic feedback coupling between the distillation strategy prototype and the optimized network structure, optimizing the evolution path through virtual trial and error injection into the network structure, and generating a coupled evolution strategy group; and an output refining unit, used in the coupled evolution strategy group to perform strategy verification and fusion, and generate an evolution control strategy group.

[0012] Optionally, the system further includes: performing multi-dimensional coupling analysis of data strategies based on the synchronized data group and the coupled evolution strategy group, integrating data synchronization characteristics with strategy evolution correlation, and generating a preliminary coupling parameter group; performing adaptive coupling optimization based on the preliminary coupling parameter group, iteratively adjusting matching weights and verifying stability, and generating an optimized coupling parameter group; and performing comprehensive index integration based on the optimized coupling parameter group, quantifying the impact of synchronization strategy consistency, and feeding back to the strategy evolution module.

[0013] Optionally, the self-healing optimization module includes: a residual calculation unit for the evolution control strategy group, performing multi-source residual pre-aggregation calculation, fusing multi-modal gap signals and filtering noise interference to generate an initial residual signal group; a scale decomposition unit for the initial residual signal group and the coupled evolution strategy group, performing multi-scale residual decomposition hierarchical structure to generate a decomposed residual group; and a reconstruction output unit for the decomposed residual group and the initial model parameter group, performing adaptive network reconstruction, inversely adjusting the multi-modal digital twin model and the hierarchical reinforcement learning network and performing verification fusion to generate a self-healing optimized model.

[0014] Optionally, the risk control output module includes: a risk quantification unit, used in the self-healing optimization model to perform dynamic risk vector mapping, quantify multi-dimensional risk correlations and calculate probability boundaries, and generate a risk quantification index group; and a ratio control unit, used in the risk quantification index group to perform adaptive ratio control, dynamically adjust the hydrogen doping ratio and output a safety assessment report.

[0015] Based on the same inventive concept, this invention also provides a method for dynamic control and safety assessment of the hydrogen blending ratio in natural gas. The method includes: collecting constraint data of the physical characteristics of a natural gas pipeline network and multimodal sensing data of the real-time operating status of the natural gas pipeline network; fusing the constraint data and the multimodal sensing data to generate a fused initial dataset; performing domain adaptation initialization and parameter optimization based on the fused initial dataset to construct a multimodal digital twin model and a hierarchical reinforcement learning network, generating an initial symbiotic decision framework; performing state synchronization and risk path exploration based on the initial symbiotic decision framework to generate a predictive state space; performing hierarchical reinforcement learning self-evolutionary training based on the predictive state space, evolving the optimal strategy through a virtual trial-and-error mechanism to generate an evolutionary control strategy group; calculating the real-world gap residual as a meta-level feedback signal to back-adjust the multimodal digital twin model and the hierarchical reinforcement learning network, generating a self-healing optimization model; and performing multi-dimensional risk quantification and dynamic ratio control based on the self-healing optimization model, outputting a safety assessment report and performing adaptive control of the hydrogen blending ratio.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. This invention achieves dynamic optimization control of the hydrogen blending ratio in natural gas through a self-evolving symbiotic architecture of multimodal digital twins and hierarchical reinforcement learning. Compared to traditional fixed-parameter models, this system can dynamically adapt to complex operating conditions based on real-time multimodal data fusion and virtual trial-and-error mechanisms, significantly improving the flexibility and stability of the control process, ensuring the maintenance of the optimal hydrogen blending ratio under conditions such as pressure fluctuations, and optimizing the response efficiency of computer-aided control.

[0018] 2. This invention employs a multi-dimensional risk quantification and dynamic feedback coupling mechanism, enhancing the computational accuracy and real-time performance of safety assessments. Compared to existing single-risk assessment methods, this system comprehensively analyzes multi-source risk correlations and dynamically adjusts proportional control strategies through data strategy coupling and self-healing optimization calculations, significantly improving the safety management capabilities of pipeline systems and avoiding the neglect of latent risks by traditional methods.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a framework diagram of a natural gas hydrogen blending ratio dynamic control and safety assessment system according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of a dynamic control and safety assessment system for the hydrogen blending ratio of natural gas according to an embodiment of the present invention.

[0023] Figure 3 This is a latent space distribution map of multimodal data fusion according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the Pareto front for multi-objective search optimization according to an embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram of the dynamic hydrogen doping ratio control process under the safety envelope of an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 One embodiment of the present invention proposes a dynamic control and safety assessment system and method for the hydrogen blending ratio of natural gas. It adopts a self-evolving symbiotic architecture of multimodal digital twin and hierarchical reinforcement learning, which can realize dynamic optimization of the hydrogen blending ratio and real-time risk assessment, significantly improving the control stability and safety reliability of natural gas pipeline networks under complex operating conditions.

[0028] like Figure 2As shown, the system in this embodiment specifically includes:

[0029] The data acquisition and fusion module is used to acquire constraint data of the physical characteristics of the natural gas pipeline network and multimodal sensing data of the real-time operating status of the natural gas pipeline network, and to fuse the constraint data and the multimodal sensing data to generate an initial fusion dataset.

