Electrochemical energy storage system real-time emergency prevention and control system and method based on digital twinning

Abstract

The invention discloses a real-time emergency prevention and control system and method for an electrochemical energy storage system based on digital twinning, and belongs to the field of energy storage safety. The system comprises a multi-source sensing unit, a digital twin deduction unit and a visual emergency decision unit. The sensing unit collects multi-source data to form a state vector; the deduction unit outputs a future thermal runaway evolution path containing a temperature field, a gas concentration field and a fire blast risk probability and uncertainty measurement thereof in real time through a pre-trained Gaussian process regression agent model; and the decision-making unit performs three-dimensional visual rendering, jointly determines a high-confidence risk area based on the risk level and the uncertainty level, and generates an emergency prevention and control instruction with accurate spatial positioning. The problems that in the prior art, disaster evolution paths cannot be predicted in real time, and prevention and control measures are extensive are solved, millisecond-level online deduction and accurate active prevention and control of the thermal runaway process are achieved, and the safety of an energy storage system is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of electrochemical energy storage system safety technology, specifically to a real-time emergency prevention and control system and method for thermal runaway of energy storage systems based on digital twins and probabilistic extrapolation. Background Technology

[0002] With the advancement of global energy structure transformation, electrochemical energy storage systems based on lithium-ion batteries are being rapidly deployed on a large scale on both the grid and user sides. Energy storage containers / prefabricated cabins contain a large number of batteries with high energy density and relatively enclosed spaces. Under conditions of misuse, aging, thermal management failure, or electrical faults, individual battery anomalies can easily occur, further inducing thermal runaway. This can lead to gas production, pressure increases, and the formation of flammable mixtures, posing a risk of localized anomalies spreading throughout the cabin and causing combustion and explosion. To mitigate these risks, existing technologies typically focus on condition monitoring, fault diagnosis, early warning, and coordinated response in research and engineering applications.

[0003] For example, patent CN116345698A (application date: May 30, 2023; publication date: June 27, 2023) proposes an operation and maintenance management solution for energy storage power stations. This solution collects monitoring data such as voltage, current, temperature, pressure, and gas, performs data analysis, prediction, and visualization at the edge, and conducts operation and maintenance management and early warning in the cloud using digital twins. While this type of solution helps achieve multi-source data access and alarm management, in practice, it often relies on thresholds, rules, or single predicted values ​​for output. This often makes it difficult to provide spatiotemporal predictions of accident evolution and their reliability in the early stages of an incident, leading to subsequent measures tending towards global cooling and fire suppression, which are less likely to be precisely matched with critical locations and timing.

[0004] For example, patent CN120372996A (application date: February 28, 2025, publication date: July 25, 2025) discloses a method for fire ventilation and explosion relief safety assessment of lithium battery energy storage systems based on multi-dimensional simulation. It can simulate and assess failure scenarios for a given design scheme. However, it relies on complex simultaneous calculations such as thermal-fluid-solid, which takes a long time to simulate. It is more inclined to offline assessment and design optimization and is difficult to provide second-level dynamic simulation support when an accident occurs.

[0005] Regarding accident response, patent CN109513135B (application date: December 7, 2018, publication date: March 26, 2019) proposes an energy storage container fire suppression system, which uses sensors to locate the thermal runaway area and coordinates isolation, atomized spraying, and ventilation. Patent CN113332640A (application date: June 3, 2021, publication date: September 3, 2021) proposes a programmable spraying strategy, which uses a fire monitoring module to obtain the failure location and coordinates valve groups to achieve tiered spraying control. While these solutions can achieve a certain degree of coordinated response and directional spraying, their triggering is largely based on the current monitoring status or fire assessment. They lack real-time, visualized predictive support for the future propagation path of the "battery failure - thermal runaway - gas production - combustion and explosion" chain process, and also lack a unified mechanism to incorporate predictive uncertainty into decision-making rules to improve command reliability.

[0006] Therefore, there is still a need for an online simulation and emergency decision-making technology that covers the entire process of energy storage accidents, so as to dynamically reproduce and predict the safety status evolution of physical entities in virtual space, and provide executable spatial positioning commands for emergency response. Summary of the Invention

[0007] The technical problem to be solved by this invention is to overcome the technical defects of existing energy storage safety technologies, such as the inability to predict the thermal runaway evolution path in real time and the rough and inaccurate prevention and control commands caused by the lack of prediction.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] On the one hand, the present invention provides a real-time emergency control system for electrochemical energy storage systems based on digital twins, including: a multi-source sensing unit, a digital twin simulation unit, and a visualized emergency decision-making unit;

[0010] The multi-source sensing unit is installed inside the energy storage container and arranged correspondingly to the battery module. It is used to collect electrical parameters, temperature parameters, combustible gas concentration parameters and cabin pressure parameters, and fuse the collected results to form a real-time state vector.

[0011] The digital twin simulation unit is communicatively connected to the multi-source sensing unit and is used to receive the real-time state vector and output a prediction of the thermal runaway evolution path within a future preset time window.

[0012] The visualization emergency decision-making unit is communicatively connected to the digital twin simulation unit and is used to perform three-dimensional dynamic visualization rendering of the predicted thermal runaway evolution path, and generate spatial positioning emergency prevention and control instructions based on preset decision rules.

[0013] As an optional implementation, the digital twin inference unit includes a high-fidelity full-chain simulation module and a probabilistic proxy inference module;

[0014] The high-fidelity full-chain simulation module is pre-set with a high-dimensional physical model of the entire chain of battery thermal runaway-gas production-combustion and explosion, and performs sampling simulation in a preset fault parameter space to generate a training sample set. The training sample set includes at least the input conditions and the corresponding output evolution sequence.

