Underground salt cavern compressed air energy storage bank monitoring system based on digital twinning

Abstract

The invention discloses an underground salt cavern compressed air energy storage bank monitoring system based on digital twinning, and relates to the technical field of intelligent monitoring. The method comprises the following steps: determining a key uncertainty parameter set influencing the multi-physics field behavior of the energy storage library, defining the prior probability distribution of the key uncertainty parameter set, sampling from the prior probability distribution by utilizing a Monte Carlo method, simulating the dynamic response of the energy storage library under a preset injection-production scheme by utilizing a multi-physics field coupling forward modeling model, and constructing an initial probabilistic digital twinborn body; collecting multivariate monitoring data of the energy storage library as observation evidence, carrying out weight updating and resampling on a model set in the initial probabilistic digital twinborn body by adopting a sequential Bayesian updating algorithm to obtain a posterior model set reflecting the current cognitive state, and carrying out parallel deduction on the state of the energy storage library in a specified time period in the future, a prediction result with a probability confidence interval is generated, and then the safe operation boundary and the injection-production strategy of the energy storage library are dynamically adjusted.

Description

Digital Twin-Based Monitoring System for Underground Salt Cavern Compressed Air Energy Storage Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and in particular to a monitoring system for underground salt cavern compressed air energy storage based on digital twins. Background Technology

[0002] Compressed gas storage (CGS) is one of the key large-scale energy storage technologies for solving peak shaving and valley filling in power grids and promoting the consumption of new energy sources. Utilizing caverns formed by the dissolution of underground salt rock layers as energy storage reservoirs has advantages such as good sealing, large energy storage capacity, and relatively low cost, and has become the mainstream site selection method for CGS power plants. However, the safety and stability of underground salt cavern energy storage reservoirs face severe challenges under extreme conditions of high pressure, high flow rate, and frequent cyclic alternation. The creep characteristics of salt rock, the integrity of the caprock and surrounding rock, and the multi-physics coupling effects during injection and extraction together constitute a highly complex and uncertain geomechanical system.

[0003] Currently, monitoring of salt cavern energy storage relies primarily on discrete sensor networks and periodic manual inspections. Conventional monitoring methods include measuring cavity pressure and temperature, conducting periodic sonar measurements to assess geometric changes, deploying microseismic monitoring networks to capture rock fracture signals, and using technologies such as InSAR to monitor surface deformation. While these methods provide valuable field data, they essentially fall under the categories of "data acquisition" and "post-analysis." Their limitations are mainly reflected in the following aspects: because various monitoring data are processed independently, there is a lack of an effective multi-source information fusion mechanism, making it difficult to form a unified and real-time understanding of the state of complex underground systems; furthermore, existing numerical simulation techniques often use deterministic geomechanical models, and their input parameters (such as salt creep parameters and formation strength) are mainly determined based on limited exploration and laboratory data, which cannot fully characterize the spatial variability and cognitive uncertainty of actual geological conditions. Therefore, model predictions often deviate from actual conditions in ways that are difficult to quantify, making the models unsuitable for high-confidence safety predictions and decision support. Furthermore, existing models typically determine the safe operating boundaries of energy storage facilities (such as maximum / minimum operating pressures) based on conservative assessments and design specifications from the geological exploration phase, failing to consider the evolution of geological conditions during long-term operation (such as cavity creep and shrinkage, damage accumulation). This can lead to insufficient safety in later stages of operation due to the accumulation of unknown risks. To address these issues, a monitoring system for underground salt cavern compressed air energy storage facilities based on digital twins is proposed. Summary of the Invention

[0004] The main objective of this invention is to provide a monitoring system for underground compressed air energy storage facilities based on digital twins. By deeply integrating multi-source monitoring data, dynamically quantifying and updating geological knowledge, and making forward-looking decisions and adaptive control based on real-time risks, this system ensures the safe, efficient, and economical operation of underground energy storage facilities in compressed air energy storage power plants throughout their entire life cycle, effectively solving the problems in the background technology.

[0005] To achieve the above objectives, the technical solution adopted by this invention is a monitoring method for underground salt cavern compressed air energy storage based on digital twins, comprising the following steps: S1: Based on geological exploration data, core experimental data, and prior geological knowledge, determine the set of key uncertainty parameters affecting the multiphysics behavior of the energy storage, and define its prior probability distribution; S2: Use the Monte Carlo method to sample from the prior probability distribution to generate a model set containing N sample parameters; Based on each sample parameter, simulate the dynamic response of the energy storage under a preset injection-production scheme through a multiphysics coupled forward model to construct an initial probabilistic digital twin; S3: Collect multivariate monitoring data of the energy storage in real time or periodically. The monitoring data is used as observation evidence through fusion and registration in a unified spatiotemporal coordinate system. The sequential Bayesian update algorithm is used to update the weights and resample the model set in the initial probabilistic digital twin to obtain a posterior model set reflecting the current cognitive state. S4: Based on the posterior model set, the state of the energy storage facility in a specified future period is extrapolated in parallel to generate prediction results with probability confidence intervals. The proportion of models in the posterior model set that trigger preset risk events is counted, and the trigger probability of the preset risk events is calculated. S5: Based on the trigger probability or the confidence interval of the prediction results, the safe operation boundary and injection-production strategy of the energy storage facility are dynamically adjusted.

[0006] Furthermore, the set of key uncertainty parameters includes at least the creep constitutive parameters of the salt rock, the fracture toughness of the caprock and surrounding rock, permeability parameters, and initial geostress field parameters; the prior probability distribution is a uniform distribution, a log-normal distribution, or an empirical distribution constructed based on geological analogy data.

[0007] Furthermore, the multi-source monitoring data includes at least two of the following: cavity pressure, temperature, microseismic events, surface deformation, and cavity geometry data.

[0008] Furthermore, the sequential Bayesian update algorithm specifically involves using a particle filter method to adjust the weights of the i-th sample model in the model set at time t. The update is performed using the following formula: ,in, Let be the parameter vector of the model for the i-th sample. Let be the observation data vector at time t. The likelihood function represents the conditional probability of the observed data under given model parameters. After normalizing the weights, the model set is resampled when the number of valid samples is lower than the preset resampling trigger threshold.

