Energy storage system full life cycle digital twin safety assessment and prediction method and system

Abstract

The present application relates to the technical field of energy storage safety evaluation, and specifically discloses a kind of energy storage system full life cycle digital twin safety evaluation and prediction method and system, the method includes the following steps: obtaining energy storage side full life cycle data and power grid side collaborative data, obtains data quality grade by three-dimensional quality grading, and the energy storage side full life cycle data and the power grid side collaborative data are weighted and fused based on cross-domain coupling coefficient, generate five-dimensional risk collaborative portrait.The present application breaks the information island of energy storage side and power grid side in the actual operation scene of energy storage power station deeply participating in power grid dispatching, realizes the accurate tracing and pre-prevention of the internal physical state evolution of energy storage system under external power grid instruction disturbance.

Description

Technical Field

[0001] This invention relates to the field of energy storage safety assessment technology, specifically to a digital twin safety assessment and prediction method and system for the entire life cycle of an energy storage system. Background Technology

[0002] Existing safety assessments of energy storage systems typically monitor them as independent physical entities, relying primarily on the single-dimensional operational history and real-time data collected from the energy storage devices, and employing fixed safety thresholds for risk warnings. At the model building level, existing technologies mostly use simple physical mechanism models or black-box big data-driven models, passively determining whether the system has safety hazards or abnormal faults by comparing characterizing parameters such as battery voltage, current, and temperature with static preset values.

[0003] However, as energy storage systems become deeply involved in grid frequency regulation, backup support, and other coordinated dispatching tasks, the internal physical states of the system, including thermal, electrical, and aging states, are subject to strong disturbances from dynamic grid commands, leading to the following problems:

[0004] If a conservative static safety threshold is adopted, the energy storage system will frequently interrupt operation due to excessive defense, and will be unable to respond to the high-frequency coordination requirements of the power grid.

[0005] However, if the early warning boundaries are relaxed in order to meet the grid commands, the lack of precise quantification of the cross-domain coupling mechanism between the underlying characteristics of energy storage and the characteristics of grid dispatch makes it easy for unknown risks to accumulate and lead to catastrophic accidents such as thermal runaway. Summary of the Invention

[0006] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a digital twin safety assessment and prediction method and system for the entire lifecycle of energy storage systems, thereby improving the operational safety of energy storage systems throughout their entire lifecycle.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for digital twin safety assessment and prediction of an energy storage system throughout its entire lifecycle, comprising the following steps:

[0008] Acquire full lifecycle data of the energy storage side and collaborative data of the grid side, perform three-dimensional quality grading to obtain data quality level, and perform weighted fusion of the full lifecycle data of the energy storage side and the collaborative data of the grid side based on cross-domain coupling coefficient to generate a five-dimensional risk collaborative profile;

[0009] A five-field coupled physical mechanism model of heat-electricity-aging-insulation-equilibrium is constructed. The cross-domain coupling coefficient is introduced as a model parameter correction term to quantify the impact of coordinated actions on risk. A data-driven prediction model is generated by reinforcement learning training combined with preset virtual accident samples. The five-field coupled physical mechanism model and the data-driven prediction model are then fused to form a hybrid model.

[0010] The risk coupling coefficient between core risks is calculated by mutual information entropy. The early warning threshold is dynamically adjusted based on the three-dimensional mapping rule composed of the risk coupling coefficient, the data quality level and the collaborative requirement level. The hybrid model outputs multi-dimensional risk classification results and risk evolution prediction results.

[0011] Based on the multi-dimensional risk classification results, a preset three-dimensional cost model is established. The three-dimensional cost model is solved with the goal of minimizing costs to obtain a combined decision scheme and output a quantitatively interpretable early warning report that includes the power grid collaborative traceability link.

[0012] When the data quality confidence level is detected to be lower than the preset confidence threshold or the model prediction uncertainty is higher than the preset uncertainty threshold, the degraded mode is triggered to dynamically correct the warning threshold and the system power boundary. The combined decision-making scheme is executed based on the three-level decision-making authority allocation system of edge layer, cloud and grid side, and feedback instructions are received to form a decision closed loop.

[0013] To achieve the above objectives, a second aspect of the present invention proposes a digital twin safety assessment and prediction system for the entire lifecycle of an energy storage system, the system comprising:

[0014] The data acquisition and fusion module is used to acquire full life cycle data of the energy storage side and collaborative data of the grid side, perform three-dimensional quality classification to obtain data quality level, and perform weighted fusion of the full life cycle data of the energy storage side and the collaborative data of the grid side based on the cross-domain coupling coefficient to generate a five-dimensional risk collaborative profile.

[0015] The hybrid model construction module is used to construct a five-field coupled physical mechanism model of heat-electricity-aging-insulation-equilibrium. The cross-domain coupling coefficient is introduced as a model parameter correction term to quantify the impact of coordinated actions on risk. The module combines preset virtual accident samples to perform reinforcement learning training to generate a data-driven prediction model. The five-field coupled physical mechanism model and the data-driven prediction model are then fused to form a hybrid model.

[0016] The risk assessment and prediction module is used to calculate the risk coupling coefficient between core risks through mutual information entropy, dynamically adjust the early warning threshold based on the three-dimensional mapping rule composed of the risk coupling coefficient, the data quality level and the collaborative requirement level, and output multi-dimensional risk classification results and risk evolution prediction results through the hybrid model.

[0017] The collaborative decision-making and early warning module is used to establish a preset three-dimensional cost model based on the multi-dimensional risk classification results, solve the three-dimensional cost model with cost minimization as the optimization objective to obtain a combined decision scheme, and output a quantitatively interpretable early warning report containing the power grid collaborative traceability link.

[0018] The degradation control and feedback closed-loop module is used to trigger a degradation mode to dynamically correct the warning threshold and system power boundary when the data quality confidence is detected to be lower than the preset confidence threshold or the model prediction uncertainty is higher than the preset uncertainty threshold. It also executes the combined decision-making scheme based on the three-level decision-making authority allocation system of edge layer, cloud and grid side, and receives feedback instructions to form a decision closed loop.

[0019] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for digital twin safety assessment and prediction of the entire life cycle of an energy storage system.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] The digital twin safety assessment and prediction method and system for the entire lifecycle of an energy storage system, as described in this invention, breaks down the information silos between the energy storage side and the grid side in actual operation scenarios where energy storage power stations are deeply involved in grid dispatch. This enables precise tracing and proactive prevention of the evolution of the internal physical state of the energy storage system under external grid command disturbances. Its effects are as follows:

[0022] By integrating multi-source quality grading data with a five-field coupled hybrid model, the system can perceive the external grid coordination needs, internal risk coupling status, and underlying data confidence in real time. It also adaptively adjusts the early warning threshold using three-dimensional mapping rules, thereby maximizing the dynamic support capability of energy storage assets for the grid while ensuring the core physical security of the system. At the same time, relying on the established three-dimensional cost model and multi-level degradation control mechanism, even when encountering severe operating conditions such as sensor data distortion or a surge in prediction uncertainty, the energy storage system can still avoid a one-size-fits-all forced shutdown of the entire station. Through a three-level decision-making system, it executes a dynamic degradation closed loop that balances minimizing economic losses and core physical defenses, effectively solving the engineering pain point of being unable to balance system security, grid coordination, and operational economic benefits in complex grid interaction environments. Attached Figure Description

[0023] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0024] Figure 1 This is a flowchart illustrating the digital twin safety assessment and prediction method for the entire lifecycle of an energy storage system provided by this invention.

[0025] Figure 2 This is a graph showing the nonlinear temperature rise response of a battery cell under high-frequency grid command disturbances in the digital twin safety assessment and prediction method for the entire life cycle of an energy storage system provided by this invention.

[0026] Figure 3 This is a graph showing the nonlinear temperature rise response of a battery cell under high-frequency grid command disturbances in the digital twin safety assessment and prediction method for the entire life cycle of an energy storage system provided by this invention.

[0027] Figure 4 This invention provides an EEMD intrinsic mode decomposition and baseline drift bitmap of the sensor's original noise signal in the digital twin safety assessment and prediction method for the entire life cycle of an energy storage system.

[0028] Figure 5 This is a convergence curve of the training loss and domain discrimination probability of the adversarial network with the introduction of the inverted gradient layer in the digital twin security assessment and prediction method for the entire life cycle of energy storage systems provided by this invention;

[0029] Figure 6 This invention provides a two-dimensional mapping heatmap of spatial uncertainty and data quality at the cluster level of energy storage power stations in the digital twin safety assessment and prediction method for the entire life cycle of energy storage systems.

[0030] Figure 7 This is a schematic diagram illustrating the implementation of the digital twin safety assessment and prediction system for the entire lifecycle of energy storage systems provided by this invention.

[0031] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0032] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0033] The following description, with reference to the accompanying drawings, outlines a method, system, and electronic device for full lifecycle digital twin safety assessment and prediction of energy storage systems according to embodiments of the present invention.