[0015] The probabilistic proxy inference module completes training based on the training sample set and infers the real-time state vector during runtime to output the predicted thermal runaway evolution path.

[0016] As an optional implementation, the thermal runaway evolution path prediction output by the probabilistic proxy inference module includes at least: a temperature field spatiotemporal sequence, a key gas concentration spatiotemporal sequence, and a combustion and explosion risk probability distribution;

[0017] Furthermore, the probabilistic proxy deduction module also outputs an uncertainty metric corresponding to the predicted thermal runaway evolution path.

[0018] As an optional implementation, the probabilistic proxy inference module employs a multi-output Gaussian process regression model;

[0019] The multi-output Gaussian process regression model uses a composite kernel function, which includes at least a Matrn kernel component, a linear kernel component, and a noise kernel component.

[0020] The hyperparameters of the composite kernel function are optimized by maximizing the marginal likelihood function.

[0021] As an optional implementation, the visualization emergency decision-making unit includes a three-dimensional rendering module, which is used to decode the predicted mean in the thermal runaway evolution path prediction into three-dimensional dynamic visualization information, and to map the uncertainty measure into credibility display information associated with the three-dimensional dynamic visualization information.

[0022] As an optional implementation, the visualized emergency decision-making unit includes a decision rule engine, which is used to jointly determine high-confidence risk areas based on at least the following conditions:

[0023] (a) The probability distribution of the combustion and explosion risk or its corresponding risk index satisfies the first threshold condition;

[0024] (b) The uncertainty measure satisfies the second threshold condition;

[0025] Emergency prevention and control instructions with spatial positioning are generated based on the high-confidence risk zone.

[0026] As an optional implementation, the visualized emergency decision-making unit further includes a strategy library. The decision rule engine matches and generates instruction combinations from the strategy library based on the physical location and hazard type of the high-confidence risk zone. The instruction combinations include at least one of the following:

[0027] Directional spray coordinates or nozzle number, ventilation intensity curve parameters, and target module isolation markings.

[0028] As an optional implementation, the system further includes an online learning module, which is used for:

[0029] When the actual monitoring data and the prediction results of the probability proxy inference module show a continuous deviation within a preset time window and the deviation exceeds the adaptive threshold, the model update process is triggered.

[0030] The new real-time state-actual evolution data pairs are used as incremental samples to update the hyperparameters or training sample set of the probabilistic proxy inference module through Bayesian updates or sliding window retraining.

[0031] As an optional implementation, the multi-source sensing unit and the digital twin simulation unit are deployed in an edge computing device local to the energy storage container. The edge computing device is communicatively connected to the visual emergency decision-making unit to complete the formation of the real-time state vector, the output of the thermal runaway evolution path prediction, and the generation and output of the emergency prevention and control instructions locally.

[0032] On the other hand, the present invention also provides a real-time emergency control method for an electrochemical energy storage system based on digital twins, applied to the aforementioned system, the method comprising:

[0033] S1. Based on the high-fidelity full-chain simulation module, sampling simulation is performed in the preset fault parameter space to generate a training sample set containing input conditions and output evolution sequences;

[0034] S2. Based on the training sample set, train the probability proxy inference module so that it can output the thermal runaway evolution path prediction within a future preset time window and its corresponding uncertainty measure;

[0035] S3. The multi-source sensing unit collects electrical parameters, temperature parameters, combustible gas concentration parameters, and cabin pressure parameters and fuses them to form a real-time state vector;

[0036] S4. Input the real-time state vector into the probability proxy inference module to obtain the spatiotemporal sequence of the temperature field, the spatiotemporal sequence of the key gas concentration and the probability distribution of combustion and explosion risk, and obtain the uncertainty measure;

[0037] S5. The visualization emergency decision-making unit performs three-dimensional dynamic rendering of the predicted mean and maps the uncertainty measure into credibility display information;

[0038] S6. The decision rule engine determines the high-confidence risk area based on the first threshold condition and the second threshold condition, and generates and outputs the spatially located emergency prevention and control instructions.

[0039] This invention integrates a digital twin framework with a probabilistic machine learning model, Gaussian process regression, to construct a second-level real-time simulation capability for the entire chain of thermal runaway accidents in electrochemical energy storage systems. This overcomes the fundamental limitations of traditional high-fidelity simulations, which cannot be applied online, and fixed-threshold alarms, which lack predictive power, enabling safety responses to seize the golden window of opportunity for accident handling. Based on this, the system transforms the uncertainty measure output by the model into a decision-making basis, forming a closed-loop prevention and control system of "prediction-decision-execution." Specifically, based on the evolution path prediction with accompanying quantified uncertainty, it automatically identifies high-confidence risk areas and generates spatially precise instructions such as directional spray coordinates and target module isolation markers. This mechanism completely changes the traditional extensive response mode of area coverage, achieving precise intervention from surface to point, significantly improving prevention and control effectiveness while greatly reducing secondary damage and operational costs.

[0040] Furthermore, the system provides operations and maintenance personnel with intuitive and reliable in-depth situational awareness and decision support by overlaying the predicted path and uncertainty in a three-dimensional dynamic visualization. The built-in online learning module ensures that the model can continuously adapt to system changes such as battery aging and maintain the accuracy of predictions in the long term.