[0009] Furthermore, the likelihood function is constructed based on the residuals between the observed data and the model prediction data, and considers the independent error distributions of different types of observed data. The calculation method is as follows: K represents the number of observed data types. Let V be the error variance of the k-th class of observations. Let be the observation operator corresponding to the model parameter set θ at time t.

[0010] Furthermore, in step S4, generating the prediction result with probability confidence intervals specifically includes: summarizing the predicted values ​​of all samples in the posterior model set for the same physical quantity in the future; calculating the statistical quantile of the predicted value of the physical quantity to determine its variation range at different confidence levels.

[0011] Furthermore, the dynamic adjustment of the safe operation boundary of the energy storage specifically includes: determining the conditions that are met through iterative solution. Maximum permissible operating pressure ,in In order to operate under pressure The probability of triggering the risk is given by α, which is a preset acceptable risk threshold.

[0012] Furthermore, the trigger probability of the preset risk event The calculation is performed using the following formula: ;in, For failure function, For indicator functions, Let be the weight of the i-th sample parameter at time t. Where N is the time interval and N is the number of sample parameters.

[0013] A monitoring system for underground salt cavern compressed air energy storage based on digital twins is provided to implement the monitoring method for underground salt cavern compressed air energy storage based on digital twins, comprising: a data acquisition and fusion module for acquiring multivariate monitoring data of the energy storage in real time or periodically, and fusing and registering the multivariate monitoring data through a unified spatiotemporal coordinate system; a probabilistic digital twin construction module for: determining the set of key uncertainty parameters affecting the multiphysics behavior of the energy storage based on geological exploration data, core experimental data, and prior geological knowledge, and defining its prior probability distribution; sampling from the prior probability distribution using the Monte Carlo method to generate a model set containing N sample parameters; simulating the dynamic response of the energy storage under a preset injection-production scheme through a multiphysics coupled forward model based on each sample parameter, and constructing an initial probabilistic digital twin; and a sequential Bayesian assimilation and update module, connected to the data acquisition and fusion module and the probabilistic digital twin construction module. The system is configured to: use the fused and registered multivariate monitoring data as observational evidence, and employ a sequential Bayesian update algorithm to update and resample the model set in the initial probabilistic digital twin to obtain a posterior model set reflecting the current cognitive state; a probability prediction and risk assessment module, connected to the sequential Bayesian assimilation and update module, is configured to: perform parallel extrapolation of the state of the energy storage facility within a specified future period based on the posterior model set, and generate prediction results with probability confidence intervals; statistically analyze the proportion of models in the posterior model set that trigger preset risk events, and calculate the trigger probability of the preset risk events; and an adaptive decision optimization module, connected to the probability prediction and risk assessment module, is configured to: dynamically adjust the safe operation boundary of the energy storage facility based on the trigger probability or the confidence interval of the prediction results; and dynamically generate or adjust the injection and extraction strategy of the energy storage facility based on the trigger probability, the confidence interval of the prediction results, and preset optimization objectives.

[0014] A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the monitoring method for underground salt cavern compressed air energy storage based on digital twins as described above. This invention has the following beneficial effects: Compared with existing technologies, this solution, by constructing a probabilistic digital twin, integrates dispersed, multi-source monitoring data (pressure, deformation, microseismic activity, etc.) into a unified virtual image of an underground system with uncertain metrics, overcoming the shortcomings of traditional methods such as isolated data and one-sided cognition, and achieving a fundamental shift from qualitative and fuzzy judgment to quantitative and probabilistic cognition.

[0015] Compared with existing technologies, this solution can generate confidence interval predictions for key future safety parameters (such as stress and deformation) through parallel probabilistic extrapolation based on a posterior model set, and accurately calculate the trigger probability of specific risk events (such as caprock rupture and leakage) within a specified future period. This changes the traditional passive mode that relies on threshold alarms and post-event analysis, and can provide probabilistic early warnings before the risk actually occurs, enabling safety management to shift from post-event remediation to pre-event prevention.

[0016] Compared with existing technologies, this solution, based on real-time updated risk probabilities and confidence intervals, can inversely solve for the maximum safe operating pressure window under the current level of understanding. It dynamically adjusts the operating pressure as understanding is updated and the system state evolves, breaking through the fixed and conservative static safety boundary management model. When the understanding is clear and the risk is low, the operating boundary can be safely widened to fully tap the peak-shaving potential of the energy storage. When uncertainty increases or the risk rises, the boundary is tightened to ensure absolute safety, thereby achieving a dynamic balance between safety and economy. Attached Figure Description

[0017] Figure 1 is a flowchart illustrating the monitoring method for underground salt cavern compressed air energy storage based on digital twins according to the present invention; Figure 2 is a structural schematic diagram illustrating the monitoring system for underground salt cavern compressed air energy storage based on digital twins according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Embodiments

[0019] Referring to Figure 1, which shows a flowchart of a monitoring method for underground salt cavern compressed air energy storage based on digital twins according to the present invention, the method includes the following steps: Step 1: Based on geological exploration data, core experimental data and prior geological knowledge, determine the set of key uncertainty parameters affecting the multi-physics behavior of the energy storage, and define its prior probability distribution.

[0020] The set of key uncertainty parameters includes at least the creep constitutive parameters of the salt rock, the fracture toughness of the caprock and surrounding rock, permeability parameters, and initial geostress field parameters; the prior probability distribution is a uniform distribution, a log-normal distribution, or an empirical distribution constructed based on geological analogy data.

[0021] In one possible implementation, when defining the prior probability distribution of the set of key uncertainty parameters using an empirical distribution constructed based on geological analogy data, the specific steps can be as follows: Step 1.1: Establish an analogy database to collect and organize "reference database" data with similar geological backgrounds to the target energy storage.

[0022] Step 1.11: Determine the analogy criteria, including: geological age and sedimentary environment: belonging to the same salt basin or similar sedimentary facies (e.g., both are marine evaporite); salt rock mineral composition: the main components (e.g., halite, potash, anhydrite) have similar contents; caprock lithology: both are mudstone, gypsum, dense carbonate rocks, etc.; tectonic setting: located in similar tectonic stress fields (e.g., both are stable platform areas or both are slightly compressed).