[0034] Example 1:

[0035] This embodiment provides a digital twin-based safety assessment and prediction method for the entire lifecycle of an energy storage system. This method relies on a digital twin system architecture that includes a data acquisition layer, a model calculation layer, a decision control layer, and a feedback execution layer. In practical industrial applications, this method is primarily deployed in large-scale energy storage power plants at the megawatt-scale or gigawatt-hour level. These power plants typically include a battery management system, an energy management system, an energy storage converter, and a liquid-cooled or air-cooled thermal management system. Furthermore, the control system of the energy storage power plant maintains real-time data interaction with the grid-side dispatch center via industrial standard communication protocols (such as IEC61850 or Modbus TCP).

[0036] The method in this embodiment specifically includes the following steps:

[0037] Phase 1: Collection, evaluation and cross-domain fusion of multi-source heterogeneous data.

[0038] Specifically, the method in this embodiment first performs the following steps: acquiring full life cycle data of the energy storage side and collaborative data of the grid side, performing three-dimensional quality grading to obtain data quality level, and weighting and fusing the full life cycle data of the energy storage side and the collaborative data of the grid side based on the cross-domain coupling coefficient to generate a five-dimensional risk collaborative profile.

[0039] For example, acquiring full lifecycle data for the energy storage side includes collecting real-time voltage, real-time current, tab temperature, surface temperature, and estimated internal resistance values ​​at each level through a high-precision sensor network distributed across individual battery cells, battery modules, and battery clusters. Simultaneously, it involves retrieving historical charge-discharge cycle counts, historical fault records, and maintenance logs from the energy storage system's system logs. Acquiring grid-side collaborative data includes receiving automatic generation control commands, automatic voltage control commands, real-time grid frequency deviation data, bus voltage transient fluctuation data, and grid short-circuit fault alarm signals issued by the grid dispatch center.

[0040] Specifically, the steps of performing three-dimensional quality grading to obtain data quality levels and weighted fusion based on cross-domain coupling coefficients include: constructing a first evaluation model for the full life cycle data of the energy storage side, including completeness, accuracy, and consistency; constructing a second evaluation model for the grid-side collaborative data, including real-time performance, accuracy, and continuity; and outputting the corresponding data quality levels respectively; calculating the mutual information entropy between the grid-side collaborative data and the energy storage risk characteristics; and using the mutual information entropy as the cross-domain coupling coefficient for the weighted fusion.

[0041] In practical applications, the integrity index in the first evaluation model is used to measure the proportion of valid data points from energy storage-side sensors within a specific sampling period to avoid data interruptions caused by communication packet loss; the accuracy index reflects the degree of approximation between sensor sample values ​​and actual physical quantities; and the consistency index is used to determine the dispersion of feedback data distribution from different individual sensors within the same battery cluster. The real-time index in the second evaluation model measures the transmission delay time from the dispatch center to the energy storage field control unit; the accuracy index checks the checksum of the power grid data packets to eliminate transmission errors; and the continuity index evaluates the temporal coherence of the power grid coordination commands.

[0042] For example, the formula for calculating the mutual information entropy algorithm for cross-domain coupling coefficients is as follows:

[0043] ;

[0044] in, Represents the cross-domain coupling coefficient; Represents the set of data variables for grid-side coordination; Represents the set of variables characteristic of energy storage risks; This represents a single data sample within the set of grid-side collaborative data variables; This represents a single feature sample in the set of energy storage risk characteristic variables; This represents the joint probability distribution density of the data sample and the feature sample; This represents the marginal probability density of the grid-side collaborative data samples; This represents the marginal probability density of energy storage risk characteristic samples.

[0045] The practical engineering significance of this step lies in the fact that by introducing cross-domain coupling coefficients for weighted fusion, the system can filter out poor-quality data containing high-amplitude measurement noise, and map the data scattered in the grid dispatch domain and energy storage physical domain to a unified high-dimensional feature space, thereby constructing a five-dimensional risk collaborative profile that includes five dimensions: thermodynamic state, electrical state, aging state, grid support response load, and system availability.

[0046] Phase 2: Hybrid construction of the five-field coupled physical mechanism model and the data-driven prediction model.

[0047] Specifically, the method involves the following steps: constructing a five-field coupled physical mechanism model of heat, electricity, aging, insulation, and equilibrium; introducing the cross-domain coupling coefficient as a model parameter correction term to quantify the impact of coordinated actions on risk; combining preset virtual accident samples for reinforcement learning training to generate a data-driven prediction model; and merging the five-field coupled physical mechanism model with the data-driven prediction model to form a hybrid model.

[0048] For example, the five-field coupling physical mechanism model is a white-box mathematical model of the energy storage cell and its system-level topology. The thermal field model follows the law of conservation of energy and Fourier's law of heat conduction to calculate the Joule heating and reversible entropy change heating of the cell; the electric field model uses a second-order equivalent circuit model to simulate the polarization response of the battery; the aging field model describes the thickening effect of the solid electrolyte phase interface film based on the Arrhenius equation; the insulation field model calculates the dynamic decay of the system's common-mode impedance; and the equilibrium field model characterizes the charge transfer path loss in the active or passive equalization circuit.

[0049] Specifically, the cross-domain coupling coefficient is introduced as a model parameter correction term to quantify the impact of coordinated actions on risk. This includes: integrating grid power command and frequency deviation parameters into the heat transfer equation and equilibrium energy transfer model of the five-field coupled physical mechanism model; calculating the temperature rise increment caused by grid coordinated actions, where the temperature rise increment is equal to the continuous product of a preset proportional coefficient, energy storage power adjustment, and the cross-domain coupling coefficient. The calculation formula is as follows:

[0050] ;

[0051] in, This indicates the increase in temperature due to coordinated operation of the power grid; This indicates a preset proportional coefficient, which is determined by calibrating the thermal capacity characteristics of the energy storage system and the specific heat capacity of the cooling medium. This represents the energy storage power adjustment calculated based on the grid frequency deviation parameter; This represents the cross-domain coupling coefficient calculated in the preceding steps.

[0052] like Figure 2 This is a graph showing the nonlinear temperature rise response of individual battery cells under high-frequency grid command disturbances. The graph is divided into two sub-graphs, which intuitively show the differences in the actual thermal response of the energy storage system when participating in high-frequency grid regulation tasks.

[0053] The horizontal axis of the upper subgraph represents time in seconds, and the vertical axis represents power adjustment in megawatts. The gray waveform curve in the figure reflects the drastic step and reversal disturbances in the power commands frequently issued by the power grid dispatch center between -8 MW and +8 MW within the operating range of 0 to 600 seconds.

[0054] The lower subplot also uses time units (seconds) on the horizontal axis and temperature rise units (degrees Celsius) on the vertical axis, comparing the temperature evolution trajectories of two different evaluation models. The blue curve in the figure represents the calculation result of the traditional mechanistic model. Because it fails to fully quantify the additional thermal shock caused by high-frequency fluctuations, its predicted temperature rise curve is relatively flat, rising only to about 35 degrees Celsius at 600 seconds.

[0055] The red curve in the figure represents the calculation results of the hybrid model of the present invention. Since the cross-domain coupling coefficient is introduced as a parameter correction term in the scheme, the model can effectively capture the nonlinear Joule heating and entropy change heat accumulation effect of battery cells under frequent charge and discharge switching.

[0056] As can be seen from the waveform transformation of the red curve, each sharp power jump results in a significant surge in the temperature rise slope on the curve, ultimately predicting a true peak temperature rise of approximately 42 degrees Celsius at 600 seconds. This data comparison objectively demonstrates that the proposed solution can effectively reduce the risk of missed thermal runaway warnings by traditional models under complex power grid interaction conditions, providing a reliable data foundation and judgment basis for early intervention and dynamic adjustment of warning thresholds in liquid cooling systems.

[0057] In practical operation, when the power grid is under high-frequency regulation, the energy storage system needs to frequently switch between charging and discharging states to respond to commands. Severe fluctuations. Traditional physical models often struggle to accurately estimate the thermal shock caused by such frequent power reversals. However, the formula above, by introducing a coupling coefficient from the grid side, can accurately feedforward calculate the nonlinear temperature rise boundary inside the battery directly triggered by grid actions, thus prompting the pre-cooling action of the thermal management system in advance.

[0058] Specifically, the data-driven prediction model is generated by training reinforcement learning based on preset virtual accident samples. This includes: inputting preset boundary condition parameters into the five-field coupling physical mechanism model to simulate risk evolution and generate the virtual accident samples containing the time series of power grid parameters, risk characteristic trajectories, and changes in cross-domain coupling coefficients; after validating the effectiveness of the virtual accident samples, splicing the virtual accident samples with real collaborative scenario accident samples to construct a training set, and training the neural network using a few-shot reinforcement learning algorithm to obtain the data-driven prediction model that meets the preset collaborative requirement objective.