[0041] Finally, by deploying the sensing and simulation units on the local edge computing devices of the energy storage container, the system achieves millisecond-level real-time calculation and response, meeting the timeliness requirements under extreme emergency conditions. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0043] Figure 1 This is a schematic diagram of the overall architecture and data flow of a real-time emergency control system for an electrochemical energy storage system based on digital twins, provided by an embodiment of the present invention. Detailed Implementation

[0044] 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, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0045] Example 1:

[0046] like Figure 1 As shown, this embodiment provides a real-time emergency control system for electrochemical energy storage systems based on digital twins. Its core logic lies in constructing a real-time mapping and advanced simulation mechanism between physical and digital spaces. From an overall architecture perspective, the system mainly includes a multi-source sensing unit, a digital twin simulation unit, and a visualized emergency decision-making unit. These three components are closely integrated around the physical entity of the electrochemical energy storage system. This physical entity can be a containerized electrochemical energy storage system, internally equipped with battery modules, power distribution and distribution components, thermal management components, and safety protection components, and equipped with sensors related to safety monitoring. To facilitate spatial positioning and the implementation of the emergency response mechanism, the physical entity of the electrochemical energy storage system can form a mappable spatial coordinate system within the container. For example, a three-dimensional coordinate system can be established using the length, width, and height of the container, or a discrete grid coordinate system can be established using the arrangement and layer numbers of the battery modules. This allows subsequent temperature fields, gas concentration fields, and risk distributions to correspond to specific physical locations. At the same time, battery modules can be divided into clusters or racks to form several separable target module units, and actuators and control units can implement differentiated spraying, ventilation and isolation control for different target module units.

[0047] Multi-source sensing units are installed inside the energy storage container and arranged correspondingly to the battery modules. They are used to collect electrical parameters, temperature parameters, combustible gas concentration parameters, and cabin pressure parameters, and fuse the collected results to form a real-time state vector. To achieve the corresponding arrangement with the battery modules, the multi-source sensing units can set temperature monitoring points near each battery module or module cluster, set voltage and current monitoring points at key electrical connections of the battery modules, set combustible gas concentration monitoring points at locations in the cabin where gases are likely to accumulate, and set pressure monitoring points inside the cabin to reflect the pressure change trend. Electrical parameters may include individual cell voltage, module voltage, cluster voltage, current, insulation status, internal resistance, or related equivalent quantities; temperature parameters may include individual cell temperature, module surface temperature, cabin ambient temperature, or related gradient information; combustible gas concentration parameters may include hydrogen, carbon monoxide, volatile organic compounds, or comprehensive combustible gas indicators; cabin pressure parameters may reflect exhaust, pre-deflagration signs, or changes in ventilation status. To ensure the consistency of multi-source data, the multi-source sensing unit can perform time alignment, missing data processing, outlier removal and normalization on various types of sensor data, and can smooth short-term fluctuations through a sliding window to form a real-time state vector that can be used as a stable input for the inference model.

[0048] The real-time state vector can be represented as an input vector containing multidimensional features. It includes at least the real-time temperature T, voltage V, gas concentration C of each monitoring point at the current moment, and the rate of change characteristics of these key states. The rate of change can be used to reflect the dynamic trend of the system state. Optionally, the rate of change can be calculated using a sliding time window, and its mathematical expression can be written as:

[0049] , , ;

[0050] in, This indicates the time span corresponding to the sliding time window. , , These represent the changes in temperature, voltage, and gas concentration within the given time span, respectively. By incorporating the rate of change into the input vector, the extrapolation model can utilize not only the instantaneous state but also the state evolution trend, thus becoming more sensitive to early changes in the thermal runaway evolution path. In addition to the above features, the real-time state vector may also include the intra-chamber pressure P and its rate of change. This may include comprehensive features derived from electrical and temperature parameters, such as heat generation per unit time and abnormal deviation, to enhance the observability of thermal runaway, gas generation, and combustion-detonation chains. The selection and combination of these features can be adjusted according to battery type, cabin structure, and sensor arrangement, but their common goal is to form a multi-dimensional input vector that can characterize the current thermal safety status of the energy storage system.

[0051] The digital twin simulation unit is communicatively connected to the multi-source sensing unit to receive real-time state vectors and output predictions of thermal runaway evolution paths within a preset time window. The digital twin simulation unit may include a high-fidelity full-chain simulation module and a probabilistic surrogate simulation module. The high-fidelity full-chain simulation module is used to offline construct a high-dimensional physical model covering the entire chain of thermal runaway, gas production, and combustion explosion, and performs sampling simulations within a preset fault parameter space to generate a training sample set. The probabilistic surrogate simulation module completes training based on the training sample set and, during runtime, infers the real-time state vectors to output predictions of the thermal runaway evolution paths and their corresponding uncertainty metrics.

[0052] The high-fidelity full-chain simulation module can pre-set coupled physical models of battery thermal runaway, gas generation and diffusion, and combustion and explosion risk evolution. These models can include internal battery heat source terms, heat conduction and convection terms, in-cabin gas generation and transport terms, ventilation boundary condition terms, and combustion and explosion criterion terms. Given input conditions and fault parameters, it outputs a spatiotemporal sequence of temperature field evolution, gas concentration field evolution, and combustion and explosion risk-related indicators over a future period. When sampling the above parameter space for simulation, random sampling, hierarchical sampling, or Latin hypercube methods can be used to cover diverse operating conditions, thereby obtaining a training sample set containing input conditions and corresponding output evolution sequences. Each sample in the training sample set can be represented as an input vector. With output vector The pairings, among which Corresponding to a certain initial simulation condition or a certain moment of state characteristics This corresponds to the evolutionary sequence within a preset time window starting from this condition. In this way, the high-fidelity full-chain simulation module transforms the high-dimensional mechanism model, which is difficult to solve in real time, into sample data that can be learned by the probabilistic proxy inference module, thus laying the foundation for subsequent millisecond-level inference.