[0023] Step 1.12: Data Sources: Existing storage facilities in the same salt basin: Publicly available research reports, monitoring data, and core test reports of other operating or abandoned salt cavern energy storage facilities and oil storage facilities in the same area.

[0024] Salt mine data: Rock mechanics and hydrogeological data accumulated during the mining of salt from the same salt layer.

[0025] Oil and gas field data: Drilling, logging, and geostress test data (such as leakage test and hydraulic fracturing data) of deep subsalt oil and gas reservoirs in the same area.

[0026] Publicly available literature and industry databases: such as the publicly available engineering parameters and research results of the US Strategic Petroleum Reserve (SPR) salt caverns, German energy storage facilities, and China's Jintan energy storage facility.

[0027] Step 1.2: Extraction and normalization of key parameters to extract effective parameter samples that can be used for statistical analysis.

[0028] Step 1.21: Parameter Extraction: Salt Rock Creep Parameters: Extract steady-state creep equations (such as the power-law creep equation ε=Aσ) from literature or reports. n Record the material constants A and n in the test. Record the confining pressure and temperature conditions.

[0029] Fracture toughness: K obtained from three-point bending test or Brazilian splitting test {IC} value.

[0030] Permeability: Data obtained from laboratory core gas permeability tests or permeability values ​​obtained through mercury intrusion porosimetry or well logging interpretation.

[0031] Initial geostress: The magnitude of the horizontal principal stress (S) obtained from tests such as hydraulic fracturing, wellbore collapse, and core testing. h ,S H ) and direction, vertical stress gradient (S) v ).

[0032] Step 1.22: Data cleaning and normalization: Outlier removal: Use statistical methods (such as the 3σ principle) to remove obviously unreasonable data points.

[0033] Depth / stress normalization: Normalizes mechanical parameters (such as strength and elastic modulus) to confining pressure or depth conditions similar to the target reservoir. Coarse corrections can be made using empirical formulas.

[0034] Temperature correction: Correct temperature-sensitive parameters such as creep parameters to the values ​​corresponding to the formation temperature of the target reservoir.

[0035] Step 1.3: Construction of empirical distribution and characterization of uncertainty, transforming the processed sample data into a prior distribution that can be used for probability sampling.

[0036] In one possible implementation, a nonparametric statistical method can be used, specifically: kernel density estimation: when the sample size is sufficient (e.g., more than 10 valid samples for each parameter), the continuous probability density function (PDF) of the parameter can be directly estimated using KDE. ;in, For sample values, Here, h is the kernel function (e.g., Gaussian kernel), and h is the bandwidth. KDE most accurately reflects the original distribution of data, including features such as multimodality and skewness.

[0037] In one possible implementation, the uncertainty envelope method can be used, specifically: when data is extremely scarce, instead of fitting an exact distribution, the minimum value of the analogous data is directly taken. and maximum value This forms a uniform distribution: .

[0038] Step 1.4: Prior Distribution Evaluation Step 1.41: Visualization and Physical Consistency Check The constructed empirical distribution is overlaid with the original data histogram to check its rationality.

[0039] Assess whether the distribution range conflicts with known geophysical laws (e.g., whether the permeability of salt rock is within a certain range). (within the typical range).

[0040] Step 1.42: Incorporating Expert Knowledge: Organize geological and rock mechanics experts for review. Experts can subjectively adjust the empirical distribution based on their understanding of the specific characteristics of the target reservoir. For example, they can appropriately shift the mean of the KDE-estimated distribution towards the direction of the expert's judgment, or appropriately increase the standard deviation.

[0041] Step 2: Use the Monte Carlo method to sample from the prior probability distribution to generate a model set containing N sample parameters.

[0042] Specifically, the above steps can be implemented through the following process: Step 2.1: Establish a standardized sampling framework.

[0043] Step 2.11: Parameter Vector Definition: Combine the m key uncertainty parameters into a vector: θ = [θ1, θ2, ..., θ m For example: θ=[A,n,K] {IC} ,k,S h ,S H ], where A is the creep coefficient, n is the stress exponent, and K {IC} Where S is the fracture toughness, k is the permeability, and S is the tensile strength. h With S H These represent the minimum and maximum horizontal principal stresses.

[0044] Step 2.12: Construction of the joint prior distribution: Assuming that all parameters are independent, the joint prior probability density function is: ,in For parameters The marginal prior distribution (derived from the empirical distribution constructed in step 1).

[0045] Step 2.2: Sampling Execution and Sample Generation Step 2.21: Determine the sample size N. Methods include: 1) Determining based on the cost calculated by the subsequent model. Typically, N is within... Magnitude.

[0046] 2) Preliminary estimation can be made using trial sampling and convergence diagnosis: First, sample a small set (e.g., N=500) and calculate the statistics of key outputs (e.g., the mean and variance of predicted pressure); gradually increase N until the change in statistics is less than the preset tolerance (e.g., 1%), at which point N can be used as the final sample size.

[0047] Step 2.22: Select and execute the sampling algorithm: In a feasible approach, basic independent sampling can be used: if the parameters are independent, directly sample from each... N independent samples are extracted from the sample and then combined into N parameter vectors.

[0048] In another feasible approach, Latin hypercube sampling can be used: to ensure that the marginal distribution of each parameter is uniformly covered, by dividing the cumulative distribution function of each parameter into N equally probable intervals, randomly sampling a point in each interval, and then randomly pairing them to generate vectors, better space filling can be achieved with fewer samples.

[0049] Step 3: Based on the parameters of each sample, simulate the dynamic response of the energy storage tank under the preset injection and production scheme through a multi-physics coupled forward model, and construct an initial probabilistic digital twin.

[0050] It should be noted that the behavior of salt cavern energy storage during the injection and production process is a typical multi-field coupling problem involving heat, fluid, solidification, and chemical processes. A complete forward model typically includes the following core physical fields and their interactions: 1) Fluid dynamic field governing equations: based on the Navier-Stokes equations or Darcy's law (for porous media flow) with conservation of mass and momentum.