[0059] Optionally, the preset boundary condition parameters include set extreme ambient temperatures, extreme short-circuit current injection pulses, and charge / discharge commands exceeding the nominal rate. The few-shot reinforcement learning algorithm employs a proximal policy optimization mechanism. Its state space includes the aforementioned generated five-dimensional risk collaboration profile, its action space is the hyperparameter adjustment matrix within the prediction model, and its reward function consists of the prediction accuracy increment and the deviation penalty term of the collaboration requirement objective. By concatenating virtual accident samples with clear evolutionary boundaries generated by the physical model with real collaborative scenario accident samples containing real electrical noise, the underfitting problem of neural networks caused by the scarcity of catastrophic accident data during real power plant operation is effectively solved, making the resulting data-driven prediction model highly robust.

[0060] Phase 3: Risk-coupled calculation and dynamic threshold adaptive adjustment.

[0061] The method executes the following steps: calculating the risk coupling coefficient between core risks through mutual information entropy; dynamically adjusting the early warning threshold based on the three-dimensional mapping rule composed of the risk coupling coefficient, the data quality level, and the collaborative requirement level; and outputting multi-dimensional risk classification results and risk evolution prediction results through the hybrid model.

[0062] Specifically, the early warning threshold is dynamically adjusted based on a three-dimensional mapping rule consisting of the risk coupling coefficient, the data quality level, and the collaborative demand level. This includes: calculating the coupling strength coefficient between thermal runaway, insulation breakdown, overcharging and over-discharging, imbalance, inter-cluster inconsistency, and the core fire risk of the energy storage system using mutual information entropy, and using this coefficient as the risk coupling coefficient; under the condition that the risk coupling coefficient is greater than or equal to a first preset threshold and the data quality level is a preset optimal quality level, if the collaborative demand level is an emergency collaborative level, the sensitivity of the early warning threshold is increased; if the collaborative demand level is a fault collaborative level, the sensitivity of the early warning threshold is decreased.

[0063] In real-world physical scenarios, energy storage risks rarely occur in isolation. For example, inconsistencies between battery clusters can lead to localized overcharging of individual cells, which in turn can cause gas generation and increased internal resistance, ultimately evolving into thermal runaway risk. By calculating the time-series mutual information entropy of multi-source risk characteristics, the risk coupling coefficient between overcharging / over-discharging and thermal runaway can be quantified.

[0064] In this step, the logic of the three-dimensional mapping rule is as follows: When the system detects that the risk coupling coefficient is extremely high, that is, a large-scale cascading fault is in the nascent stage, such as being greater than or equal to the first preset threshold, and the data collected by the current system sensors is at the best quality level (indicating that the monitoring data is extremely reliable), if the power grid is in an emergency coordination level (for example, a primary frequency regulation condition, where the power grid urgently needs energy storage to provide high-rate power support but the overall power grid has not collapsed), in order to prevent the high voltage and high current impact from directly inducing the accumulated risks, the system will actively increase the judgment sensitivity of the warning threshold, making it easier for the system to trigger the protection mechanism, with equipment safety as the primary guarantee.

[0065] Conversely, if the power grid is at a fault coordination level, such as when a local power grid experiences a short circuit fault and faces the extreme collapse of widespread power outages, the energy storage system, as an islanded power source or the last black start support point, has a value in maintaining the grid that far outweighs the life loss of individual batteries. The system will then lower the sensitivity of the warning threshold, allowing the battery to forcibly output power under certain over-temperature or over-charge conditions, and perform degraded fault-tolerant operation to respond to the urgent external demand for grid fault ride-through.

[0066] Optionally, the multi-dimensional risk classification result includes a degree of synergistic correlation. The quantification steps of the degree of synergistic correlation include: calculating the probability value of the current risk being triggered by the coordinated action of the power grid using conditional probability. The probability value is the degree of synergistic correlation, and its value is equal to the quotient obtained by dividing the risk increment caused by the coordinated action of the power grid by the total risk increment. Based on the magnitude of the probability value, the degree of synergistic correlation is classified into three judgment levels: strong correlation risk, medium correlation risk, and weak correlation risk.

[0067] For example, the quantitative formula for collaborative association degree is expressed as follows:

[0068] ;

[0069] in, This indicates the degree of collaborative association, used to characterize this probability value; This represents the risk increment base extracted from fluctuations in pure power grid coordinated action commands; It represents the total risk increment calculated from the overall state change of the system, which is the sum of the natural aging risk inside the energy storage and the risk imposed by the external power grid.

[0070] The classification of this correlation degree has clear significance for assigning responsibility and providing scheduling guidance in practical applications: if it is determined to be a strong correlation risk, it means that the current abnormal temperature rise or insulation decay of the energy storage is mainly caused by the grid side issuing too frequent frequency regulation commands. In subsequent decision-making, the energy storage system should send a request to the grid dispatch center to restrict the issuance of subsequent high-frequency commands, rather than simply shutting down the energy storage hardware itself.

[0071] Phase 4: Three-dimensional cost model and quantitatively interpretable decision-making.

[0072] Specifically, a preset three-dimensional cost model is established based on the multi-dimensional risk classification results. The combined decision-making scheme is obtained by solving the three-dimensional cost model with cost minimization as the optimization objective. This includes: constructing the three-dimensional cost model containing the costs of missed reports, false reports, penalties for not meeting grid coordination requirements, operating depreciation costs, and source tracing and repair costs; dynamically adjusting the weight coefficients of the above costs according to preset normal coordination levels, emergency coordination levels, or fault coordination levels; and solving the three-dimensional cost model using a non-dominated sorting genetic algorithm with coordination constraints and a Bayesian optimization algorithm to generate the combined decision-making scheme containing the early warning threshold, liquid cooling configuration parameters, balance control strategy, and grid coordination adjustment suggestions.

[0073] For example, the objective function formula for the three-dimensional cost model is expressed as:

[0074] ;

[0075] in, This represents the total optimization cost target value; This refers to the cost of missed reporting caused by actual system failures without warning, such as the reset cost resulting from a fire in the energy storage compartment; This indicates the cost of false alarms caused by the system issuing alarms incorrectly while in a safe state, such as the cost of reduced system availability due to unexplained downtime. This indicates the penalty cost for failing to meet grid coordination requirements, such as the fine for failing to respond to AGC commands. This indicates operating depreciation costs, such as the energy consumption and hardware lifespan loss caused by frequent start-stop of the liquid-cooled compressor; Indicates the cost of tracing and repairing the source; to These represent the dynamically adjusted weight coefficients, and the system constraints require that the sum of the weight coefficients be constant.

[0076] During the optimization process, the Non-Dominated Sorting Genetic Algorithm with Cooperative Constraints (NSGA-II) encodes the combined decision schemes as chromosomes in the population. By simulating crossover, mutation, and non-dominated sorting operations, the algorithm searches for a Pareto optimal solution set among multiple objectives. Simultaneously, a Bayesian optimization algorithm is introduced to perform local Gaussian process regression probing on the Pareto front to reduce suboptimal decisions caused by the genetic algorithm getting trapped in local optima. The final quantitatively interpretable early warning report not only outputs the alarm level but also uses time-series graphs to demonstrate the complete tracing chain of how grid commands gradually increase battery polarization resistance and trigger heat accumulation.

[0077] like Figure 3 This is a spatial scatter plot of the Pareto front for three-dimensional cost optimization based on a non-dominated sorting genetic algorithm with collaborative constraints. In the three-dimensional coordinate system of this plot, the horizontal axis represents the cost of missed reports (in ten thousand yuan), the vertical axis represents the cost of penalties (in ten thousand yuan), and the vertical axis represents the cost of running depreciation (in ten thousand yuan).

[0078] The gray scattered points in the figure represent the initial decision population generated in the early stages of the algorithm's operation. The costs of such decision schemes are high and divergent. For example, the penalty cost of some initial decisions is close to 500,000 yuan and the depreciation cost exceeds 600,000 yuan.

[0079] The colored scatter points in the figure, which are distributed on a curved surface, represent the Pareto optimal solution set obtained after multi-objective optimization iteration. The color of the scatter points transitions from blue to dark red, which intuitively reflects the numerical change of operating depreciation costs from low to high.

[0080] The semi-transparent surface passing through the colored scattering area represents the local Gaussian process regression probe surface constructed after introducing the Bayesian optimization algorithm. The concave convergence shape of this probe surface objectively reflects the interrelationships among various costs in the actual operation of the energy storage system.

[0081] As can be seen from the data in the figure, when the system compresses both the cost of underreporting and the cost of penalties to a low level of 100,000 yuan in order to meet the needs of external collaboration, frequent hardware adjustments will cause the operating depreciation cost to increase sharply to about 1.15 million yuan; conversely, if the cost of penalties is allowed to be relaxed to the level of 500,000 yuan, the operating depreciation cost can fall back to below 200,000 yuan.

[0082] This spatial distribution characteristic intuitively demonstrates that the three-dimensional cost model established in this invention can effectively quantify the economic and safety game boundary under complex working conditions. By avoiding local suboptimal solutions, it provides a decision-making optimization basis with engineering feasibility for the system to generate quantitatively interpretable early warning reports and execute dynamic degradation control at different coordination levels.

[0083] Phase 5: Degradable mode triggering and three-level linkage closed loop.