[0030] The model building module is used to perform domain adaptation initialization and parameter optimization based on the fused initial dataset, build a multimodal digital twin model and a hierarchical reinforcement learning network, and generate an initial symbiotic decision framework;

[0031] The state synchronization exploration module is used to perform state synchronization and risk path exploration based on the initial symbiotic decision framework to generate a predictive state space;

[0032] The strategy evolution module is used to perform hierarchical reinforcement learning self-evolution training based on the predictive state space, evolve the optimal strategy through a virtual trial and error mechanism, and generate an evolutionary control strategy group.

[0033] The self-healing optimization module is used to calculate the reality gap residual as a meta-level feedback signal based on the evolutionary control strategy group, and to adjust the multimodal digital twin model and the hierarchical reinforcement learning network in reverse to generate a self-healing optimization model.

[0034] The risk control output module is used to perform multi-dimensional risk quantification and dynamic ratio control based on the self-healing optimization model, output a safety assessment report, and perform adaptive control of the hydrogen doping ratio.

[0035] This invention employs a self-evolving symbiotic architecture of multimodal digital twins and hierarchical reinforcement learning, which enables dynamic optimization of hydrogen doping ratio and real-time risk assessment, significantly improving the control stability and safety reliability of natural gas pipeline networks under complex operating conditions.

[0036] Optionally, the data acquisition and fusion module includes:

[0037] The data acquisition unit is used to acquire the constraint data of the physical characteristics of the natural gas pipeline network and the multimodal sensing data of the real-time operating status of the natural gas pipeline network, and generate an initial data set.

[0038] Specifically, this step aims to provide all the raw information needed for decision-making and analysis. Constraint data consists of engineering data describing the static physical properties of the pipeline network, such as pipe topology, diameter, wall thickness, and rated yield strength determined by the material grade. Multimodal sensing data consists of real-time measurements describing the dynamic operating state of the pipeline network, such as time-series data collected by pressure sensors, temperature sensors, turbine flow meters, electrochemical hydrogen concentration sensors, and acoustic emission (AE) sensors used to monitor the development of microcracks deployed along the pipeline. This data acquisition unit combines these two types of data to form the initial data set.

[0039] An adaptive synchronization unit is used to perform edge computing adaptive data synchronization based on the initial data group, dynamically adjust the synchronization frequency to match data heterogeneity, and generate a synchronized data group.

[0040] Specifically, this step aims to address the time alignment challenges caused by the diverse sources, varying sampling frequencies, and differing importance of multimodal sensing data. This unit, deployed in an edge computing gateway close to the sensor, implements an adaptive data synchronization strategy. This strategy dynamically adjusts the synchronization and upload frequency of data packets based on the information change rate of each data stream. For example, a mechanism is used to determine the synchronization frequency. The calculation can be defined by the following formula:

[0041] ,

[0042] in, It is a basic synchronization frequency. It is the first The normalized signal value of a sensor data stream, It is its signal change rate. This is the preset importance weight of the data stream. It is a sensitivity adjustment coefficient. When the pipeline network is running smoothly, the rate of change of each signal is low, and the synchronization frequency is maintained at a basic level. When transient events such as pressure fluctuations occur, the rate of change surges, and the synchronization frequency is dynamically increased accordingly. This greatly optimizes the efficiency of data transmission and processing while ensuring that key information is not lost, and ultimately generates a synchronized data set that is precisely aligned in time.

[0043] The fusion optimization unit is used to perform variational autoencoder fusion optimization based on the synchronous data group, quantify uncertainty and refine features, and generate a fusion feature group;

[0044] Specifically, this step aims to fuse and compress heterogeneous, multimodal synchronization data sets into a unified, low-dimensional feature space. In this process, a variational autoencoder (VAE) is employed for fusion optimization. The encoder part of this network receives the synchronization data sets as input and maps them to a probability distribution in the latent space. The mean vector of this probability distribution is extracted as a highly refined and information-condensed fusion feature set that can characterize the current network state; while the variance vector of this probability distribution is used as a direct quantification of the "uncertainty" of the current data fusion result, which can be used by downstream modules, such as... Figure 3 As shown in the figure, the distribution of the two-dimensional latent space learned by the variational autoencoder is displayed in the form of a scatter plot. The markers of different shapes represent the fusion feature groups under different pipeline network operating conditions. It can be seen that the fusion optimization unit has successfully separated the different states. The ellipse around each point intuitively represents the uncertainty quantification result of the data point.

[0045] The output generation unit is used to perform standardization processing based on the fused feature group to generate a fused initial dataset.

[0046] Specifically, this step aims to provide a standardized input with uniform dimensions for subsequent model building modules. During this process, for each feature dimension in the fused feature set, its mean and standard deviation are calculated over a certain historical time window. Then, each feature value is subtracted from its corresponding mean and divided by its standard deviation to complete Z-score standardization. After this step, a fused initial dataset is finally generated, with all features in each dimension following a standard normal distribution, serving as the final output.

[0047] Optionally, the model building module includes:

[0048] A data input unit is used to perform data preprocessing and feature mapping on the fused initial dataset to generate a preprocessed dataset.

[0049] Specifically, this unit serves as the interface for the model building module, responsible for receiving the initial fused dataset generated by the data acquisition and fusion module. Within this unit, final data cleaning and transformation operations are performed on the dataset, such as detecting and removing potential outliers using algorithms like Isolation Forest. Simultaneously, to improve the training efficiency of subsequent neural networks, this unit can also perform a feature mapping, projecting the input data into a feature space that is easier for the model to learn, ultimately generating a preprocessed dataset.