[0051] Simulation content: The compression, expansion, and flow process of high-pressure air (or nitrogen) in the cavity; gas seepage through porous / fractured media such as salt rock / caprock.

[0052] Key outputs: pressure distribution, gas density distribution, and flow velocity field within the cavity.

[0053] 2) Thermodynamic field governing equations: energy conservation equations.

[0054] Simulation content includes: compression and heating effect during gas injection, expansion and cooling during gas extraction (Joule-Thomson effect), and unsteady heat exchange between gas and surrounding rock.

[0055] Key output: Evolution of the temperature field of the gas and surrounding rock inside the cavity.

[0056] 3) Geomechanical field governing equations: equilibrium equations based on momentum conservation, combined with constitutive relations.

[0057] Constitutive model: Salt rock: A viscoelastic-plastic model that considers transient and steady-state creep (such as Norton power-law creep, Lubby2 model, etc.) is used to simulate its rheological behavior related to time, stress and temperature.

[0058] Overburden and surrounding rock: Elastic-plastic or damage-plastic models are usually used to simulate their deformation, damage and fracture behavior.

[0059] Simulation content: The evolution of stress and strain field in salt cavern, wellbore and surrounding strata under pressure, temperature changes and self-weight.

[0060] Key outputs: displacement field, stress field, strain field, damage variable, and plastic zone range.

[0061] 4) Chemical field governing equations: transport and reaction equations.

[0062] Simulation content includes: water vapor condensation / evaporation, dissolution and recrystallization of salt rock, and chemical reactions of impurity gases. For long-term operation, this field is valuable for assessing the evolution of the mechanical properties of salt rock and its sealing performance.

[0063] Coupling mechanism: Fluid-structure interaction: Gas pressure acts on the cavity wall and cracks, causing deformation; the deformation, in turn, changes the porosity and permeability, affecting the flow.

[0064] Thermo-mechanical coupling: Temperature changes affect gas physical properties (density, viscosity) and rock mechanical parameters (creep rate, strength); flow and deformation processes are accompanied by thermal effects.

[0065] Fully coupled thermal-fluid-solid modeling is the most faithful simulation method, simultaneously solving the aforementioned equations, but it is computationally extremely expensive. In engineering, sequential coupling or partial coupling approximation methods are often used to improve efficiency.

[0066] The dynamic response content of the forward model simulation output, namely the dynamic response of the energy storage system, is the bridge connecting the virtual model and the real monitoring. It mainly includes the following four categories: 1) Physical quantities that can be directly observed, which may include: overall cavity response; pressure-time history; pressure change curve at the wellhead or top of the cavity.

[0067] Temperature-time history: Temperature change curves at key points (such as the bottom of the well or the middle of the cavity).

[0068] Cavity volume change: The change in effective working volume over time due to salt rock creep and pressure deformation.

[0069] Surface and near-field response: Surface displacement field: vertical settlement / uplift, horizontal displacement and their changes over time at each monitoring point.

[0070] Near-field strain: The strain history of the well casing or near-cavity formation.

[0071] Microseismic activity response: Event rate: The number of microseismic events occurring per unit time.

[0072] Source parameter distribution: the set of spatial location (three-dimensional coordinates), magnitude, and time of occurrence of the simulated microseismic events.

[0073] 2) Intrinsic state quantities that are not directly observable but are critical for safety assessment may include: stress field evolution: the evolution of the magnitude and direction of principal stresses in key areas such as the cavity periphery, caprock, and faults.

[0074] Stress concentration factor, yield proximity.

[0075] Damage and failure indicators: expansion of the plastic strain region.

[0076] Spatial distribution and evolution of damage factors (e.g., from 0 to 1, representing integrity to complete destruction).

[0077] Fracture driving factors are used to determine whether a crack has started and is propagating.

[0078] Sealing index: Changes in the equivalent permeability field of the caprock and wellbore cement sheath.

[0079] The migration path and extent of the gas seepage front.

[0080] 3) Performance indicators used for economic and efficiency evaluation may include: Cycle efficiency: the ratio of output energy to input energy in a complete injection-production cycle.

[0081] Working gas volume: The volume of gas available within the safe pressure window.

[0082] Maximum injection / production power / rate: The maximum capability that the equipment can achieve under the current conditions.

[0083] 4) Risk event trigger flags used for probability calculation may include: "event tags" automatically recorded by the model during the simulation process according to preset criteria, such as: cap layer shear failure = TRUE / FALSE (triggered when the shear stress of a certain element exceeds the threshold).

[0084] Cavity instability = TRUE / FALSE (triggered when shrinkage exceeds limit or shape is distorted).

[0085] Seal failure = TRUE / FALSE (triggered when gas seepage channels penetrate the cap layer).

[0086] Specifically, this can be achieved through the following steps: Step 3.1: Run the physical model for each parameter sample to generate a predicted response database.

[0087] Step 3.11: Setting up an automated simulation process: Develop a script (such as a Python script) to automatically execute the following loop: 1) Input generation: Generate the i-th parameter vector Write the input files required for the forward model (such as ABAQUS .inp files, COMSOL .java files, or configuration files for custom code).

[0088] 2) Model Invocation and Execution: Invoke the corresponding multiphysics solver (such as finite element method or finite volume method) and run the simulation on the specified computing resources.

[0089] 3) Output Extraction and Storage: Extract key spatiotemporal response data from the simulation result file, such as: cavity pressure-time series, displacement and stress history of key points in the caprock, cavity volume change curves caused by salt rock creep, and microseismic activity indicators (such as plastic strain regions). 4) Data Storage: θ (i) and its corresponding output response d (i) As a record, it is stored in a high-performance database (such as HDF5 format) or an array.

[0090] Step 3.2: High-performance computing strategy: Parallelization: Since the samples are independent of each other, this process can be highly parallelized. Use a cluster computing job scheduling system (such as Slurm, PBS) to distribute N tasks to multiple computing nodes for simultaneous execution.

[0091] Fault tolerance mechanism: Set checkpoints. If a simulation of a sample fails (e.g., due to mesh distortion causing non-convergence), it will be automatically recorded and replaced with a new sampling point for recalculation.

[0092] Step 3.3: Ensure that the generated model set is valid and can be used for subsequent assimilation.