[0084] It should be noted that when the data quality confidence level is detected to be lower than the preset confidence threshold or the model prediction uncertainty is higher than the preset uncertainty threshold, the degraded mode is triggered to dynamically correct the warning threshold and the system power boundary, and the combined decision-making scheme is executed based on the three-level decision-making authority allocation system of edge layer, cloud and grid side, and feedback instructions are received to form a decision closed loop.

[0085] Specifically, when the data quality confidence level is detected to be lower than a preset confidence threshold or the model prediction uncertainty is detected to be higher than a preset uncertainty threshold, a degradable mode is triggered to dynamically correct the warning threshold and the system power boundary. This includes: triggering the degradable mode when the data quality confidence level is detected to be lower than a second preset threshold, when there is a time-series break in the full lifecycle data, or when the model prediction uncertainty is detected to be higher than a third preset threshold; in normal collaborative scenarios, the thermal runaway safety threshold in the warning threshold is lowered; in emergency collaborative scenarios or fault collaborative scenarios, the thermal runaway safety threshold and the upper limit of energy storage power output in the system power boundary are quantitatively and progressively lowered.

[0086] In real-world energy storage operations, if strong electromagnetic interference on the communication bus causes widespread data corruption in some battery clusters, or if the confidence interval of the hybrid model's prediction results is too large (indicating extreme uncertainty about the current state), the system will force a degradation. The physical significance of this tiered downscaling in emergency collaborative scenarios is to prevent the entire power station from being directly tripped and shut down due to partial distortion of the underlying data. By gradually reducing the total control power limit according to a safety margin, for example, from 100% full power to 80% and then 60%, the system can maintain some usable capacity for the grid while ensuring that the batteries in the lower-level dead zones are not damaged by excessive current.

[0087] Specifically, the combined decision-making scheme is executed based on a three-level decision-making authority allocation system of edge layer, cloud, and grid side, and feedback instructions are received to form a decision-making closed loop, including: the edge layer is responsible for executing liquid cooling parameter and equalization loop control decisions in normal collaborative scenarios, as well as local degradable optimization execution when data quality is abnormal; the cloud is responsible for executing power limiting decisions in emergency collaborative scenarios or fault collaborative scenarios, and performing overall collaborative decision-making when the cross-domain coupling coefficient is greater than or equal to the fourth preset threshold; after receiving the quantified interpretable early warning report and grid collaborative adjustment suggestions, the grid side feeds back scheduling instructions to the cloud to iteratively update the hybrid model.

[0088] The edge layer is deployed directly in the local controller within the energy storage container, with computational latency typically controlled in the millisecond range. It is responsible for executing direct hardware actions such as adjusting water pump speed and starting / stopping specific cluster equalization boards, ensuring rapid convergence of the underlying physical state. The cloud layer is deployed on a remote high-performance server, which has cross-site level data and long-sequence historical models. It is responsible for calculating global power limits and interacting with the external power grid platform. The power grid side, as the external excitation source of the closed-loop system, receives collaborative adjustment suggestions reported by the cloud and modifies its own active and reactive power scheduling algorithms, forming a complete information flow closed loop from monitoring-evaluation-prediction to software and hardware collaborative control-scheduling strategy reconstruction.

[0089] In summary, existing technologies using fixed safety thresholds for risk warning can only passively address single-dimensional faults. Furthermore, single physical models or data-driven models are prone to false alarms, missed alarms, or even system crashes and network outages when faced with complex power grid disturbances and missing sensor data. The technical solution provided in this embodiment, however, generates a five-dimensional feature profile through multi-source heterogeneous data quality grading and cross-domain coupling coefficients, reducing uncertainty at the data source. Through hybrid modeling of physical mechanisms and data-driven approaches combined with reinforcement learning using virtual accident samples, the system maintains strong generalization prediction capabilities against external disturbances even with scarce samples.

[0090] Furthermore, by relying on mutual information entropy to identify risk coupling states and dynamically map and adjust early warning thresholds, combined with cost optimization decisions and hierarchical degradation control closed loops adapted to harsh operating conditions, this solution not only ensures that the underlying physical safety red line of the energy storage system is not violated in practical applications, but also maintains the available capacity of energy storage in response to grid commands at a high level, resolving the conflict between safety and economic operation and dispatch, and has high industrial promotion value and feasible prospects.

[0091] Example 2:

[0092] In the actual engineering construction and deployment of digital twin models for energy storage systems, virtual data generated solely based on the five-field coupling physical mechanism model of heat-electricity-aging-insulation-equilibrium often exhibits highly idealized characteristics, lacking the complex electromagnetic harmonic interference and sensor hardware aging characteristics of actual large-capacity energy storage power stations. If these idealized virtual data are directly mixed with real data to train the prediction model, the model will perform well in laboratory evaluation environments, but exhibit a significant "sim-to-real gap" phenomenon when deployed at the edge computing nodes of real power stations. That is, the model is extremely sensitive to real-world noise, and its generalization prediction ability drops sharply.

[0093] To address the discrepancy between simulated and real-world data distribution, this embodiment provides a deep training mechanism based on adversarial virtual-real feature alignment and dynamic gradient cosine similarity weighting, specifically including the following:

[0094] Phase 1: Extraction of real-world disturbance features and construction of dynamic noise distribution matrix.

[0095] Specifically, the method steps provided in this embodiment include: extracting sensor operating noise and numerical drift features from the real collaborative scenario accident sample, and constructing a dynamic noise distribution matrix.

[0096] For example, in the actual operation environment of a centralized energy storage power station, the underlying physical data collected by the battery management system mainly relies on high-precision analog-to-digital converter chips, negative temperature coefficient thermistors, and Hall closed-loop current sensors. In actual operation, these hardware components are inevitably affected by electromagnetic harmonic interference generated by the high-frequency switching of the insulated-gate bipolar transistors (IGBTs) in the energy storage converter, as well as the aging effects of components operating in a high-temperature, high-pressure, and enclosed environment for extended periods. Therefore, the voltage, current, and temperature signals actually collected are inevitably superimposed with random high-frequency oscillation noise and low-frequency baseline drift that changes slowly over time or with the number of charge-discharge cycles. Extracting these non-ideal features first requires acquiring a massive amount of real-world collaborative scenario accident samples and historical normal operation samples. The system introduces an ensemble empirical mode decomposition algorithm to adaptively decompose the time-series signals from various sensors that are actually collected.

[0097] It is also important to note that the ensemble empirical mode decomposition algorithm can decompose the original non-stationary time signal into several intrinsic mode function (IMF) components in different frequency bands and a residual term representing the global trend. High-frequency operational noise components are mainly concentrated in low-order IMFs, while long-term numerical drift characteristics are intuitively reflected in the residual trend term. The system obtains the independent noise and drift time series of each independent sensor by calculating the residual time series after removing the fundamental frequency component of the ideal signal. Subsequently, the system performs statistically-oriented parametric modeling on these extracted feature sequences. For operational noise, the system calculates its mean variance and spectral density under different ambient temperatures and different charge / discharge current rates; for numerical drift characteristics, the system extracts the drift slope and nonlinear offset as a function of time or cycle number through polynomial regression fitting.

[0098] like Figure 4 The diagram shows the empirical mode decomposition and baseline drift bitmap of the original noise signal from the sensor. The diagram is divided into three sub-diagrams from top to bottom, which intuitively demonstrate the signal processing process of extracting non-ideal hardware disturbance features from accident samples in real collaborative scenarios.

[0099] The horizontal axis of the first sub-graph represents time (seconds), and the vertical axis represents voltage (volts). The gray waveform curve in the graph represents the raw acquisition signal actually collected by the battery management system within the 0 to 500 second range. The overall voltage reference of this signal slowly decreases from 3.3 volts, but its waveform is superimposed with a large number of dense spikes and irregular fluctuations, truly reflecting the working state of the sensor in the complex electromagnetic environment of the energy storage power station.

[0100] The horizontal axis of the second subplot represents time (seconds), and the vertical axis represents amplitude (volts). The blue high-frequency oscillation curve in the figure represents the first-order intrinsic mode function component, i.e., the high-frequency noise component, extracted by the ensemble empirical mode decomposition algorithm. The amplitude of this component exhibits high-frequency random fluctuations between -0.2 volts and +0.2 volts, objectively quantifying the electromagnetic harmonic interference intensity induced on the sampling line by the high-frequency switching action of the energy storage converter.

[0101] The horizontal axis of the third subplot represents time in seconds, and the vertical axis represents offset in volts. The smooth red curve in the figure represents the residual trend term extracted by the algorithm, i.e., the baseline drift component. This component curve exhibits a low-frequency nonlinear shape, slowly decreasing from -0.05 volts to -0.13 volts and then gradually rising again, characterizing the numerical drift offset of the sensor components under long-term operation and high-temperature stress.

[0102] Through the data decomposition and waveform transformation of this series of multi-axis sub-graphs, it is clear that the technical solution of this invention can accurately separate the chaotic on-site physical disturbances into independent error sequences that can be parameterized and modeled, thereby providing high-confidence underlying data support for the subsequent construction of dynamic noise distribution matrix and generation of augmented virtual samples.