[0050] The domain adaptation initialization unit is used for the preprocessed dataset to perform variational inference domain adaptation initialization, quantify data uncertainty and adapt to changes in operating conditions, and generate an initialization model parameter set.

[0051] Specifically, this unit aims to provide a high-quality, information-rich "starting point" for subsequent parameter optimization, rather than traditional random initialization. In this process, a variational inference domain adaptive initialization method is employed. This method does not directly solve for a fixed set of initial model parameters, but instead trains an inference network to learn a posterior probability distribution that characterizes the model parameters. By training on a preprocessed dataset containing various historical operating conditions, this posterior probability distribution inherently possesses adaptability to different operating conditions. Simultaneously, the variance of this distribution directly quantifies the parameter uncertainty caused by data noise or incompleteness. Ultimately, the initial model parameter set output by this unit is this set of hyperparameters describing the posterior probability distribution of the parameters.

[0052] The parameter optimization unit is used to initialize the model parameter group, perform multi-objective search to optimize the parameters, construct the multimodal digital twin model and the hierarchical reinforcement learning network, and generate an optimized network structure.

[0053] Specifically, this unit aims to find a specific parameter instance with optimal comprehensive performance across multiple dimensions from the parameter probability distribution obtained in the previous step. In this process, a multi-objective search algorithm is executed, such as an evolutionary search strategy. This algorithm samples from the probability distribution defined by the initial model parameter set, generating a large number of candidate network structures and parameter configurations. Each candidate configuration is evaluated in a virtual environment based on multiple conflicting performance objectives, such as: simulation accuracy of the multimodal digital twin model, initial learning efficiency of the hierarchical reinforcement learning network, and overall model computational complexity. Through multiple generations of iterative optimization, this unit ultimately finds a network configuration located on the Pareto front that achieves the optimal balance among the aforementioned objectives, serving as the optimized network structure, such as... Figure 4 As shown in the figure, the performance distribution of a large number of candidate network structures in the multi-objective search optimization process is displayed in the form of a scatter plot. The horizontal axis represents the computational complexity of the model, and the vertical axis represents the simulation accuracy of the model. The figure clearly shows the Pareto front curve that achieves the optimal trade-off between the two conflicting objectives, as well as the optimal network structure solution that is finally selected to build the symbiotic decision framework.

[0054] The framework output unit is used to optimize the network structure, perform integration verification and framework assembly, and generate an initial symbiotic decision framework.

[0055] Specifically, this unit is the final stage of model building, aiming to assemble and validate two independently optimized network models into a collaborative, organic whole. During this process, the multimodal digital twin model and the hierarchical reinforcement learning network defined in the optimized network structure are instantiated. This unit performs an integration validation, placing both models in a simulated environment for joint debugging to ensure the correctness and stability of their interface calls, data transmission, and interaction logic. After successful validation, this unit performs the final framework assembly, encapsulating them into a unified initial symbiotic decision framework that can be called by subsequent modules as the final output.

[0056] Optionally, the state synchronization exploration module includes:

[0057] The state synchronization unit is used to perform adaptive state alignment optimization based on the initial symbiotic decision framework, dynamically calibrate the multi-source operating state matching risk scenario, and generate a synchronized state group.

[0058] Specifically, this unit aims to ensure that the internal state of the multimodal digital twin model accurately and in real-time reflects the true health status of the physical pipeline network. In this process, the unit continuously inputs real-time acquired multimodal sensor data into the multimodal digital twin model within the initial symbiotic decision framework, and compares the deviation between the model's simulation output and the actual sensor data to form a state alignment error. Using an online optimization algorithm, such as particle swarm optimization, the internal state parameters in the multimodal digital twin model that cannot be directly measured, such as the virtual fatigue damage degree or hydrogen embrittlement sensitivity coefficient of the pipeline, are iteratively adjusted with the goal of minimizing this state alignment error. After dynamic calibration, this set of internal parameters that represents the current true state of the physical pipeline network constitutes the synchronization state set.

[0059] The risk exploration unit is used in the synchronized state group to perform probabilistic risk path mining, simulate evolution trajectory and quantify boundary uncertainty, and generate a predictive state space.

[0060] Specifically, this unit aims to proactively and probabilistically explore potential future failure paths using a multimodal digital twin model synchronized with reality. In this process, the unit executes thousands of Monte Carlo accelerated simulations using a synchronized state set as initial conditions. In each simulation, random perturbations conforming to the uncertainty distribution of key parameters in the model are applied, thereby simulating a unique future evolution trajectory. To accurately quantify potential extreme risks, this unit further employs the Conditional Value at Risk (CVaR) method to assess boundary uncertainties, the calculation formula of which is:

[0061] ,

[0062] in, It is a loss function that represents the approximation of a failure state. At a confidence level of Value at risk at that time This is a preset high confidence level. By calculating the conditional value at risk, the unit can accurately assess the worst-case scenario. In this context, the expected value of the loss is calculated. Ultimately, by combining all the simulated evolutionary trajectories with the calculated boundary uncertainty index, a predictive state space is constructed as the final output, which includes both future trends and the quantification of extreme risks.

[0063] Optionally, the system further includes:

[0064] Based on the synchronization data group and the synchronization state group, heterogeneous data state coupling analysis is performed, multimodal data characteristics are integrated with the operating state, and a preliminary calibration parameter group is generated.