[0053] The probabilistic surrogate inference module can employ a multi-output Gaussian process regression model as the probabilistic surrogate model. Given the current state vector, it outputs an m×n dimensional vector consisting of the predicted values ​​of m key state variables for the next n time steps, and simultaneously outputs the corresponding uncertainty metric. To describe the organization of the multi-output prediction, the output vector can be denoted as... Alternatively, y can contain sequential representations of indicators such as temperature field distribution, key gas concentration distribution, and explosion risk probability over multiple future time steps. To facilitate the mapping of output to spatial location, the temperature field distribution can be organized according to several spatial sampling points after the cabin is gridded, the gas concentration distribution can also be organized according to the same spatial sampling points, and the explosion risk probability can be organized according to spatial sampling points or module units. This ensures that the output vector includes not only time dimensions but also spatial and variable dimensions. Through this organization, the prediction of thermal runaway evolution path is no longer limited to single-point temperature or single-point concentration but can reflect the direction, speed, and key risk locations of the disaster in space.

[0054] The prior mean function of a Gaussian process regression model can be expressed in constant form as follows:

[0055] ;

[0056] in, The covariance function can be estimated from the mean of the training data, which simplifies the model and improves numerical stability. A composite kernel function can be used to characterize the correlation between different input features and improve robustness to noise. Optionally, the composite kernel function may include at least a Matrn kernel component, a linear kernel component, and a noise kernel component. The Matrn kernel component is used to capture the nonlinear correlation between state variables, the linear kernel component is used to characterize the approximately linear trend of changes with the input, and the noise kernel component is used to model observed noise or unmodeled disturbances. For ease of explanation, the composite form including the Matrn kernel and the noise kernel can be given first, which can be written as:

[0057] ;

[0058] in, This represents the signal variance, used to control the output amplitude of the function. This is a length scale vector used to control the range of influence of different input dimensions on similarity; For the Matern 5 / 2 kernel function; For noise variance; The Dirac function is used to... and When the parameters are the same or similar, a noise term is introduced. The above expression illustrates the combination of the nonlinear correlation kernel and the noise kernel. When a linear kernel component is introduced, the composite kernel function can be further written as the sum of the Matrn kernel component, the linear kernel component, and the noise kernel component, thus enabling the model to simultaneously possess nonlinear fitting capability and linear trend fitting capability. The linear kernel component can optionally take the form of the inner product of the input vector or a linear mapping form equivalent to the inner product, to match the approximately linear coupling trend of electrical parameters, temperature parameters, and gas concentration parameters under certain operating conditions.

[0059] The Matrn 5 / 2 kernel function can be written as:

[0060] Matern 5 / 2(r) = (1 + +5 )exp(− ), r = ||x - x'||;

[0061] Where r represents the Euclidean distance or the scale-normalized distance metric between input vectors. x and x' represent two different samples or the current input and the corresponding training samples, respectively. A 5D input vector. By employing the Matrn 5 / 2 kernel function, the model can capture nonlinear changes between input and output while maintaining a certain level of smoothness, and it also exhibits some robustness to input noise. Noise kernel components. Used for modeling observation noise or sensor noise, this term improves the model's adaptability to fluctuations in actual sensor data. To adapt the composite kernel function to specific data distributions, its hyperparameters can be optimized, and the hyperparameters can be expressed as:

[0062] ={ ,l, };

[0063] in, 、l and These correspond to the signal variance, length scale vector, and noise variance, respectively. The choice of [the appropriate parameter] determines the model's sensitivity to input similarity and its tolerance to noise.

[0064] The probabilistic surrogate inference module can optimize the hyperparameters of the composite kernel function by maximizing the marginal likelihood function. The logarithmic form of the marginal likelihood can be written as:

[0065] ;

[0066] in, The input matrix for the training samples, The output vector of the training samples or the expanded output matrix. The kernel matrix and its elements , Let be an n×n identity matrix, where n is the number of training samples. The above log-marginal likelihood comprehensively considers both the data fitting term and the model complexity term, achieving a balance between fitting accuracy and generalization ability. To solve for the optimal hyperparameters, the probabilistic surrogate inference module can use gradient descent or gradient ascent optimization strategies to update them. For example, the following update rules can be used:

[0067] ;

[0068] in, For learning rate, This indicates the number of iterations. Training can be terminated when the marginal likelihood change is less than a threshold, the maximum number of iterations is reached, or the gradient norm is lower than a set value. Through the above training process, the probabilistic surrogate inference module can obtain the kernel matrix correlation quantities and hyperparameter set that can be used for online prediction, thus preparing for rapid calculation in the real-time inference stage.