[0103] Optionally, the process of constructing the dynamic noise distribution matrix is ​​as follows: the total number of sensors participating in data acquisition in the energy storage system is globally set as a first constant, denoted as a natural number. The total number of sampling steps for the time series within a single prediction window is set globally to a second constant, denoted as a natural number. The dynamic noise distribution matrix is ​​a matrix with dimension 1. OK, A two-dimensional real matrix. Each element of this matrix contains the expected real value of the overall error for a specific sensor at a specific time step.

[0104] Specifically, the formula for calculating the elements within the dynamic noise distribution matrix is ​​as follows:

[0105] ;

[0106] in, The first element in the constructed dynamic noise distribution matrix represents the... Line number The actual values ​​of the elements in the column; This represents the global natural number index number of the sensor, with a value ranging from 1 to... ; The global natural number index number representing the time step, with values ​​ranging from 1 to 1. ; This represents the random noise basis generation function based on a Gaussian distribution; This indicates the statistical result of the first The mathematical mean of the historical operating noise of each sensor; This indicates the statistical result of the first Standard deviation of historical operating noise of each sensor; The state-scale scaling factor is used to characterize the amplification effect of local electromagnetic noise caused by a surge in the amplitude of the current operating current of the current energy storage system or an abnormal increase in the local ambient temperature. This factor is a dynamic real number that is positively correlated with the real-time absolute power of the system. Indicates the first The sensor at the first The numerical drift over a given time step is calculated directly using the aforementioned polynomial regression fitting function.

[0107] The engineering significance of the above extraction and construction process lies in accurately transforming the random physical disturbances and non-ideal attenuation characteristics of sensor hardware that are difficult to analyze in the real physical world into mathematical matrix entities that can be read, mapped and operated on by computer programs in a standardized manner.

[0108] Phase 2: Dynamic noise injection and augmented virtual sample generation.

[0109] Specifically, the method steps provided in this example further include: mapping and injecting the dynamic noise distribution matrix into the virtual accident sample to generate an augmented virtual sample with real operational disturbances.

[0110] For example, the five-field coupled physical mechanism model, based on preset boundary condition parameters such as preset external short-circuit resistance and set coolant flow rate, solves a system of high-dimensional ordinary differential equations to output a smooth, ideal time series trajectory that conforms to fundamental physical laws and contains no acquisition errors—the original virtual accident sample. To ensure that the virtual accident sample closely resembles the real sensor feedback signal in terms of data appearance and underlying statistical distribution, the system must perform a rigorous noise injection process during the algorithm preprocessing stage.

[0111] For example, the calculation formula for mapping and injecting the dynamic noise distribution matrix into the virtual accident sample is as follows:

[0112] ;

[0113] in, This indicates that the augmented virtual sample generated after processing is in the 1st... The sensor dimension, the first The final signal value at each time step; This represents the ideal signal value of the original virtual accident sample, directly output from the five-field coupled physical mechanism model and without any noise processing, at the corresponding sensor dimension and time step. This represents the corresponding coordinate element value of the dynamic noise distribution matrix constructed by the aforementioned calculation; A function representing a physical state mask.

[0114] Optionally, the purpose of introducing a physical state masking function is to avoid irrational noise injection actions that violate the underlying physical logic. For example, when the original virtual accident sample shows that the main contactor of a battery cluster is in the open state, i.e., the actual physical current value is strictly 0, the thermal noise and servo interference noise of the current sensor should also be significantly suppressed, and tens of amperes of Gaussian noise should not be added arbitrarily. Physical state masking function Based on the input The signal characteristics are analyzed, and a real-valued weighting factor between 0 and 1 is output to ensure that the augmented virtual sample after noise injection not only has the fluctuations and disorder of the real world, but also strictly follows the electrochemical energy conservation and thermodynamic heat transfer laws of the underlying energy storage battery under boundary conditions.

[0115] Phase 3: Construction of adversarial training architecture and forward propagation of hybrid features.

[0116] Specifically, the method steps provided in this embodiment further include: constructing an adversarial training architecture that includes a feature extraction network and a domain discriminator, and inputting the augmented virtual sample and the real collaborative scene accident sample into the feature extraction network in parallel to obtain a hybrid feature representation.

[0117] For example, in the deep neural network backbone design for few-shot reinforcement learning, the feature extraction network is a composite deep learning model containing bidirectional long short-term memory (Bi-LSTM) neural network layers and one-dimensional causal convolutional neural network layers (1D-CNN). The set of all trainable weights and bias parameters of the feature extraction network is denoted as the first parameter vector set. The domain discriminator is a sub-network that branches in parallel at the output of the feature extraction network. It contains three fully connected layers and connects a logistic regression activation function (Sigmoid function) to the final output. The set of all trainable weights and bias parameters of the domain discriminator is denoted as the second parameter vector set.

[0118] It's also important to note that in the actual adversarial network forward propagation process, the system executes computations in batch processing mode. The total number of samples in a single training batch is set to a natural number. This batch contains the same number of augmented virtual sample data streams and real collaborative scenario accident sample data streams. These two data streams are packaged and fed into the feature extraction network in parallel. The one-dimensional causal convolutional kernel of the feature extraction network is responsible for extracting local mutation features of the signal within a very short time window, such as hundreds of milliseconds; subsequently, bidirectional long short-term memory units are responsible for extracting long-range temporal dependencies of the signal over charge-discharge cycles lasting several hours.

[0119] Optionally, the high-dimensional tensor output by the feature extraction network in the last hidden layer is the hybrid feature representation. The hybrid feature representation is a highly abstract set of mathematical vectors that encodes not only the electrochemical risk features extracted from real samples but also the boundary accident evolution features extracted from augmented virtual samples. In this training batch, the cyclic index variable of the samples is set to... Its value ranges from 1 to , No. The tensor representing the hybrid features generated after forward propagation of each sample through the feature extraction network is denoted as . .

[0120] Phase 4: Calculation of neighborhood divergence loss and backpropagation of joint gradient.

[0121] Specifically, the method steps provided in this embodiment further include: calculating the domain divergence loss of the hybrid feature representation through the domain discriminator, and combining the domain divergence loss with the policy feedback loss of the few-shot reinforcement learning algorithm to perform gradient backpropagation, so as to drive the feature extraction network to generate domain-independent features that are insensitive to the source of virtual and real data.

[0122] For example, the core task of the domain discriminator is to act as a referee node, whose input is the hybrid feature representation tensor output by the feature extraction network. The output is a real probability value between 0 and 1. This probability value is specifically used to determine whether the currently input feature tensor originates from data collected from a real power plant or from simulation data from a twin virtual model. Through this continuous determination process, the system calculates the neighborhood divergence loss.

[0123] For example, the formula for calculating the neighborhood divergence loss is:

[0124] ;

[0125] in, This represents the calculated scalar value of the neighborhood divergence loss; This indicates the total number of samples in the current training batch; Represents the global natural index variable for samples within a batch; Indicates the first The real data source for each sample is a hard label. When the sample is a real collaborative scenario accident sample, the label value is set to a constant one; when the sample is an augmented virtual sample, the label value is set to a constant zero. Indicates the first The hybrid feature representation tensor generated after each sample passes through the feature extraction network; This represents the predicted source probability value output by the neighborhood discriminator for the input tensor; This represents the natural logarithm function.

[0126] Optionally, in the aforementioned adversarial training architecture, the core component connecting the feature extraction network and the neighborhood discriminator is the Gradient Reversal Layer (GRL). The GRL is concatenated between the output of the feature extraction network and the input of the neighborhood discriminator. During the forward propagation inference phase of the network, the GRL acts as an identity mapping, meaning it does not modify the mixed feature representation tensor and directly passes the data as is. However, during the backpropagation differentiation phase, the GRL multiplies the gradient values ​​flowing through it by a negative scaling factor. The overall optimized loss function for its joint gradient backpropagation can be expressed as:

[0127] ;

[0128] in, This represents the joint total loss scalar value that the feature extraction network uses during parameter update iterations; This represents the policy feedback loss of a few-shot reinforcement learning algorithm. This loss measures the absolute deviation between the network's output decision action (such as the adjustment of the warning threshold) and the optimal collaborative requirement objective, which drives the model to learn how to accurately assess risk. This represents the gradient penalty constant term of the inverted gradient layer, which is usually a preset scalar with a value greater than zero, such as 0.1; This represents the scalar value of the neighborhood divergence loss obtained from the aforementioned calculation.

[0129] In actual network parameter optimization iterations, this is a typical min-max game. The neighborhood discriminator continuously optimizes its second parameter vector set using conventional gradient descent, striving to distinguish between real and fake data as accurately as possible, thus achieving... Minimize. However, when the feature extraction network receives the backpropagation gradient and updates the first parameter vector set, due to the negative sign processing of the gradient reversal layer, the parameter optimization direction of the feature extraction network is not only to minimize the policy feedback loss. This is also to maximize the domain divergence loss. The physical significance of this technical feature lies in its ability to force the feature extraction network, from a mathematical foundation, to actively forget and ignore superficial electromagnetic noise features that can only distinguish whether data comes from simulation or reality. Instead, it focuses all its attention on extracting the essential features of electrochemical polarization and thermodynamic accumulation that lead to the core risks of energy storage systems. When the mixed feature representation output by the feature extraction network makes it difficult for the domain discriminator to distinguish between true and false data, that is, when the output predicted source probability value... When the random prediction baseline is around 50%, it indicates that the system has successfully generated domain-independent features that are highly insensitive to the source of virtual and real data.