[0093] Step 3.31: Sample distribution verification: Visually examine the sample histogram for each parameter and compare it with the target prior distribution to ensure that the sampling correctly covers the range of uncertainty.

[0094] Step 3.32: Physical rationality filtering: Set simple physical rules (such as "the minimum principal stress of the cap layer must be greater than the maximum pressure in the cavity") to automatically remove parameter samples that obviously violate basic physical principles (even though they may be drawn from the statistical distribution).

[0095] Step 3.33: Construct the proxy model (optional but crucial): As an input-output pair, train a Gaussian process regression model or a deep neural network. This surrogate model can approximate... The mapping is faster than the original physics model, and the evaluation speed is several orders of magnitude faster.

[0096] Final output: A structured initial probabilistic digital twin, the core of which includes: 1) the parameter sample matrix Θ N×M Each row is a parameter vector θ (i) .

[0097] 2) Corresponding predictive response database D N×(t×s) Where t is the time step and s is the number of observation points.

[0098] 3) The trained surrogate model is used for fast prediction.

[0099] 4) Sample weight vector W0 = [1 / N, 1 / N, ..., 1 / N] T : Initialize all weights to be equal.

[0100] Step 4: Collect multi-dimensional monitoring data of the energy storage facility in real time or periodically, fuse and register them through a unified spatiotemporal coordinate system, and use the monitoring data as observation evidence.

[0101] The multi-source monitoring data includes at least two of the following: cavity pressure, temperature, microseismic events, surface deformation, and cavity geometry data.

[0102] Specifically, the process includes the following steps: Step 4.1: Data synchronization and timestamp alignment: All monitoring devices are connected to a unified network time protocol server to ensure that data timestamps are synchronized to the millisecond level.

[0103] Set a uniform data sampling / reporting period (e.g., 1 minute, 1 hour, 1 day), periodically resample continuous data (pressure, temperature), and summarize event data (micro-seismic events) over time periods.

[0104] Step 4.2: Spatial Coordinate System and Data Mapping: Establishing a Global Coordinate System: Define a three-dimensional rectangular coordinate system (or combine it with a geographic coordinate system) with the cavity center as the origin.

[0105] Sensor position calibration: Accurately determine the spatial coordinates (x, y, z) of each sensor (bottom hole pressure gauge, microseismic detector, surface InSAR reference point) in the global coordinate system.

[0106] Data spatialization: Point data (pressure, temperature): directly associated with its sensor coordinates.

[0107] Surface data (InSAR deformation map): interpolates deformation raster data onto preset surface grid point coordinates.

[0108] Volume data (sonar scan): Transform the point cloud on the cavity surface into the global coordinate system.

[0109] Event data (microseismic): Each event includes the source coordinates (x, y, y). e ,y e ,z e ), time of occurrence, and magnitude.

[0110] Step 4.3: Construct the normalized observation vector: at each assimilation time t k The registered multivariate data are packaged into an observation vector. Example: Given an assimilation period of 1 day, the observation vector might contain: ;in, These are the wellhead pressure and temperature, respectively. This represents the number of microseismic events in a specific area over the past 24 hours. Let be the cumulative InSAR line-of-sight deformation at the r-th surface grid point. This represents the change in cavity volume from the previous sonar scan.

[0111] Step 5: Use the sequential Bayesian update algorithm to update the weights and resample the model set in the initial probabilistic digital twin to obtain the posterior model set that reflects the current cognitive state.

[0112] Specifically, the sequential Bayesian update algorithm involves using a particle filter method to update the weights of the i-th sample model in the model set at time t. The update is performed using the following formula: ,in, Let be the parameter vector of the model for the i-th sample. Let be the observation data vector at time t. Let be the likelihood function, representing the conditional probability of observed data given model parameters. The likelihood function is constructed based on the residuals between observed data and model-predicted data, and considers the independent error distributions of different types of observed data. The calculation method is as follows: K represents the number of observed data types. Let V be the error variance of the k-th class of observations. For time t, the observation operator corresponding to the model parameter set θ; after normalizing the weights, when the number of valid samples is lower than the preset resampling trigger threshold, the model set is resampled.

[0113] Step 6: Based on the posterior model set, perform parallel extrapolation of the state of the energy storage facility within a specified future period, generate prediction results with probability confidence intervals, count the proportion of models in the posterior model set that trigger preset risk events, and calculate the trigger probability of preset risk events.

[0114] Specifically, generating prediction results with probability confidence intervals includes: summarizing the predicted values ​​of all samples in the posterior model set for the same physical quantity in the future; and calculating the statistical quantiles of the predicted value of the physical quantity to determine its range of variation at different confidence levels.

[0115] Preset the probability of risk events The calculation is performed using the following formula: ;in, For failure function, For indicator functions, Let be the weight of the i-th sample parameter at time t. Where N is the time interval and N is the number of sample parameters.

[0116] Specifically, the following process can be used to achieve this: Step 6.1: Prediction scenario setting and initialization Step 6.11: Determine the prediction time domain: Short-term prediction: the next few hours to days, used to guide real-time injection and collection operations.

[0117] Medium-term forecast: Over the next few weeks to months, this will be used to develop maintenance and testing plans.

[0118] Long-term forecasting: Over the next few years to decades, used to assess the lifespan and long-term stability of the storage facility.

[0119] Step 6.12: Define the predicted operating conditions: Specify the injection and production plan for the future time period (pressure and flow rate change curves over time).

[0120] Consider different scenarios, such as baseline scenario, maximum load scenario, emergency shutdown scenario, etc.

[0121] Step 6.13: Define the predicted output quantities: safety-related physical quantities: maximum principal stress of the cavity, minimum principal stress of the caprock, cavity shrinkage volume, maximum surface settlement, and well casing stress.

[0122] Performance-related physical quantities: effective working volume of the cavity, circulating gas volume, injection and extraction efficiency.

[0123] Risk event indicator: An indicator function used to determine whether a preset risk event has occurred.

[0124] Step 6.2: Parallel inference based on the posterior set: Using the updated cognition (posterior set), predict the system's various possible behaviors in the future.