[0069] In the real-time simulation phase, given the current real-time state vector The future n-step output vector predicted by a multi-output Gaussian process regression model It can follow the following posterior distribution:

[0070] p( |X,y, )~N( , );

[0071] in, , which is the posterior mean vector, representing the most likely evolutionary path prediction result; Let be the posterior covariance matrix, used to quantify the uncertainty of the prediction. The posterior mean can be written as:

[0072] ;

[0073] The posterior covariance can be written as:

[0074] =K( , )- K(X, );

[0075] Where X and y are the input matrix and output vector of the training sample, respectively. To train the kernel matrix between inputs, To train the kernel vector or kernel matrix between the input and the current input, K( , ) represents the currently input kernel term. The above expression indicates that... Essentially, it is the output of training samples. The weighted combination is determined by the similarity of the inputs; the closer the inputs are to the training samples, the more the prediction depends on the corresponding samples. This reflects the degree of coverage of the current input relative to the distribution of the training data. When the difference from the training samples is large, An increase indicates increased uncertainty in the forecast. The main diagonal elements can serve as a measure of uncertainty at each prediction point, while the off-diagonal elements can reflect the uncertainty coupling relationships between different output variables or different spatial points. Through the output... and The system can not only provide the most likely outcome of the future evolutionary path, but also the credibility of that outcome, thus providing a basis for subsequent decision-making.

[0076] To achieve millisecond-level online inference, the probabilistic surrogate inference module can utilize the data obtained during the training phase. The results of the correlation matrix factorization are cached, and matrix-vector multiplication is used for fast calculation in real-time prediction. and Real-time prediction may include steps such as input preprocessing, kernel vector calculation, posterior distribution solution, and output decoding. Input preprocessing is used to convert the raw state data acquired in real time into a d-dimensional input vector consistent with that of the training phase. Kernel vector computation is used to calculate The posterior distribution is solved to obtain and The output decoder is used to... Mapped to specific physical quantity prediction sequences and from Extracting uncertainty metrics. Through the above design, the digital twin simulation unit can quickly output the spatiotemporal sequence of temperature field, the spatiotemporal sequence of key gas concentration, and the probability distribution of combustion and explosion risks within a preset future time window without directly solving high-dimensional physical models, and simultaneously output the corresponding uncertainty metrics, thereby forming simulation results that can be visualized and used for decision-making.

[0077] Furthermore, the visualized emergency decision-making unit is communicatively connected to the digital twin simulation unit for performing three-dimensional dynamic visualization rendering of the predicted thermal runaway evolution path and generating spatially positioned emergency control instructions based on preset decision rules. The visualized emergency decision-making unit may include a three-dimensional rendering module and a decision rule engine. The three-dimensional rendering module is used to decode the predicted mean into three-dimensional dynamic visualization information and map the uncertainty measure into confidence display information associated with the three-dimensional dynamic visualization information. The decision rule engine is used to determine high-confidence risk areas based on the probability distribution of combustion and explosion risks or their corresponding risk indicators and uncertainty measures, and generate spatially positioned emergency control instructions based on the high-confidence risk areas.

[0078] The 3D rendering module can map the spatiotemporal sequences of the temperature field and the gas concentration field into 3D spatial mesh data based on the cabin's spatial coordinate system, and dynamically render and display them along the time axis. Optionally, the 3D rendering module can map the temperature field into a 3D thermal distribution, the gas concentration field into a concentration cloud map or isosurface, and can overlay the probability distribution of combustion and explosion risks into a risk thermal layer, thus presenting a multi-dimensional risk situation within the same spatial coordinate framework. Simultaneously, the 3D rendering module can... The prediction variance corresponding to the main diagonal is mapped to confidence level information, such as color depth, blur transparency, or confidence level bars, and overlaid with the most likely evolution path. This overlay method allows operations personnel to see both the magnitude of risk and the confidence level of the prediction simultaneously, preventing overly aggressive actions based solely on high-risk but high-uncertainty predictions, and avoiding overlooking critical areas with high risk but low uncertainty.

[0079] The decision rule engine can jointly determine high-confidence risk zones based on at least the following conditions: The first condition is that the probability distribution of the explosion risk or its corresponding risk indicator meets a first threshold condition, meaning the risk has reached a level requiring intervention; the second condition is that the uncertainty measure meets a second threshold condition, meaning the prediction variance is below a preset upper limit or the confidence level is above a preset lower limit. By simultaneously satisfying the risk and confidence conditions, the decision rule engine can identify areas with "high risk and reliable prediction" as high-confidence risk zones. To achieve spatial positioning, high-confidence risk zones can correspond to several spatial units in the cabin grid or to one or more battery module units, and can further combine spatial connectivity to cluster discrete points or merge regions to form an executable set of target areas. For areas with high risk but high uncertainty, the decision rule engine can choose to generate instructions to strengthen monitoring, improve ventilation, or implement conservative prevention; for areas with high risk and low uncertainty, the decision rule engine can prioritize generating more targeted instructions such as directional spraying and rapid isolation, thereby achieving a joint trade-off between risk and uncertainty.

[0080] The visualized emergency decision-making unit may also include a strategy library. The decision rule engine matches and generates instruction combinations from the strategy library based on the physical location and hazard type of high-confidence risk areas. Hazard types can be determined based on the variable characteristics of the predicted output, such as the risk of thermal runaway expansion characterized by a rapid rise in temperature and local accumulation, the risk of flammable gas accumulation characterized by a rapid rise in gas concentration and reaching the flammability threshold, and the risk of combustion and explosion characterized by a risk probability distribution showing a critical trend of deflagration. The strategy library can pre-set emergency action templates for different hazard types and locations. The instruction combination must include at least one of the following:

[0081] Directional spray coordinates or nozzle number, ventilation intensity curve parameters, and target module isolation markings.

[0082] Directional spray coordinates or nozzle numbers are used to spray extinguishing or cooling media to locations corresponding to high-confidence risk areas; ventilation intensity curve parameters are used to control the ventilation intensity of fans or dampers over time, thereby achieving coordinated exhaust, dilution and cooling at different stages; target module isolation identifiers are used to indicate or trigger electrical, thermal or mechanical isolation of specific module units to block the thermal runaway propagation chain.