[0130] like Figure 5 This is a convergence curve of training loss versus domain discrimination probability for an adversarial network with an inverted gradient layer. The horizontal axis represents the number of training iterations (unit: times), ranging from 0 to 1000. The vertical axis includes two axes: the left axis represents the policy feedback loss (unit: dimensionless), and the right axis represents the predicted source probability value (unit: dimensionless).

[0131] The figure mainly presents two convergence curves. The blue curve represents the policy feedback loss curve of the feature extraction network during the optimization process, and the red curve represents the predicted source probability value output by the neighborhood discriminator. The gray dashed line represents the baseline of the blind guessing state with a value of 0.5. The trend of the blue curve shows that as the number of training iterations increases, the policy feedback loss steadily decreases from an initial value of around 2.5 and eventually converges to around 0.5. This indicates that the data-driven prediction model is continuously learning how to accurately assess the internal risks of the energy storage system and gradually approaching the target of collaborative requirements.

[0132] Meanwhile, in the early stages of training, due to the significant difference between real-world data and simulated virtual data, the domain discriminator could accurately identify the data source with a probability close to 0.9. However, as the inverted gradient layer continuously applied gradients with negative scaling factors to the feature extraction network during the error backpropagation phase, the feature extraction network gradually generated domain-independent features that were insensitive to the source of the virtual or real data. This is reflected in the red curve as a waveform trend of oscillating decay in the predicted source probability value, which, after approximately 600 iterations, eventually converged and fluctuated closely around the blind guessing baseline of 0.5.

[0133] This data objectively demonstrates that the game-theoretic training mechanism employed in this invention can effectively reduce the distribution difference between virtual simulation samples and real power plant samples, enabling the final trained small-sample reinforcement learning model to be insensitive to on-site interference environments and robust in inferring risk states.

[0134] Phase 5: Dynamic sample weighting based on gradient cosine similarity.

[0135] Specifically, the method steps provided in this embodiment further include: calculating the cosine similarity of the network parameter gradient of each augmented virtual sample based on the gradient descent direction of the network parameters of the real collaborative scenario accident sample, and assigning dynamic training weights to each augmented virtual sample according to the cosine similarity, thereby completing the training of the data-driven prediction model.

[0136] For example, although the aforementioned steps have minimized the macroscopic distribution differences between virtual and real data through dynamic noise injection and domain adversarial training, in rare edge cases, the rigid solution of the internal ordinary differential equations of the five-field coupled physical mechanism model may still diverge when inputting extremely harsh boundary condition parameters, thus generating a small number of singular virtual data packets that violate basic electrochemical principles. If these hidden anomalous augmented virtual samples participate in the gradient update of the neural network indiscriminately, they will cause severe negative transfer to the model, destroying the correct physical mapping laws that the model has learned. Therefore, it is necessary to quantify and redistribute the training gain value of each augmented virtual sample in the batch before backpropagation. Let the total number of augmented virtual samples in the current batch be a natural number. .

[0137] For example, the formula for calculating gradient cosine similarity is:

[0138] ;

[0139] in, Indicates the first The cosine similarity scalar value of the network parameter gradient of an augmented virtual sample; This represents a specific natural number index of the augmented virtual sample within the current training batch, with a value ranging from 1 to... ; This represents a multi-dimensional vector of network parameter gradient descent directions calculated by summarizing accident samples in the current feature extraction network parameter space in real collaborative scenarios. This vector represents a reasonable guiding direction for the model to evolve in accordance with the objective laws of the real physical world. Indicates the first Each augmented virtual sample is a multidimensional vector of network parameter gradient directions independently computed in the current feature extraction network parameter space. The numerator represents the dot product algebra of the two gradient vectors, and the denominator represents the product of the Euclidean norms (lengths) of the two gradient vectors.

[0140] Optionally, the calculated cosine similarity scalar value is distributed within the range of -1 to +1. When the value is close to +1, it indicates that the direction of the augmented virtual sample guiding the optimization of model parameters is highly consistent with the direction guided by the real data, and the sample has extremely high reinforcement training gain value. Conversely, when the value is negative, it indicates that the gradient direction provided by the augmented virtual sample is contrary to the physical laws of the real world, and directly merging its gradients will seriously interfere with the normal convergence of the model.

[0141] For example, the formula for calculating the dynamic training weights assigned to each of the augmented virtual samples based on the cosine similarity is as follows:

[0142] ;

[0143] in, Indicates the final assignment to the first The real values ​​of the dynamic training weights of an augmented virtual sample; This represents the linear rectified cutoff function (ReLU logic), whose core function is to force the weights of all anomalous virtual samples that generate negative gradients to 0, directly blocking their participation in the parameter updates of the current training cycle at the underlying level, thus achieving physical isolation of anomalous samples. This represents the preset weight normalization sum constant, used to maintain the stability of the global learning rate and avoid gradient explosion caused by sudden weight changes; This represents a natural number variable specifically used as an inner loop index in the summation operation, iterating through all virtual samples. This summation term is used to calculate the sum of all valid positive cosine similarities within the same batch, thereby completing the probabilistic normalization of the weights.

[0144] Through the aforementioned iterative dynamic gradient direction monitoring and sample weight shuffling mechanism, the prediction model can retain the advantage of expanding the system boundary cognition with massive virtual samples, while ensuring that the core network parameters always converge smoothly and efficiently along the direction close to the real physical scene, and finally complete the high-quality training closed loop of the data-driven prediction model.

[0145] Comparative analysis with existing technologies reveals that traditional digital twin systems, when combining physical simulation data with AI data-driven algorithms, generally employ simple hybrid data pool stacking training or coarse fine-tuning pre-training modes. This existing technological architecture is highly susceptible to feature overfitting when faced with multi-source high-frequency electromagnetic noise, sensor baseline drift, and sporadic abrupt acquisition errors in real-world operating environments. This results in predictive models achieving extremely high evaluation metrics on cleaned laboratory datasets, but frequently outputting erroneous safety assessment conclusions in real-world scenarios due to their oversensitivity to real-world noise, leading to serious false negatives and false positives.

[0146] In this embodiment, the technical solution completes the low-level data alignment by pre-extracting and injecting a real sensor dynamic noise distribution matrix based on natural state evolution, then constructs a domain adversarial discriminant network with an inverted gradient layer to complete the high-dimensional feature alignment, and finally uses gradient cosine similarity to dynamically adjudicate the training weights of virtual samples to perform gradient-level isolation in the backpropagation stage.

[0147] This comprehensive, multi-level joint training mechanism not only achieves a rigorous closed loop in terms of mathematical algorithm principles and deep network architecture, but also fundamentally solves the core technical bottleneck of virtual-real distribution differences when energy storage digital twins are deployed across domains. This enables the trained small-sample reinforcement learning model to possess strong insensitivity to high-intensity electromagnetic interference and robustness in state inference. In practical applications, this solution significantly improves the accuracy and reliability of identifying internal electrochemical risks throughout the entire lifecycle of energy storage systems under complex grid collaborative operating conditions, effectively safeguarding the underlying operational safety of massive energy assets in complex grid interaction scenarios.

[0148] Example 3:

[0149] In the operating environment of large-capacity energy storage power stations at the megawatt-scale or gigawatt-hour level, the entire energy storage system is typically composed of dozens or even hundreds of energy storage clusters connected in parallel in terms of physical topology. When faced with issues such as loose local sensor sampling lines, strong electromagnetic interference affecting communication, or a significant decrease in the confidence level of local state predictions, traditional battery management systems and energy management systems often adopt a strategy of globally and proportionally reducing the total active power command and reactive power command of the entire energy storage power station for system-level safety considerations. This coarse control logic leads to the forced idle of the available remaining capacity of a large number of healthy energy storage clusters in the system, resulting in a precipitous drop in the energy storage power station's responsiveness to frequency regulation and peak shaving commands issued by the grid.

[0150] To overcome this technical deficiency, this embodiment provides a spatially aware asymmetric cluster-level degradation and physical forced compensation mechanism, specifically including the following:

[0151] Phase 1: Extraction and construction of the cluster-level spatial uncertainty distribution matrix.

[0152] Specifically, the method steps provided in this embodiment include: extracting the local data quality confidence and local prediction uncertainty of each physically parallel energy storage cluster in the energy storage system, and constructing a cluster-level spatial uncertainty distribution matrix.