[0065] Specifically, this step aims to construct a parallel verification path for evaluating "self-consistency," that is, verifying the consistency between the lower-level sensor data and the higher-level digital twin state. During this process, a heterogeneous data state coupling analysis is performed. This analysis quantifies the strength of the correlation between the lower-level statistical characteristics in the synchronized data set and the higher-level states inferred from the digital twin model in the synchronized state set by calculating the mutual information or correlation coefficient between them. All these quantified correlation strength indicators collectively constitute the initial calibration parameter set.

[0066] Based on the initial calibration parameter set, dynamic calibration optimization is performed, the data state matching weights are iteratively adjusted and the calibration stability is verified to generate an optimized calibration parameter set.

[0067] Specifically, this step aims to refine the initial coupling correlation indicators to eliminate spurious correlations caused by random noise or transient disturbances. During this process, a sliding time window-based verification mechanism is used to check the stability of each parameter in the initial calibration parameter set. Only parameters exhibiting stable and consistent strong correlations across multiple consecutive time windows have their matching weights retained or enhanced; while weak correlations that fluctuate wildly are suppressed or eliminated. Through this iterative verification and adjustment, an optimized calibration parameter set that more robustly reflects the true correlation between data and state is ultimately generated.

[0068] Based on the optimized calibration parameter set, a comprehensive index integration is performed to quantify the impact of multi-source consistency and generate a comprehensive dynamic calibration index.

[0069] Specifically, this step aims to integrate the optimized, multi-dimensional calibration parameters obtained in the previous step into a single, comprehensive indicator that intuitively reflects the current overall consistency. In this process, a pre-trained lightweight neural network or a weighted summation model can be used to fuse all parameters in the optimized calibration parameter set into a standardized single value. This value is the comprehensive dynamic calibration indicator. A higher indicator value indicates a high degree of consistency between the underlying sensor data and the high-level model state, suggesting reliable cognition; conversely, a lower indicator value indicates a discrepancy between the two, possibly stemming from sensor anomalies or model mismatch, requiring more careful decision-making from downstream modules.

[0070] Optionally, the strategy evolution module includes:

[0071] The state input unit is used to perform initial state mapping and calibration injection on the predictive state space and the integrated dynamic calibration index to generate a calibration state space.

[0072] Specifically, this unit aims to provide a rich and confidence-calibrated input to the hierarchical reinforcement learning network. In this process, "initial state mapping" refers to mapping the complex, high-dimensional probabilistic data in the predictive state space into a fixed-length feature vector representing the current environmental state through an encoder network. "Calibration injection" is a crucial and innovative step, using a comprehensive dynamic calibration index as an additional input channel, concatenating or gating it with the state feature vector. In this way, "self-consistency" or "confidence level" is directly injected into the state representation, enabling the subsequent reinforcement learning agent to perceive the quality of information upon which the current decision is based, thereby generating a calibrated state space.

[0073] A cross-layer distillation unit is used in the calibration state space to perform cross-layer knowledge distillation, transferring macroscopic safety objectives to hydrogen doping ratio actions and generating a distillation strategy prototype.

[0074] Specifically, this unit is the core training stage of the hierarchical reinforcement learning network, aiming to efficiently transform abstract macroscopic goals into concrete microscopic actions. In this process, a cross-layer knowledge distillation technique is employed. This technique comprises a "teacher network" responsible for evaluating the macroscopic security goal and a "student network" responsible for outputting specific hydrogen doping ratios. The training of the "student network" not only relies on traditional reinforcement learning rewards but also learns to fit the soft-label probability distribution output by the "teacher network" through a knowledge distillation loss function. The total loss function of the student network is... It can be represented as:

[0075] ,

[0076] in, It is a standard reinforcement learning loss function. It is used to measure the distribution of teachers' network outputs. Student network output distribution Knowledge distillation loss due to differences between them It is a hyperparameter that balances the two losses. By minimizing this total loss function, macroscopic safety knowledge is efficiently "distilled" and transferred to specific action policies, generating distillation policy prototypes.

[0077] The feedback coupling unit is used to dynamically couple the distillation strategy prototype and the optimized network structure, injecting virtual trial and error into the network structure optimization evolution path to generate a coupled evolution strategy group.

[0078] Specifically, this unit aims to deeply optimize policies through virtual trial and error, and creatively introduces the evolution of the network structure itself. In this process, the distilled policy prototype serves as the initial policy, undergoing extensive trial-and-error exploration in the virtual environment provided by the multimodal digital twin model. Furthermore, "dynamic feedback coupling" is reflected not only in optimizing the weight parameters of the policy network, but also in fine-tuning the optimized network structure itself based on specific challenging scenarios encountered in virtual trial and error. For example, through Neural Architecture Search (NAS) technology, neuronal connections are dynamically added or pruned to optimize its ability to model specific evolutionary paths, thereby generating coupled evolutionary policy sets.

[0079] The output refining unit is used for the coupled evolutionary strategy group to perform strategy verification and fusion, and generate an evolutionary control strategy group.

[0080] Specifically, this unit is the final stage of policy evolution, designed to ensure the robustness and stability of the output policy. In this process, "policy verification" refers to testing the optimal policy in the coupled evolutionary policy group under a series of pre-defined, stringent virtual boundary conditions to verify its safety and generalization ability. "Fusion" is an optional enhancement step that integrates multiple high-performing policies that emerge during evolution, for example, through weighted averaging or voting mechanisms, to form a final policy with stronger overall performance and greater stability. This verified and fused final policy is the evolutionary control policy group.