[0083] By introducing a strategy library into the system, instruction generation is no longer a single action, but can form a combination of multiple actions and stages to match the temporal characteristics of thermal runaway evolution.

[0084] The actuators and control units may include sprinkler actuators, ventilation actuators, isolation actuators, and their controllers. Sprinkler actuators may include sprinkler heads, piping, and pump / valve assemblies; ventilation actuators may include fans, dampers, and exhaust duct assemblies; and isolation actuators may include contactors, circuit breakers, or disconnect switches, and their drive components. After receiving spatial positioning commands from the visual emergency decision-making unit, the actuators and control units can parse the commands into executable control quantities, such as sprinkler head numbers and sprinkler duration, fan speed changes over time, and the action sequence of the disconnect switch. Optionally, the actuators and control units can feed back the execution results to the multi-source sensing unit and the digital twin simulation unit, enabling the simulation model to consider the state changes caused by sprinkler cooling, ventilation dilution, and isolation cutoff in subsequent predictions, thereby forming a more realistic closed-loop simulation and decision-making process.

[0085] The system may also include an online learning module to maintain the predictive accuracy and adaptability of the probabilistic proxy inference module during long-term system operation. The online learning module continuously monitors the deviation between the actual monitoring data and the prediction results of the probabilistic proxy inference module. When a persistent deviation occurs within a preset time window and exceeds an adaptive threshold, a model update process is triggered. To quantify the prediction deviation, the prediction mean at each time step can be denoted as... The actual evolutionary observations at the corresponding time are denoted as The deviation can then be written as:

[0086] Deviation = ;

[0087] in, This represents the L2 norm, used to comprehensively measure the overall deviation across multiple outputs, spatial points, and variable dimensions. To avoid false triggering due to single-point noise, the online learning module can use a sliding window to statistically average the deviation, for example, it can be written as:

[0088] Average deviation = Here, a time step represents the number of discrete time steps contained within the sliding window, or the equivalent length of a statistical window. The initial threshold can be set based on the prediction error distribution of the training set and can be adaptively adjusted according to the statistical characteristics of recent deviations. For example, the threshold can be updated based on the mean and variance of the average deviation, ensuring that the system maintains reasonable trigger sensitivity during battery aging, environmental changes, or operational condition transitions. When the average deviation of multiple consecutive time steps exceeds the current threshold and the duration is greater than the preset window, the online learning module can trigger the model update process.

[0089] The model update process can begin by determining the incremental sample construction method, collecting real-time state and actual evolution data pairs during the update trigger period, and forming a new set of incremental data pairs {( The online learning module can then select an update strategy, which may include a Bayesian update strategy or a sliding window retraining strategy. The Bayesian update strategy is suitable for scenarios with small incremental data volumes or where maintaining model stability is crucial. It uses new data to supplement prior information and obtains updated hyperparameters by maximizing marginal likelihood. The Bayesian update strategy can be represented as follows: ;in, and This represents the input and output set of the incremental sample. This represents the updated set of hyperparameters. Through Bayesian updates, the system can integrate new and old knowledge and reduce the risk of catastrophic forgetting. The sliding window retraining strategy is suitable for scenarios with large incremental data volumes or significant changes in system characteristics. It maintains a fixed-size training data window; when new data is added, the oldest data is removed, keeping the window size unchanged, and the model is retrained using all data within the window. This allows for rapid adaptation to the latest system state and reduces the negative impact of historical data. To avoid system instability caused by frequent updates, the online learning module can also set a minimum update interval, executing updates only when the triggering condition is met and the minimum interval is exceeded. Multiple historical model versions can be retained to support rapid rollback. Validation set testing can be performed before updating to ensure the new model's performance is not lower than the old model, and large-scale retraining can be performed during system idle periods to avoid impacting real-time predictions.

[0090] To meet millisecond-level real-time computing and response requirements, multi-source sensing units and digital twin simulation units can be deployed in edge computing devices located locally within the energy storage container. These edge computing devices communicate with the visualized emergency decision-making unit to locally generate real-time state vectors, predict thermal runaway evolution paths, and generate and output emergency control commands. The edge computing devices can be industrial computers, embedded edge servers, or computing modules with acceleration capabilities. They can form low-latency communication links with sensor networks and actuator controllers, thereby reducing response lag caused by cloud round-trip latency. The visualized emergency decision-making unit can be deployed locally on the edge computing device or on a monitoring terminal, accessing the edge computing device via a local area network or dedicated link to achieve 3D situational awareness and command review. For scenarios requiring higher reliability, the visualized emergency decision-making unit can also have local offline operation capabilities, enabling rendering and decision output even when the external network is unavailable.

[0091] Based on the above system structure, during operation, the multi-source sensing unit continuously collects and integrates electrical parameters, temperature parameters, combustible gas concentration parameters, and cabin pressure parameters to form a real-time state vector, which is then sent to the digital twin simulation unit. The digital twin simulation unit uses the probabilistic proxy simulation module to output the spatiotemporal prediction results of the temperature field, gas concentration field, and explosion risk probability within a future preset time window, and outputs the corresponding uncertainty measure. The visualization emergency decision-making unit performs three-dimensional dynamic rendering of the predicted mean and maps the uncertainty measure to confidence display information. At the same time, the decision rule engine determines the high-confidence risk zone based on the risk threshold and uncertainty threshold and generates spatially positioned emergency prevention and control instructions. When necessary, it combines with the strategy library to form instruction combinations such as spraying, ventilation, and isolation. The actuators and control units implement corresponding actions according to the instruction combinations and can provide feedback on the execution status, thereby realizing proactive intervention in the thermal runaway process.