[0153] For example, a centralized energy storage system consists of multiple energy storage units connected in parallel to a multi-branch energy storage converter via DC busbars. Each series-connected battery branch, which independently performs the charging and discharging physical process and has an independent master-slave control battery management architecture, is called a physically parallel energy storage cluster. The total number of energy storage clusters in the system is globally set to a constant, denoted as . Local data quality confidence refers to the weighted quantitative index of the effective data frame integrity rate and signal-to-noise ratio of the underlying physical data collected by the internal temperature and voltage sensors of a single specific energy storage cluster within a specific time window. Local prediction uncertainty refers to the mathematical variance of the output risk assessment probability distribution of the aforementioned trained data-driven prediction model when predicting the risk state evolution of the specific energy storage cluster within the next hour.

[0154] It is also important to note that the calculation of local prediction uncertainty relies on the Monte Carlo forward propagation sampling algorithm. During the inference phase of the neural network model, the algorithm targets the first... A physically parallel energy storage cluster is used. A random neuron inactivation mechanism is introduced into the fully connected layer of the network to perform multiple independent forward propagation calculations, and the dispersion of multiple risk prediction assessment values ​​is statistically analyzed. The formula for calculating this local prediction uncertainty is:

[0155] ;

[0156] in, Indicates the first The local prediction uncertainty of an energy storage cluster; The natural number index represents the global energy storage cluster, with values ​​ranging from 1 to 1. ; This represents the set constant for the total number of samples taken during the Monte Carlo forward propagation; The natural number index number representing a single Monte Carlo sample; Indicates the first In the second sampling calculation, the output of the data-driven prediction model is for the first... Real-valued risk prediction and assessment data for each energy storage cluster; Indicates in total In the second sampling calculation, the model output is for the first... The mathematical average expectation of the real values ​​of the risk prediction assessment for an energy storage cluster.

[0157] The above formula allows the system to transform the black-box characteristics of data-driven prediction models into quantifiable confidence indices. When the value is too large, the model lacks a precise predictive grasp of the current internal electrochemical state of the energy storage cluster.

[0158] Optionally, the process of constructing the cluster-level spatial uncertainty distribution matrix involves the structured integration of the state quantification indicators of all parallel energy storage clusters in the entire energy storage system along a spatial dimension. This cluster-level spatial uncertainty distribution matrix is ​​a matrix with dimension... A two-dimensional real matrix with two rows and two columns.

[0159] For example, the structural formal representation of this cluster-level spatial uncertainty distribution matrix is ​​as follows:

[0160] ;

[0161] in, This represents the generated cluster-level spatial uncertainty distribution matrix; the first column of the matrix represents the local data quality confidence score of each energy storage cluster, for example... The first column represents the real-valued local data quality confidence score of the energy storage cluster with index number one; the second column of the matrix represents the local prediction uncertainty of each energy storage cluster calculated above, for example... This represents the local prediction uncertainty value for the energy storage cluster with index number one.

[0162] The role of this matrix in practical engineering is to provide a matrix mapping table with the ability to perceive the physical topology of equipment for the subsequent reconstructing of the scheduling strategy of the energy management system. This enables the control module to accurately locate the specific physical branch nodes that are data monitoring blind spots caused by aging hardware lines or communication bus interference.

[0163] like Figure 6 This is a two-dimensional heatmap of spatial uncertainty and data quality at the cluster level in energy storage power stations. The horizontal axis of the graph is dimensionless, representing the column number of the energy storage cluster, containing discrete location coordinates from 1 to 10. The vertical axis is dimensionless, representing the row number of the energy storage cluster, containing discrete location coordinates from 1 to 8. The color bars on the side of the graph represent dimensionless local prediction uncertainty, with values ​​ranging from 0 to 1.

[0164] The graph uses a grid of squares transitioning from blue to dark red to visually map the real-time data quality monitoring and risk assessment status of eighty physically parallel energy storage clusters within a large-scale energy storage power station. The large areas of dark blue and light blue in the matrix represent the second type of energy storage clusters with lower local prediction uncertainties, whose uncertainties are stably distributed between 0.1 and 0.3. This objectively indicates that the system has sufficient predictive control over the internal electrochemical and thermodynamic states of these nodes.

[0165] The few highlighted square nodes in the matrix, appearing in deep red and orange, indicate a surge in their local prediction uncertainty values, exceeding the third preset threshold, and are thus classified as Class I energy storage clusters. For example, the red nodes in the 3rd row and 4th column and the 3rd row and 5th column of the diagram have prediction uncertainty values ​​reaching dangerous levels of 0.85 and 0.92, respectively. This clearly reflects that this specific physical branch has become a blind spot for system data monitoring due to underlying sensor failures or communication disruptions.

[0166] By constructing and displaying a matrix mapping table with device spatial topology awareness, it can be clearly demonstrated that the system's energy management module can accurately locate risky physical branch nodes. Based on this two-dimensional mapping data, the system no longer adopts the conventional and coarse-grained global unified power reduction strategy. Instead, it can individually execute local power limiting and physical forced compensation actions for the maximum rated flow of liquid cooling pipelines for red high-risk nodes, while dynamically increasing the power allocation weight for blue healthy nodes to fill the system's total power deficit. This spatially aware asymmetric scheduling mechanism effectively isolates the cascading risks caused by local hardware anomalies, ensuring that the energy storage system can still output sufficient supporting power to the external grid under degraded fault-tolerant operation, significantly enhancing the fault tolerance and availability of the implementation plan in complex industrial environments.

[0167] Phase Two: Abandon the strategy of globally unified downsampling and depth restriction of the first type of energy storage cluster.

[0168] Specifically, the method steps provided in this embodiment further include: based on the cluster-level spatial uncertainty distribution matrix, abandoning the globally unified power downscaling command, and performing an asymmetric cluster-level power redistribution step.

[0169] For example, abandoning the globally unified power reduction command means that the system's total energy management unit no longer issues equal charging and discharging current allocation targets to all energy storage converter power modules. In traditional logic, if the system's total power needs to be reduced by 20% due to a cluster failure, the current commands for all clusters are reduced by 20% proportionally. However, in this embodiment, the system will break the equal allocation mechanism and implement differentiated scheduling based on the real-time uncertainty state of each cluster in the cluster-level spatial uncertainty distribution matrix. The core logic of the asymmetric cluster-level power reallocation step is to divide the energy storage cluster group into different categories based on uncertainty thresholds, thereby performing isolation control at both the physical and algorithmic levels.

[0170] Specifically, in the asymmetric cluster-level power redistribution step, for the first type of energy storage cluster whose local prediction uncertainty is higher than the third preset threshold, its local power output upper limit is limited to a preset safe power value and its local thermal runaway safety threshold is simultaneously lowered.

[0171] It is also important to note that the third preset threshold is a fixed real number determined through backtesting of historical energy storage power station accident samples and calibrated by expert experience. During the system's operation cycle, the energy management system scans the second column of the cluster-level spatial uncertainty distribution matrix in real time. When the local prediction uncertainty value corresponding to a row in the matrix is ​​strictly greater than the third preset threshold, it indicates that the system can no longer accurately grasp the true polarization resistance and thermal accumulation state within the corresponding energy storage cluster, and this energy storage cluster is classified as a Class I energy storage cluster. The set of Class I energy storage clusters is globally denoted as […]. The preset safe power value refers to the stable charge and discharge power extreme constant that can still ensure that the internal temperature rise rate of the battery pack does not trigger the separator melting and thermal runaway chain reaction under the most severe external ambient temperature and the extreme conservative heat dissipation conditions of liquid cooling system failure.

[0172] For example, the formula for limiting the local power output upper limit of the first type of energy storage cluster to a preset safe power value is calculated as follows:

[0173] ;

[0174] in, This represents the first [unit / item] after processing by the constraint algorithm. A storage cluster (and the cluster satisfies) The actual active power command scalar quantity executed under the condition; This represents the original demand power command scalar that the cluster should have undertaken if the system had not triggered degradation and continued to execute the global unified power allocation; This indicates the calibrated preset safe power value; This represents a mathematical function that takes the minimum value of two input parameters.

[0175] The physical significance of this step is that, for the first type of energy storage cluster where data monitoring is in a blind spot, by forcibly and significantly reducing its operating current, the Joule heating rate caused by the internal ohmic resistance and polarization resistance of the battery is reduced, thereby artificially constructing an isolation zone at the physical level to block the spread of heat.

[0176] Optionally, the process of simultaneously lowering the local thermal runaway safety threshold for the first type of energy storage cluster is a dynamically quantified defense mechanism. Due to abnormally high uncertainty in data prediction, the pre-set static temperature alarm threshold of the system can no longer provide sufficient advance response time. The threshold temperature red line must be dynamically lowered in a non-linear decay manner according to the degree of uncertainty deterioration.

[0177] For example, its dynamic down-adjustment formula is defined as:

[0178] ;

[0179] in, Indicates the calculation of the updated number of... One energy storage cluster (meeting) The local thermal runaway safety threshold temperature, with the physical dimension in degrees Celsius; This represents the baseline thermal runaway safety threshold temperature constant for an energy storage system under normal healthy conditions. This represents the preset threshold decay sensitivity coefficient, used to adjust the descent slope of the logarithmic function; Represented by natural constant A logarithmic function with base 0; Indicates the first The actual local prediction uncertainty value of an energy storage cluster; This represents the set uncertainty reference constant, and the algorithm constraints. Always greater than .