[0125] Step 6.21: Posterior set preparation: Load the current assimilation time t k The posterior model set: ,in For model parameters, Set its normalized weights.

[0126] Step 6.22: Parallel Forward Simulation: For each sample i=1,2,...,N in the set: 1) Set parameters: Set the model parameters to... 2) Set the initial state: Set t k 3) Apply the predicted conditions: load the predefined future injection and extraction plan as boundary conditions; 4) Execute the simulation: run the multiphysics forward model (or call its high-precision proxy model), starting from t k Integral to t k +Δt (predicted time domain endpoint); 5) Extract the predicted trajectory: Record the sequence of changes in the key physical quantities predicted by the model over time, denoted as , where t∈[t k ,t k +Δt];Step 6.23: Generate a weighted prediction set: the prediction result for each sample Each comes with its own weight. .

[0127] The set of all weighted predictions Discrete approximation of the probability distribution that constitutes the future behavior of the system.

[0128] Step 6.3: Calculate the probability confidence interval. Extract easily understandable statistics from the weighted prediction set to quantify the uncertainty of the prediction.

[0129] Step 6.31: At the specified prediction time point t p Summarize: For each physical quantity of interest Y (such as cavity volume), at time point t pCollect predicted values ​​from all samples. and their weights .

[0130] Step 6.32: Calculate weighted statistical quantiles: 1) Sort: Arranged in ascending order of value, we obtain the sequence {Y}. (1) ,Y (2) ,...,Y (N)}, and remember the original weight w corresponding to each value. (j) .

[0131] 2) Calculate the cumulative weight: Calculate the cumulative weight of the sorted sequence. .

[0132] 3) Find the quantile position: For a confidence level α (e.g., α=0.9 corresponds to a 90% confidence interval), the lower bound corresponds to the position with a cumulative weight of (1−α) / 2, and the upper bound corresponds to the position with a cumulative weight of 1−(1−α) / 2.

[0133] 4) Interpolation: Since the cumulative weight is a step function, linear interpolation is usually used to calculate the accurate quantile values.

[0134] Example formula (lower bound): .

[0135] Step 6.33: Generate a confidence interval sequence: For a series of discrete time points {t1, t2, ..., t...} in the future prediction time domain... m Repeat the above process.

[0136] Output the predicted median (50th percentile) for each physical quantity at each time step and the upper and lower bounds for the specified confidence level (e.g., 90%).

[0137] Visualization: With time t as the horizontal axis and physical quantity Y as the vertical axis, a "prediction cone" can be drawn. The cone-shaped area is the confidence interval, which intuitively shows the propagation of uncertainty over time.

[0138] Step 6.4: Calculate the probability of risk event triggering, quantifying the likelihood of a specific catastrophic event occurring in the future.

[0139] Step 6.41: Pre-determine risk events and criteria: Clearly define what constitutes a "risk event occurrence". For example: caprock shear failure: When the effective shear stress τ of a certain element in the caprock exceeds its shear strength τ f When, i.e., g1=τ / τ f -1 > 0; Excessive cavity contraction: When the cavity volume loss exceeds the allowable limit V lim At that time, g2 = ΔV - V lim>0; Gas leakage: When a penetrating seepage channel is formed in the caprock (permeability surges in the simulation), i.e., g3=I(connectivity)−1=0.

[0140] Step 6.42: Traverse the set to perform risk assessment: For each posterior sample i and its predicted trajectory :1) Throughout the entire prediction time domain, check whether the above criterion (i.e., g) is triggered at any time and any spatial location. j (Is >0 true?)

[0141] 2) Record a binary flag This indicates whether risk event j will occur in the future as predicted by the model.

[0142] Step 6.43: Calculate the trigger probability: the posterior predicted probability of risk event j occurring, which is the weighted average of the occurrence markers of that event across all models. Output: A set of probability values, for example, p 盖层破坏 =0.003, p 过度收缩 =0.12.

[0143] Step 6.5: Results Synthesis and Report Generation Step 6.51: Generate Prediction and Risk Assessment Report: Includes: predicted median curves and confidence intervals of key physical quantities, and a list of trigger probabilities for each risk event.

[0144] Step 6.52: Spatiotemporal risk heatmap generation: For certain risks (such as damage), the proportion of each spatial unit that will be damaged during the prediction period can be statistically analyzed (weighted sum) to generate a "spatial distribution map of the probability of damage in the future period".

[0145] Step 6.53: Decision-ready format output: Encapsulate the prediction results and risk probabilities to trigger subsequent optimization or alarm processes.

[0146] Step 7: Based on the confidence interval of the trigger probability or prediction results, dynamically adjust the safe operation boundary and injection / production strategy of the energy storage facility.

[0147] Specifically, dynamically adjusting the safety operation boundary of the energy storage facility includes: determining the conditions that are met through iterative solutions. Maximum permissible operating pressure ,in In order to operate under pressure The probability of triggering the risk is as follows. This is a preset acceptable risk threshold.

[0148] Specifically, it can be implemented according to the following process: Step 7.1: Dynamically adjust the safe operation boundary. Calculate and update the allowable upper and lower limits of energy storage pressure / flow based on the real-time risk assessment results.

[0149] Step 7.11: Input and Threshold Definition Input: Trigger probabilities of each risk event calculated in the above steps. and confidence interval.

[0150] Define the acceptable risk threshold α j Set a maximum tolerable probability (e.g., α) for each type of risk event. 盖层破坏 =10 −4 α 泄漏 =10 −3 ).

[0151] Step 7.12: Real-time safety pressure window calculation problem construction: The maximum / minimum allowable operating pressure (P) is calculated. max ,P min As the variable to be determined, we search for the largest feasible interval that satisfies all risk constraints.

[0152] Algorithm flow: 1) Initialization: Set the current operating pressure P current Based on the existing digital twin posterior set.

[0153] 2) One-sided search (with P) max For example): 2.1) Set the pressure test value P test , usually from P current Initially, increase the increment by a certain step size.

[0154] 2.2) In digital twins, rapid simulation (usually using a surrogate model) of all posterior samples at P test Response under pressure levels.

[0155] 2.3) Calculate the probability of triggering all risk events j under this pressure. .

[0156] 2.4) Check if the conditions are met. .