[0081] Optionally, the system further includes:

[0082] Based on the synchronized data group and the coupled evolution strategy group, perform multidimensional coupling analysis of data strategy, integrate data synchronization characteristics and strategy evolution correlation, and generate a preliminary coupling parameter group;

[0083] Specifically, this step aims to build an analysis module to assess the interaction between "data quality" and "strategy behavior." During this process, a multidimensional coupling analysis of data and strategy is performed. This analysis aims to answer a core question: will the performance of the control strategy learned by the strategy evolution module fluctuate drastically due to small changes in the characteristics of the input data? This analysis quantifies this potential coupling relationship by calculating a multidimensional correlation matrix between the statistical characteristics of the synchronized data set and the performance indicators exhibited by the coupled evolution strategy set in the virtual environment. This correlation matrix constitutes the initial coupling parameter set.

[0084] Based on the initial coupling parameter set, adaptive coupling optimization is performed, the matching weights are iteratively adjusted and the stability is verified to generate an optimized coupling parameter set;

[0085] Specifically, this step aims to refine the initially quantified coupling relationships to ensure their stability and reliability. In this process, a time-series cross-validation method is employed, repeatedly executing multi-dimensional coupling analysis of the data strategy across different historical data time windows. Only those coupling relationships that stably reproduce across multiple time windows and possess statistical significance are confirmed and strengthened in terms of matching weight. Through this iterative validation and optimization, accidental spurious couplings caused by specific data fragments can be filtered out, thereby generating an optimized set of coupling parameters that more stably and fundamentally reflects the correlation between data characteristics and strategy performance.

[0086] Based on the optimized coupling parameter group, a comprehensive index integration is performed to quantify the impact of synchronization strategy consistency, and the results are fed back to the strategy evolution module.

[0087] Specifically, this step is the closing loop of another higher-dimensional feedback loop. In this process, the optimized coupling parameter set is integrated into one or more feedback adjustment signals and fed back to the policy evolution module. These feedback signals are used to dynamically and adaptively adjust the learning process of the policy evolution module itself. For example, when the optimized coupling parameter set shows that the current control policy is highly sensitive to noise in the input data, the feedback signal will instruct the cross-layer distillation unit in the policy evolution module to increase the weight of the regularization term used to suppress overfitting in its loss function, or instruct the feedback coupling unit to increase the proportion of noisy training samples in virtual trial and error. In this way, direct feedback from data quality to the policy learning process is achieved, enabling the policy evolution to proactively adapt to the current data environment.

[0088] Optionally, the self-healing optimization module includes:

[0089] The residual calculation unit is used by the evolution control strategy group to perform multi-source residual pre-aggregation calculation, fuse multi-modal difference signals and filter noise interference to generate an initial residual signal group.

[0090] Specifically, this unit aims to accurately quantify the discrepancy between the "virtual world" and the "physical world" within the symbiotic decision-making framework, i.e., the "reality gap." In this process, the unit first uses an evolutionary control strategy group to drive a multimodal digital twin model for simulation, obtaining a set of simulation outputs corresponding to the multimodal sensing data. Subsequently, this simulation output is compared one by one with synchronized, real multimodal sensing data, calculating the differences between the two to form a multidimensional residual vector. This unit further filters this residual vector to remove interference from random measurement noise, ultimately generating an initial residual signal set that accurately reflects the model's bias.

[0091] The scale decomposition unit is used to perform multi-scale residual decomposition hierarchical structure on the initial residual signal group and the coupled evolution strategy group to generate decomposed residual groups.

[0092] Specifically, this unit aims to perform in-depth analysis of the initial residual signal set to determine whether its root cause stems from short-term strategic bias or long-term model drift. In this process, a multi-scale decomposition technique, such as the Continuous Wavelet Transform (CWT), is employed to decompose the initial residual signal set. The calculation of the Continuous Wavelet Transform can be defined by the following formula:

[0093] ,

[0094] in, It is a time series of the initial residual signal group. It is the mother wavelet function. It is a scale factor. It is the translation factor. Indicates conjugation. This is achieved through different scale factors. By performing a transformation, this unit can decompose the total residuals across different time scales. For example, residuals associated with high-frequency, instantaneous behavior will be concentrated at smaller scales, while residuals associated with slow, drifting behavior will be concentrated at larger scales. All these residual components at different scales together constitute the decomposed residual set.

[0095] The reconstructed output unit is used to perform adaptive network reconstruction on the decomposed residual group and the initialized model parameter group, and to reverse adjust the multimodal digital twin model and the hierarchical reinforcement learning network and perform verification fusion to generate a self-healing optimized model.

[0096] Specifically, this unit is the core of the "self-healing" operation, designed to intelligently and selectively feed back the residuals at different scales obtained from the previous step to the corresponding models. During this process, this unit performs an adaptive network reconstruction. The basic principle is that the residual components concentrated at smaller time scales, obtained from the scale decomposition unit, are mainly attributed to short-term policy biases in the hierarchical reinforcement learning network and are therefore used to generate correction signals to adjust the hierarchical reinforcement learning network. Meanwhile, the residual components concentrated at larger time scales are mainly attributed to physical mechanism mismatches or parameter drifts in the multimodal digital twin model and are therefore used to generate correction signals to adjust the multimodal digital twin model. Through this "divide and conquer" feedback mechanism, accurate and efficient collaborative optimization of the two core models is achieved. The resulting set of models, doubly calibrated, is the self-healing optimized model.