[0092] In summary, by integrating multi-source sensing, probabilistic inference, visualization, and rule-based decision-making into a closed-loop chain, this system enables maintenance personnel or automatic control systems to obtain spatial path information and credibility information of future evolution in the early stages of thermal runaway. Based on this, they can output emergency actions that can be located to specific spatial locations or specific module units, thereby forming a clearer target orientation and a more controllable action sequence in emergency response.

[0093] Example 2:

[0094] This embodiment provides a real-time emergency control method for electrochemical energy storage systems based on digital twins, which can be applied to the above-mentioned system architecture. The method process may include a model building stage, a real-time monitoring and prediction stage, and a visualization and decision-making stage.

[0095] During the model building phase, a high-fidelity full-chain simulation module can perform sampling simulations within a preset fault parameter space, generating a training sample set containing input conditions and output evolution sequences. Input conditions may include the initial simulation state, fault location and intensity, ventilation boundary conditions, and ambient temperature, while the output evolution sequence may include temperature fields, gas concentration fields, and risk probability sequences for multiple future time steps. Subsequently, a probabilistic surrogate inference module is trained based on the training sample set, enabling it to output predictions of the thermal runaway evolution path within a preset time window, along with corresponding uncertainty metrics. The aforementioned methods can be used during the training process. Prior mean setting, composite kernel function The construction of the hyperparameter θ and the marginal likelihood optimization are carried out. The appropriate set of hyperparameters and pre-computed matrices can be obtained through log marginal likelihood and gradient update rules for subsequent real-time derivation.

[0096] During the real-time monitoring and prediction phase, multi-source sensing units can collect electrical parameters, temperature parameters, combustible gas concentration parameters, and cabin pressure parameters, and fuse them to form a real-time state vector. This may include T, V, C and their rates of change. , , Dynamic features are then identified. Subsequently, the real-time state vector is input to the probabilistic proxy inference module to obtain the spatiotemporal sequence of the temperature field, the spatiotemporal sequence of key gas concentrations, and the probability distribution of combustion and explosion risks, as well as to obtain an uncertainty measure. This prediction can be based on the posterior distribution p( |X,y, )~N( , ) calculate, where As the output of the most likely evolutionary path prediction result, Used to output prediction variance and covariance information. For easier subsequent decision-making, it can be... The main diagonal elements are extracted as the prediction variance of each spatial point or each module unit, and output as an uncertainty measure. At the same time, the risk probability distribution can be output by spatial grid or by module unit to form risk distribution layer data.

[0097] During the visualization and decision-making phase, the visualization emergency decision-making unit can perform 3D dynamic rendering of the predicted mean and map uncertainty measures into confidence display information. The 3D dynamic rendering can scroll along a timeline to display changes in temperature and gas concentration fields over multiple future time steps, and overlay the probability of combustion and explosion risks as thermal layers or isosurfaces to create an intuitive presentation of future evolution. Confidence display information can be linked to risk presentation; for example, locations with high risk and low variance can be presented more clearly, brighter, or with higher saturation, while locations with high risk but high variance can be presented more vaguely or transparently, thus indicating differences in prediction reliability. Subsequently, the decision rule engine determines high-confidence risk areas based on a first threshold condition and a second threshold condition, and generates and outputs spatially positioned emergency prevention and control instructions. The first threshold condition can constrain risk probability or risk indicators, and the second threshold condition can constrain prediction variance or confidence level; their combined use can select priority intervention areas. If the system is configured with a strategy library, the decision rule engine can match and generate instruction combinations from the strategy library based on the physical location and hazard type of high-confidence risk areas. These instruction combinations can include at least one or more of the following: directional sprinkler coordinates or sprinkler head numbers, ventilation intensity curve parameters, and target module isolation identifiers. Multiple instructions can be time-programmed according to the predicted evolution sequence, enabling sprinkler, ventilation, and isolation actions to coordinate at different stages. Finally, the actuators and control units execute sprinkler, ventilation, and isolation actions according to the instruction combinations and can feed back the execution status to the monitoring and simulation link so that subsequent predictions can take into account the status changes caused by the executed actions.

[0098] The method may also include online learning and adaptive update processes to maintain long-term predictive accuracy. These processes can calculate prediction biases during real-time monitoring. The system calculates the average deviation of the sliding window. When the average deviation continuously exceeds an adaptive threshold within a preset time window, an update is triggered. The new real-time state and the actual evolution data are added to the incremental sample set. The hyperparameters or training sample set of the probabilistic proxy inference module are then updated through Bayesian updates or sliding window retraining. Bayesian updates can be achieved through... The updated hyperparameters are determined in the form of sliding window retraining, which can achieve rapid adaptation by maintaining a fixed-size data window and retraining the model. At the same time, a minimum update interval and version rollback mechanism can be set to ensure system stability.

[0099] In summary, this method combines the mechanistic data generated by high-fidelity simulation with the rapid extrapolation capability of the probabilistic surrogate model, enabling the system to output the spatial path and credibility information of future thermal runaway evolution in an interpretable manner during actual operation. Furthermore, this information is converted into emergency instructions for spatial positioning, thus forming an executable closed loop from prediction to action in the handling process. The model can also be adaptively adjusted as the system state changes through online updates.