[0157] 2.5) Find the highest pressure P that satisfies all constraints. max That is, the boundary point. Similarly, P can be found. min .

[0158] Step 7.13: Output the dynamic security window: [P min (t),P max (t)]. This window changes dynamically with time t and cognitive state.

[0159] Step 7.14: Confidence interval driven margin management scenario: If the upper edge of the confidence interval for a key predictor (such as cavity shrinkage) is close to the design limit L.

[0160] Adjustment strategy: Conservative strategy: Further tighten the safety pressure boundary so that under the new pressure, the upper edge of the prediction confidence interval also satisfies UpperBound < L.

[0161] Formulaic expression: Finding P * Make Q 95 (Y(P * ))<L, where Q 95 This represents the 95th percentile predicted value.

[0162] Step 7.2: Dynamic optimization of injection and extraction strategy. Within the updated safety boundary, generate or adjust the injection and extraction plan for the next few hours to days to balance safety, efficiency and grid demand.

[0163] Step 7.21: Optimization Problem Modeling Decision Variables: Injection-production rate sequence Q=[Q1,Q2,...,Q] for the next M control time periods (e.g., each time period is 15 minutes). M ] T Or the target pressure sequence P=[P1,P2,...,P M ] T .

[0164] Objective function J(Q): Typically, it aims to maximize economic benefits or power grid service indicators, such as: ;where λ t Let η be the time-of-use electricity price, η be the cycle efficiency (which may be a function of pressure / rate), and γ be the risk penalty coefficient. This is a risk penalty term, the specific form of which can be designed according to risk characteristics and decision preferences, such as a threshold penalty function, a linear / proportional penalty function, an exponential penalty function, or a penalty function based on a confidence interval.

[0165] Constraints: 1) Safety boundary constraints: P min (t)≤P t (Q)≤P max (t), where P t (Q) is the cavity pressure calculated from the injection-production rate using a dynamic model of the energy storage tank; 2) Equipment capacity constraints: Q min ≤Q t ≤Q max ;3) Power grid dispatch instructions: Total gas injection and production volume requirements.

[0166] 4) Risk probability constraint: p failure,j (t;Q)≤α j p failure,j (t;Q) represents the risk probability at time t during the injection-production process using the injection-production rate sequence.

[0167] Step 7.22: Fast Policy Evaluation and Optimization Based on Digital Twins: The optimization algorithm (such as Model Predictive Control (MPC) or evolutionary algorithm) will propose multiple candidate policies Q. candidate .

[0168] Rapid simulation: For each candidate strategy, using the posterior set of digital twins and the surrogate model, its execution consequences are simulated in parallel within seconds, and its objective function value J and risk probability p are calculated. failure .

[0169] Optimization solution: Choose the policy Q that satisfies all constraints and maximizes J. ∗ As the optimal strategy.

[0170] Step 7.23: Strategy Classification and Execution Mode Green Mode (Low Risk, Narrow Confidence Interval): Adopt an aggressive strategy to track high grid price signals as much as possible within the safety window to improve returns.

[0171] Yellow mode (medium risk, or wide confidence interval): adopt a robust strategy, reduce the variation in injection and production rates, increase the pressure balance period, and prioritize safety.

[0172] Red mode (high risk, probability exceeds threshold): Triggers protection strategy, immediately stops current injection and production, and automatically switches to pressure stabilization or safe pressure relief procedure.

[0173] Step 7.3: Adjustment Instruction Generation and Human-Machine Collaboration Step 7.31: Instruction Packaging and Verification Transform the adjustment decision into a specific, executable set of instructions, for example: Setpoint instruction: Pressure target curve P for the next 2 hours ∗ .

[0174] Control mode command: Switch to "Constant rate gas injection" or "Pressure control mode".

[0175] Equipment start / stop command: Start / stop a compressor.

[0176] Pre-verification: Quickly simulate the execution of the instruction set in the digital twin to perform a final security check.

[0177] Step 7.32: Human-computer interaction and automatic authorization execution: For fine-tuning (such as minor optimization of the pressure target within the green range) or emergency protection (red mode), the system can directly issue instructions to the station control system for execution.

[0178] Manual confirmation: For major strategy mode changes (such as switching from green to yellow) or special operations (such as diagnostic testing), the system will push adjustment suggestions and risk assessment reports to the console, which must be confirmed by the operators before execution.

[0179] Step 7.33: Closed-loop learning and adaptive execution monitoring: The system monitors the actual response of the energy storage tank after adjustment.

[0180] Bias analysis: Compare the actual response with the digital twin's prediction. If a persistent and significant deviation occurs, a special update to the digital twin model is triggered (e.g., adding new uncertainty parameters), or the weight parameters in the optimization algorithm are adjusted (e.g., the risk penalty coefficient γ). Example

[0181] This invention also provides a monitoring system for underground salt cavern compressed air energy storage based on digital twins for implementing the above-mentioned method. Referring to the system structure diagram shown in Figure 2, it includes: a data acquisition and fusion module for real-time or periodic acquisition of multivariate monitoring data of the energy storage, and for fusing and registering the multivariate monitoring data through a unified spatiotemporal coordinate system; a probabilistic digital twin construction module for: determining the set of key uncertainty parameters affecting the multiphysics behavior of the energy storage based on geological exploration data, core experimental data, and prior geological knowledge, and defining its prior probability distribution; sampling from the prior probability distribution using the Monte Carlo method to generate a model set containing N sample parameters; simulating the dynamic response of the energy storage under a preset injection-production scheme using a multiphysics coupled forward model based on each sample parameter, and constructing an initial probabilistic digital twin; and a sequential Bayesian assimilation and update module, which works in conjunction with the data acquisition and fusion module and the probabilistic digital twin... The system is divided into three modules: a construction module and a probabilistic modeling module. The construction module uses fused and registered multivariate monitoring data as observational evidence, employs a sequential Bayesian update algorithm to update and resample the model set in the initial probabilistic digital twin, and obtains a posterior model set reflecting the current cognitive state. The probabilistic prediction and risk assessment module, connected to the sequential Bayesian assimilation and update module, performs parallel extrapolation of the energy storage's state over a specified future period based on the posterior model set, generating prediction results with probability confidence intervals. It also calculates the probability of triggering preset risk events by statistically analyzing the proportion of models in the posterior model set that trigger preset risk events. The adaptive decision optimization module, connected to the probabilistic prediction and risk assessment module, dynamically adjusts the energy storage's safe operating boundary based on the trigger probability or the confidence interval of the prediction results. It also dynamically generates or adjusts the energy storage's injection and extraction strategies based on the trigger probability, the confidence interval of the prediction results, and preset optimization objectives. (Example)

[0182] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the above-described monitoring method for underground salt cavern compressed air energy storage based on digital twins.