[0097] Optionally, the risk control output module includes:

[0098] The risk quantification unit is used in the self-healing optimization model to perform dynamic risk vector mapping, quantify multi-dimensional risk correlations and calculate probability boundaries, and generate a risk quantification index group.

[0099] Specifically, this unit is the final diagnosis and quantification stage, aiming to transform the complex internal state of the self-healing optimization model into a multi-dimensional risk insight that provides clear guidance for operations and maintenance personnel. In this process, this unit performs a dynamic risk vector mapping. This mapping process includes: first, directly extracting the final failure probability from the self-healing optimization model as the core "probabilistic risk" indicator; second, using a feature attribution algorithm to analyze and calculate the contribution of each input feature to the failure probability, forming a "causal risk" indicator; and third, analyzing the variance of the probability distribution output by the self-healing optimization model to form an "uncertainty risk" indicator. This unit integrates these three dimensions of risk indicators to generate a set of quantitative risk indicators.

[0100] The proportional control unit is used for the risk quantification index group to perform adaptive proportional control, dynamically adjust the hydrogen doping ratio, and output a safety assessment report.

[0101] Specifically, this unit is the final decision-making and output stage, responsible for generating two key outputs: one for people and one for equipment. On one hand, based on a risk quantification index set, this unit generates a visualized, multi-level safety assessment report. On the other hand, this unit executes adaptive ratio control. In this process, it first receives the optimal hydrogen doping ratio recommendation value output by a hierarchical reinforcement learning network, and then uses a "safety safeguard strategy" to verify and smooth this recommendation value using the risk quantification index set. For example, when the "uncertainty risk" index is high, the safeguard strategy will appropriately lower the recommendation value to increase the safety margin, ultimately outputting a smart and safe dynamic hydrogen doping ratio control command, such as... Figure 5 As shown in the figure, the dynamic process of adaptive ratio control is illustrated. The theoretically optimal hydrogen doping ratio generated by the hierarchical reinforcement learning network is smoothed and constrained by a safety control envelope that dynamically tightens based on uncertainty when an increase in risk is predicted. The final output is a practically controlled hydrogen doping ratio that is both efficient and absolutely safe.

[0102] Based on the same inventive concept, this invention also provides a method for dynamic control and safety assessment of the hydrogen blending ratio in natural gas, the method comprising:

[0103] Constraint data of the physical characteristics of the natural gas pipeline network and multimodal sensing data of the real-time operating status of the natural gas pipeline network are collected, and the constraint data and the multimodal sensing data are fused to generate an initial fused dataset.

[0104] Based on the aforementioned fused initial dataset, domain adaptation initialization and parameter optimization are performed to construct a multimodal digital twin model and a hierarchical reinforcement learning network, thereby generating an initial symbiotic decision framework.

[0105] Based on the initial symbiotic decision-making framework, state synchronization and risk path exploration are performed to generate a predictive state space;

[0106] Based on the predictive state space, hierarchical reinforcement learning self-evolutionary training is performed, and the optimal strategy is evolved through a virtual trial-and-error mechanism to generate an evolutionary control strategy group;

[0107] Based on the evolutionary control strategy group, the reality gap residual is calculated as a meta-level feedback signal, and the multimodal digital twin model and the hierarchical reinforcement learning network are adjusted in reverse to generate a self-healing optimization model.

[0108] Based on the self-healing optimization model, multi-dimensional risk quantification and dynamic ratio control are performed, a safety assessment report is output, and adaptive control of the hydrogen doping ratio is carried out.

[0109] To verify the feasibility and advancement of this invention in practice, it was applied to a regional hydrogen blending demonstration project for a city-level natural gas pipeline network. This project aims to gradually increase the proportion of clean energy in the natural gas supply to a large industrial park and surrounding residential areas by injecting green hydrogen produced through photovoltaic electrolysis. The regional pipeline network has varying pipe materials (new and old) and a complex topology, presenting the core challenge of maximizing the hydrogen blending ratio while ensuring absolute safety, thus achieving a balance between economic and environmental benefits.

[0110] In this embodiment, dynamic control and safety assessment of the pipeline network in the demonstration area were conducted for six months. Data was collected from multimodal sensors deployed along the pipeline, including those for pressure, flow rate, temperature, hydrogen concentration, and acoustic emission, and was integrated with constraint data such as the pipeline network's GIS topology and material properties.

[0111] First, the heterogeneous data is processed by the data acquisition and fusion module. For example, during a transient fluctuation in upstream pressure on January 20, 2025, the adaptive synchronization unit deployed at the edge gateway automatically increased the synchronization frequency of the pressure and acoustic emission sensors from the usual 1Hz to 50Hz, generating a high-fidelity synchronization data set. Subsequently, the fusion optimization unit transforms this data set into an initial fusion dataset that includes uncertainty quantization through a variational autoencoder.

[0112] The model building module constructs a multimodal digital twin model and a hierarchical reinforcement learning network that includes the physical mechanism of the pipeline network through the domain adaptation initialization unit and the parameter optimization unit, thus forming an initial symbiotic decision framework.