Claims

1. A real-time emergency prevention and control system for an electrochemical energy storage system based on digital twinning, characterized in that, The system comprises: a multi-source perception unit, a digital twin inference unit, and a visualized emergency decision-making unit; The multi-source perception unit is arranged in the energy storage container and corresponds to the battery module, is used for collecting electrical parameters, temperature parameters, flammable gas concentration parameters, and cabin pressure parameters, and fuses the collection results to form a real-time state vector; The digital twin inference unit is in communication connection with the multi-source perception unit, is used for receiving the real-time state vector and outputting a thermal runaway evolution path prediction in a future preset time window; The visualized emergency decision-making unit is in communication connection with the digital twin inference unit, is used for performing three-dimensional dynamic visualized rendering on the thermal runaway evolution path prediction, and generating a spatially positioned emergency prevention and control instruction based on a preset decision-making rule.

2. The real-time emergency prevention and control system for electrochemical energy storage systems based on digital twinning according to claim 1, characterized in that, The digital twin inference unit comprises a high-fidelity full-chain simulation module and a probabilistic agent inference module; The high-fidelity full-chain simulation module preloads a full-chain high-dimensional physical model of battery thermal runaway-gas production-explosion, and performs sampling simulation in a preset fault parameter space to generate a training sample set, the training sample set at least comprising input conditions and corresponding output evolution sequences; The probabilistic agent inference module is trained based on the training sample set, and infers the real-time state vector at runtime to output the thermal runaway evolution path prediction.

3. The real-time emergency prevention and control system for electrochemical energy storage systems based on digital twinning according to claim 2, characterized in that, The thermal runaway evolution path prediction output by the probabilistic agent inference module at least comprises a temperature field time-space sequence, a key gas concentration time-space sequence, and an explosion risk probability distribution; And the probabilistic agent inference module also outputs an uncertainty measure corresponding to the thermal runaway evolution path prediction.

4. The real-time emergency prevention and control system for electrochemical energy storage systems based on digital twinning according to claim 2, characterized in that, The probabilistic agent inference module adopts a multi-output Gaussian process regression model; The multi-output Gaussian process regression model adopts a composite kernel function, the composite kernel function at least comprising a Matern kernel component, a linear kernel component, and a noise kernel component; And the hyperparameters of the composite kernel function are optimized by maximizing the marginal likelihood function.

5. The real-time emergency prevention and control system for electrochemical energy storage systems based on digital twinning according to claim 3, characterized in that, The visualized emergency decision-making unit comprises a three-dimensional rendering module, which is used for decoding a predicted mean in the thermal runaway evolution path prediction into three-dimensional dynamic visualized information, and mapping an uncertainty measure into credibility display information associated with the three-dimensional dynamic visualized information.

6. The real-time emergency prevention and control system for electrochemical energy storage systems based on digital twinning according to claim 1, characterized in that, The visualized emergency decision-making unit comprises a decision-making rule engine, which is used for jointly determining a high-confidence risk area based on at least the following conditions: (a) the explosion risk probability distribution or a corresponding risk indicator thereof satisfies a first threshold condition; (b) the uncertainty measure satisfies a second threshold condition; And generating a spatially positioned emergency prevention and control instruction based on the high-confidence risk area.

7. The real-time emergency prevention and control system for electrochemical energy storage systems based on digital twinning according to claim 6, characterized in that, The visualized emergency decision-making unit further comprises a strategy library, and the decision-making rule engine matches and generates an instruction combination from the strategy library according to the physical location and the danger type of the high-confidence risk area, the instruction combination at least comprising one of the following: directed spraying coordinates or nozzle number, ventilation intensity curve parameters, and target module isolation identifier. 8.The real-time emergency prevention and control system based on digital twinning of electrochemical energy storage system according to claim 1, wherein, The system further comprises an online learning module, which is used for: When the actual monitoring data and the prediction results of the probability agent inference module continuously deviate within a preset time window and the deviation exceeds an adaptive threshold, a model updating process is triggered; The new real-time state-actual evolution data pair is taken as an incremental sample, and the hyperparameters or training sample set of the probability agent inference module are updated through Bayesian updating or sliding window retraining. 9.The real-time emergency prevention and control system based on digital twinning of electrochemical energy storage system according to claim 1, wherein, The multi-source perception unit and the digital twin inference unit are deployed in an edge computing device local to the energy storage container, and the edge computing device is in communication connection with the visual emergency decision unit to locally complete formation of the real-time state vector, output of the thermal runaway evolution path prediction, and generation and output of the emergency prevention and control instructions.

10. A real-time emergency prevention and control method for an electrochemical energy storage system based on digital twinning, characterized in that, The method is applied to the system of any one of claims 1-9, and the method comprises: S1. Based on the high-fidelity full-chain simulation module, sampling simulation is performed in a preset fault parameter space to generate a training sample set containing input conditions and output evolution sequences; S2. The probability agent inference module is trained based on the training sample set, so that it can output thermal runaway evolution path prediction within a preset time window in the future and the uncertainty measure corresponding thereto; S3. The multi-source perception unit collects electrical parameters, temperature parameters, flammable gas concentration parameters, and cabin pressure parameters and fuses them to form a real-time state vector; S4. The real-time state vector is input into the probability agent inference module to obtain a temperature field time-space sequence, a key gas concentration time-space sequence, and a combustion risk probability distribution, and to obtain an uncertainty measure; S5. The visual emergency decision unit performs three-dimensional dynamic rendering on the prediction mean, and maps the uncertainty measure to credibility display information; S6. The decision rule engine determines a high-confidence risk area based on the joint of the first threshold condition and the second threshold condition, generates a spatially positioned emergency prevention and control instruction, and outputs the same.

Images