[0183] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A monitoring method for underground salt cavern compressed air energy storage based on digital twins, characterized in that, Includes the following steps: S1: Based on geological exploration data, core experimental data, and prior geological knowledge, determine the set of key uncertainty parameters affecting the multiphysics behavior of the energy storage reservoir, and define its prior probability distribution; S2: Use the Monte Carlo method to sample from the prior probability distribution to generate a model set containing N sample parameters; Based on each sample parameter, simulate the dynamic response of the energy storage reservoir under a preset injection-production scheme through a multiphysics coupled forward model, and construct an initial probabilistic digital twin; S3: Collect multi-dimensional monitoring data of the energy storage reservoir in real time or periodically, fuse and register them through a unified spatiotemporal coordinate system, and combine the monitoring data... As observational evidence, a sequential Bayesian update algorithm is used to update and resample the model set in the initial probabilistic digital twin to obtain a posterior model set reflecting the current cognitive state; S4: Based on the posterior model set, the state of the energy storage facility within a specified future time period is extrapolated in parallel to generate prediction results with probability confidence intervals. The proportion of models in the posterior model set that trigger preset risk events is statistically analyzed, and the trigger probability of the preset risk events is calculated; S5: Based on the trigger probability or the confidence interval of the prediction results, the safe operation boundary and injection / production strategy of the energy storage facility are dynamically adjusted.

2. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 1, characterized in that, In step S1, the set of key uncertainty parameters includes at least the creep constitutive parameters of the salt rock, the fracture toughness of the caprock and surrounding rock, the permeability parameters, and the initial geostress field parameters; the prior probability distribution is a uniform distribution, a log-normal distribution, or an empirical distribution constructed based on geological analogy data.

3. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 1, characterized in that, In step S3, the multi-source monitoring data includes at least two of the following: cavity pressure, temperature, microseismic events, surface deformation, and cavity geometry data.

4. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 1, characterized in that, In step S3, the sequential Bayesian update algorithm specifically involves using a particle filter method to adjust the weights of the i-th sample model in the model set at time t. The update is performed using the following formula: ,in, Let be the parameter vector of the model for the i-th sample. Let be the observation data vector at time t. The likelihood function represents the conditional probability of the observed data under given model parameters. After normalizing the weights, the model set is resampled when the number of valid samples is lower than the preset resampling trigger threshold.

5. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 3, characterized in that, The likelihood function is constructed based on the residuals between the observed data and the model-predicted data, and takes into account the independent error distributions of different types of observed data. The calculation method is as follows: K represents the number of observed data types. Let V be the error variance of the k-th class of observations. Let be the observation operator corresponding to the model parameter set θ at time t.

6. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 1, characterized in that, In step S4, generating a prediction result with a probability confidence interval specifically includes: summarizing the predicted values ​​of all samples in the posterior model set for the same physical quantity in the future; and calculating the statistical quantile of the predicted value of the physical quantity to determine its variation range at different confidence levels.

7. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 1, characterized in that, In step S4, the dynamic adjustment of the safe operation boundary of the energy storage specifically includes: determining the conditions that are met through iterative solution. Maximum permissible operating pressure ,in In order to operate under pressure The probability of triggering the risk is given by α, which is a preset acceptable risk threshold.

8. The monitoring method for underground salt cavern compressed air energy storage based on digital twins according to claim 1, characterized in that, In step S4, the trigger probability of the preset risk event is... The calculation is performed using the following formula: ;in, For failure function, For indicator functions, Let be the weight of the i-th sample parameter at time t. Where N is the time interval and N is the number of sample parameters.

9. A monitoring system for underground salt cavern compressed air energy storage based on digital twins for implementing the method as described in any one of claims 1-8, characterized in that, include: The data acquisition and fusion module is used to acquire multivariate monitoring data of the energy storage facility in real time or periodically, and to fuse and register the multivariate monitoring data through a unified spatiotemporal coordinate system. The probabilistic digital twin construction module is used to: determine the set of key uncertainty parameters affecting the multiphysics behavior of the energy storage facility based on geological exploration data, core experimental data, and prior geological knowledge, and define its prior probability distribution; sample from the prior probability distribution using the Monte Carlo method to generate a model set containing N sample parameters; and, based on each sample parameter, simulate the dynamic response of the energy storage facility under a preset injection-production scheme through a multiphysics coupled forward model, constructing an initial probability model. The system includes: a probabilistic digital twin; a sequential Bayesian assimilation and update module, connected to the data acquisition and fusion module and the probabilistic digital twin construction module, used to: use the fused and registered multivariate monitoring data as observation evidence, and employ a sequential Bayesian update algorithm to update and resample the model set in the initial probabilistic digital twin to obtain a posterior model set reflecting the current cognitive state; and a probability prediction and risk assessment module, connected to the sequential Bayesian assimilation and update module, used to: perform parallel extrapolation of the state of the energy storage facility within a specified future period based on the posterior model set, and generate prediction results with probability confidence intervals. The proportion of models in the posterior model set that trigger the preset risk event is statistically analyzed, and the trigger probability of the preset risk event is calculated. An adaptive decision optimization module, connected to the probability prediction and risk assessment module, is used to: dynamically adjust the safe operation boundary of the energy storage facility based on the trigger probability or the confidence interval of the prediction result; and dynamically generate or adjust the injection and extraction strategy of the energy storage facility based on the trigger probability, the confidence interval of the prediction result, and a preset optimization target.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the monitoring method for underground salt cavern compressed air energy storage based on digital twins as described in any one of claims 1 to 8.

Images