[0113] During real-time operation, the state synchronization exploration module continuously operates. The state synchronization unit aligns real-time sensor data with the digital twin model, dynamically calibrating implicit states such as the virtual hydrogen embrittlement damage level inside the pipeline. Based on this calibrated state, the risk exploration unit performs probabilistic risk path mining and generates a predictive state space on February 5th. This space shows that if the hydrogen doping ratio is increased from 8% to 15%, the failure probability of a certain aging pipeline section will exceed the safety threshold within the next 72 hours.

[0114] Meanwhile, a parallel verification module generated a comprehensive dynamic calibration index that remained above 0.9 during this period, indicating that the underlying data was highly consistent with the high-level twin state and the cognition was reliable.

[0115] The strategy evolution module receives the aforementioned predictive state space and high-reliability calibration metrics. The interlayer distillation unit efficiently transforms the macroscopic objective of "maintaining pipe safety margin" into a specific hydrogen doping ratio adjustment strategy. The feedback coupling unit then conducts trial and error in the virtual environment provided by the digital twin, optimizing the agent's network structure and ultimately generating the evolutionary control strategy set.

[0116] The self-healing optimization module demonstrated its powerful adaptive capabilities. In early April, due to slight drifts in the physical properties of the pipeline caused by seasonal temperature changes, the calculated "residual gap" began to slowly increase. The scaling unit attributed this residual to long-term model drift, and the reconstruction output unit accordingly fine-tuned the material thermal expansion coefficient of the multimodal digital twin model, restoring the model to high accuracy, thus generating a self-healing optimized model.

[0117] Ultimately, the risk control output module produced a highly intelligent result. The generated safety assessment report not only quantified the risks of each pipeline segment but also, through dynamic risk vector mapping, indicated that the dominant risk factor was the localized enrichment of hydrogen concentration downstream of valve 2. Based on this report and through a "safety protection strategy," the proportional control unit adaptively and smoothly reduced the hydrogen blending ratio from 12% to 10.5% during peak gas consumption periods, ensuring a safety margin.

[0118] Data shows that this invention can achieve performance far exceeding that of traditional fixed-ratio or passive-response control. Through multi-level intelligent optimization and feedback, the average hydrogen doping ratio was increased from a conservative fixed value of 5% to a dynamically changing 11.8% over a 6-month operation period, without ever triggering any safety alarms.

[0119] Table 1. Examples of Dynamic Hydrogen Filtration Ratio and Pipeline Status Monitoring Data

[0120]

[0121] Table 2 Comparison of Safety Assessment and Dynamic Control Performance

[0122]

[0123] Table 3 Examples of Risk Event Prediction and Verification

[0124]

[0125] As can be seen from the data recorded in Tables 1-3 above, this invention performs excellently in practical applications. Table 1 demonstrates the ability to dynamically adjust the hydrogen blending ratio based on operating conditions and data quality. Table 2, through a comparison with traditional methods, clearly proves that this invention can significantly improve the benefits of hydrogen blending while better ensuring safety. Table 3 verifies the core predictive safety assessment capability; its projections of potential risks are strongly confirmed by virtual simulation, providing strong technical support for achieving proactive and preventative pipeline safety management.

[0126] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or parameters that can be superimposed in the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. These are conventional technical methods and will not be elaborated further. The electrical connections between the various units described above do not necessarily represent direct or indirect connections; any indirect connection method is applicable to the embodiments of this invention as long as it achieves the purpose of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of this invention.

[0127] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A dynamic control and safety assessment system for the hydrogen blending ratio of natural gas, characterized in that, The system includes: The data acquisition and fusion module is used to acquire constraint data of the physical characteristics of the natural gas pipeline network and multimodal sensing data of the real-time operating status of the natural gas pipeline network, and to fuse the constraint data and the multimodal sensing data to generate an initial fusion dataset. The model building module is used to perform domain adaptation initialization and parameter optimization based on the fused initial dataset, build a multimodal digital twin model and a hierarchical reinforcement learning network, and generate an initial symbiotic decision framework; The state synchronization exploration module is used to perform state synchronization and risk path exploration based on the initial symbiotic decision framework to generate a predictive state space; The strategy evolution module is used to perform hierarchical reinforcement learning self-evolution training based on the predictive state space, evolve the optimal strategy through a virtual trial and error mechanism, and generate an evolutionary control strategy group. The self-healing optimization module is used to calculate the reality gap residual as a meta-level feedback signal based on the evolutionary control strategy group, and to adjust the multimodal digital twin model and the hierarchical reinforcement learning network in reverse to generate a self-healing optimization model. The risk control output module is used to perform multi-dimensional risk quantification and dynamic ratio control based on the self-healing optimization model, output a safety assessment report, and perform adaptive control of the hydrogen doping ratio.

2. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 1, characterized in that, The data acquisition and fusion module includes: The data acquisition unit is used to acquire the constraint data of the physical characteristics of the natural gas pipeline network and the multimodal sensing data of the real-time operating status of the natural gas pipeline network, and generate an initial data set. An adaptive synchronization unit is used to perform edge computing adaptive data synchronization based on the initial data group, dynamically adjust the synchronization frequency to match data heterogeneity, and generate a synchronized data group. The fusion optimization unit is used to perform variational autoencoder fusion optimization based on the synchronous data group, quantify uncertainty and refine features, and generate a fusion feature group; The output generation unit is used to perform standardization processing based on the fused feature group to generate a fused initial dataset.

3. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 2, characterized in that, The model building module includes: A data input unit is used to perform data preprocessing and feature mapping on the fused initial dataset to generate a preprocessed dataset. The domain adaptation initialization unit is used for the preprocessed dataset to perform variational inference domain adaptation initialization, quantify data uncertainty and adapt to changes in operating conditions, and generate an initialization model parameter set. The parameter optimization unit is used to initialize the model parameter group, perform multi-objective search to optimize the parameters, construct the multimodal digital twin model and the hierarchical reinforcement learning network, and generate an optimized network structure. The framework output unit is used to optimize the network structure, perform integration verification and framework assembly, and generate an initial symbiotic decision framework.

4. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 3, characterized in that, The state synchronization exploration module includes: The state synchronization unit is used to perform adaptive state alignment optimization based on the initial symbiotic decision framework, dynamically calibrate the multi-source operating state matching risk scenario, and generate a synchronized state group. The risk exploration unit is used in the synchronized state group to perform probabilistic risk path mining, simulate evolution trajectory and quantify boundary uncertainty, and generate a predictive state space.

5. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 4, characterized in that, The system also includes: Based on the synchronization data group and the synchronization state group, perform heterogeneous data state coupling analysis, integrate multimodal data characteristics with operating state correlation, and generate a unit for a preliminary calibration parameter group; Based on the initial calibration parameter set, dynamic calibration optimization is performed, the data state matching weights are iteratively adjusted and the calibration stability is verified, and a unit with an optimized calibration parameter set is generated. Based on the optimized calibration parameter set, a comprehensive index integration is performed to quantify the impact of multi-source consistency and generate a unit of comprehensive dynamic calibration index.

6. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 5, characterized in that, The strategy evolution module includes: The state input unit is used to perform initial state mapping and calibration injection on the predictive state space and the integrated dynamic calibration index to generate a calibration state space. A cross-layer distillation unit is used in the calibration state space to perform cross-layer knowledge distillation, transferring macroscopic safety objectives to hydrogen doping ratio actions and generating a distillation strategy prototype. The feedback coupling unit is used to dynamically couple the distillation strategy prototype and the optimized network structure, injecting virtual trial and error into the network structure optimization evolution path to generate a coupled evolution strategy group. The output refining unit is used for the coupled evolutionary strategy group to perform strategy verification and fusion, and generate an evolutionary control strategy group.

7. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 6, characterized in that, The system also includes: Based on the synchronized data group and the coupled evolution strategy group, perform multidimensional coupling analysis of data strategy, integrate data synchronization characteristics and strategy evolution correlation, and generate units of preliminary coupling parameter group; Based on the initial coupling parameter set, adaptive coupling optimization is performed, the matching weights are iteratively adjusted and the stability is verified, and a unit with an optimized coupling parameter set is generated. Based on the optimized coupling parameter set, a comprehensive index integration is performed to quantify the impact of synchronization strategy consistency, and the results are fed back to the unit of the strategy evolution module.

8. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 6, characterized in that, The self-healing optimization module includes: The residual calculation unit is used by the evolution control strategy group to perform multi-source residual pre-aggregation calculation, fuse multi-modal difference signals and filter noise interference to generate an initial residual signal group. The scale decomposition unit is used to perform multi-scale residual decomposition hierarchical structure on the initial residual signal group and the coupled evolution strategy group to generate decomposed residual groups. The reconstructed output unit is used to perform adaptive network reconstruction on the decomposed residual group and the initialized model parameter group, and to reverse adjust the multimodal digital twin model and the hierarchical reinforcement learning network and perform verification fusion to generate a self-healing optimized model.

9. The natural gas hydrogen blending ratio dynamic control and safety assessment system according to claim 1, characterized in that, The risk control output module includes: The risk quantification unit is used in the self-healing optimization model to perform dynamic risk vector mapping, quantify multi-dimensional risk correlations and calculate probability boundaries, and generate a risk quantification index group. The proportional control unit is used for the risk quantification index group to perform adaptive proportional control, dynamically adjust the hydrogen doping ratio, and output a safety assessment report.

10. A method for dynamic control and safety assessment of the hydrogen blending ratio in natural gas, applied to a dynamic control and safety assessment system for the hydrogen blending ratio in natural gas as described in any one of claims 1-9, characterized in that, The method includes: Constraint data of the physical characteristics of the natural gas pipeline network and multimodal sensing data of the real-time operating status of the natural gas pipeline network are collected, and the constraint data and the multimodal sensing data are fused to generate an initial fused dataset. Based on the aforementioned fused initial dataset, domain adaptation initialization and parameter optimization are performed to construct a multimodal digital twin model and a hierarchical reinforcement learning network, thereby generating an initial symbiotic decision framework. Based on the initial symbiotic decision-making framework, state synchronization and risk path exploration are performed to generate a predictive state space; Based on the predictive state space, hierarchical reinforcement learning self-evolutionary training is performed, and the optimal strategy is evolved through a virtual trial-and-error mechanism to generate an evolutionary control strategy group; Based on the evolutionary control strategy group, the reality gap residual is calculated as a meta-level feedback signal, and the multimodal digital twin model and the hierarchical reinforcement learning network are adjusted in reverse to generate a self-healing optimization model. Based on the self-healing optimization model, multi-dimensional risk quantification and dynamic ratio control are performed, a safety assessment report is output, and adaptive control of the hydrogen doping ratio is carried out.

Images