[0180] By introducing the natural logarithm function, the system constructs a steep defense logic: when the prediction uncertainty just exceeds the third preset threshold, the pressure on the safety threshold is relatively gentle; however, as the uncertainty deteriorates sharply, the safety threshold will be pressured down extremely rapidly, ensuring that when the underlying sensor data causes serious delays or significant distortions, the system-level fire-fighting linkage mechanism or physical fuse protection mechanism can be triggered earlier.

[0181] Phase 3: Asymmetric power compensation scheduling of the second type of energy storage cluster.

[0182] Specifically, the method steps provided in this embodiment further include: for the second type of energy storage cluster whose local prediction uncertainty is lower than or equal to the third preset threshold, dynamically increasing the local power output allocation weight of the second type of energy storage cluster without touching its physical operating limit, so as to compensate for the power deficit of the first type of energy storage cluster.

[0183] For example, in the aforementioned control cycle, when the output power of the first type of energy storage cluster is forcibly suppressed by the energy management system, if the entire energy storage power station is to maintain overall coordination in fulfilling grid dispatch commands, such as Automatic Generation Control (AGC) commands, an active power execution gap will inevitably occur within the system. To prevent the grid dispatch platform from determining that the energy storage power station has defaulted, this power gap must be dynamically filled by the second type of energy storage cluster, which is under clear monitoring, has a stable internal thermodynamic state, and whose local prediction uncertainty is lower than or equal to a third preset threshold. The set of the second type of energy storage clusters is globally denoted as […]. .

[0184] For example, the formula for calculating the total system power deficit resulting from the limitation of the first type of energy storage cluster is as follows:

[0185] ;

[0186] in, This represents the total power deficit scalar value within the current control cycle. This represents the set of all energy storage clusters classified as Class I. This represents a natural number variable specifically used as a loop index for the first type of energy storage cluster in the summation function; This represents the scalar value representing the original power demand command undertaken by the first type of energy storage cluster; This represents the scalar value of the actual active power command after the limitation process.

[0187] Optionally, the process of dynamically adjusting the local power output allocation weight of the second type of energy storage cluster needs to comprehensively consider the current remaining available charge capacity of each second type of energy storage cluster and its real-time thermal management heat dissipation margin to prevent the risk of secondary overload during the compensatory power output process. The physical operating limits are strictly defined here as the maximum continuous operating current and the maximum operating temperature that a single energy storage cluster can withstand.

[0188] For example, the formula for calculating dynamically assigned weights is:

[0189] ;

[0190] in, This indicates that the target is specified and the index number is... The second type of energy storage cluster (satisfying) The calculated real number of the dynamic allocation weight for compensation power; Indicates the first The current state of charge percentage (SOC) of a Class II energy storage cluster is estimated by a combination of ampere-hour integration and Kalman filtering. A constant representing the upper limit of the maximum operating temperature specified in the physical operating limits; Indicates the first The current actual battery surface temperature monitoring value of a Class II energy storage cluster; The natural number variable represents the second type of energy storage cluster specifically used as a loop index in the summation function, traversing... All available clusters in the set. The numerator characterizes the overall regulation margin of a single healthy energy storage cluster: the lower the state of charge, i.e., The larger the value, the larger the rechargeable absorption space, or the more reverse-like correction coefficient is used under discharge conditions. The further the current actual temperature is from the limit temperature, the more sufficient the heat dissipation margin, the stronger the ability of the energy storage cluster to share and compensate for the power gap, and the higher the allocation weight.

[0191] For example, the formula for the final additional power allocation target executed by the second type of energy storage cluster is:

[0192] ;

[0193] in, Indicates the first The actual physical active power command value that a Class II energy storage cluster ultimately needs to output or absorb; This represents the original base power command value of the cluster before degradation occurred; This represents the real number representing the dynamic allocation weight of the compensation power calculated above; This represents the total power deficit scalar calculated above. In the actual control link, the energy management system will send these recalculated asymmetric commands to the bidirectional DC / DC converter control motherboards at the back end of each corresponding energy storage cluster via industrial Ethernet. Through the above-mentioned soft algorithm-level reconstruction, this solution realizes intelligent spatial transfer of internal response loads, ensuring that the energy storage power station can still stably output sufficient supporting power to the grid side even when some faulty units are limited.

[0194] Phase 4: Physically mandated compensation control closed loop for the first type of energy storage cluster.

[0195] Specifically, the method steps provided in this embodiment further include: synchronously triggering a physical forced compensation step for the first type of energy storage cluster, including increasing the flow rate of the liquid cooling pipeline of the branch where the first type of energy storage cluster is located to a preset maximum rated flow rate, and opening the active balancing loop at a preset maximum balancing operating frequency, thereby suppressing internal thermal risks through forced physical heat dissipation and charge balancing actions.

[0196] For example, in addition to reducing the power input of the first type of energy storage cluster at the pure software algorithm level, the system also activates the underlying hardware defense intervention mechanism. In the thermal management architecture of a large containerized energy storage system, the liquid cooling system typically includes an external main refrigeration compressor circuit and parallel liquid cooling branches entering each energy storage cluster. Each branch is equipped with an electronic proportional regulating valve independently driven by the control board. On the other hand, the active balancing circuit is a high-frequency power electronic conversion network, such as a bidirectional flyback converter or a switched capacitor array, installed on the slave control board of the battery management system. It is specifically used for high-current charge transfer between adjacent individual cells.

[0197] It is also important to note that increasing the flow rate of the liquid cooling pipeline in the branch containing the first type of energy storage cluster to the preset maximum rated flow rate means that regardless of the value returned by the surface temperature sensor of the first type of energy storage cluster, the control board of the liquid cooling system will overstep its authority and ignore the conventional proportional-integral-derivative (PID) temperature closed-loop tracking algorithm. The system will directly send a full-scale duty cycle pulse width modulation signal to the electronic proportional control valve of the branch corresponding to the first type of energy storage cluster.

[0198] For example, its electronic control valve opening duty cycle output matrix is:

[0199] ;

[0200] in, Indicates that it is issued to the first The duty cycle value of the electronic regulating valve opening of each energy storage cluster corresponding to the cooling branch; and These represent the first and second type energy storage clusters, respectively; 1.0 represents the fully open undamped constant of the electronic proportional control valve when it is in the preset maximum rated flow state; This represents the proportional-integral-derivative abstract control function under conventional liquid-cooled logic. Indicates the first Real-time ambient temperature measurement points for each energy storage cluster; This indicates the target set temperature for thermal management. By physically forcibly intervening through the control logic to fully open the valves of specific branches, the system forcibly removes localized abnormal heat accumulation in individual units that may be masked by data blind spots, using the maximum flow rate of the cooling medium (such as ethylene glycol aqueous solution), thus avoiding individual unit thermal runaway by utilizing extreme fluid convection heat transfer.

[0201] Optionally, a control mechanism that activates the active balancing loop at a preset maximum balancing operating frequency is used primarily to prevent severe charge inconsistency between battery cells caused by poor data quality. When a cluster is in the first type of energy storage cluster state, because the system can no longer accurately observe the charge and discharge depth of individual cells, there is a high risk of individual high internal resistance cells being hidden overcharged or over-discharged at the end of the charge and discharge cycle.

[0202] For example, the instruction setting mechanism for its active equalization switching frequency is as follows:

[0203] ;

[0204] in, Indicates the first The final operating switching frequency value of the active balancing circuit within each energy storage cluster; This represents the preset maximum equalization operating frequency constant allowed by the design of magnetic components in the hardware circuit. At this frequency, the equalization inductor or transformer is at the critical saturation edge, and the equalization current transmission power reaches the physical limit of the circuit. This represents a dynamic frequency adjustment function based on the differential pressure between individual cells as the input variable. Indicates the first The maximum voltage range of individual cells within each energy storage cluster is monitored in real time. This step abandons the conventional gentle balancing logic that is activated on demand, and directly uses the highest frequency and highest intensity continuous charge transfer to forcibly smooth out potentially dangerous voltage spikes within the energy storage cluster, effectively blocking the source risk link of local overcharging-induced thermal runaway from the underlying electrical characteristics.

[0205] In summary, existing energy storage power stations typically employ a highly rigid and globally consistent safety degradation strategy. This means that when a very small number of battery cells within the power station experience communication failures or a decrease in data confidence, the entire energy management system, prioritizing safety, will forcibly reduce the rated power of the entire station or even directly execute AC-side disconnection from the grid. This crude control not only results in significant economic waste of energy storage assets but also causes the energy storage system to lose its continuity and reliability when participating in critical grid frequency regulation during primary or secondary frequency regulation periods.

[0206] The technical solution described in this embodiment transforms the global unified degradation thinking into precise asymmetric cluster-level spatial scheduling. By dynamically quantifying and evaluating the prediction uncertainty of each cluster, the solution implements deep power current limiting and nonlinear reduction of the safety alarm threshold for the first type of energy storage cluster with damaged data at the software algorithm level; simultaneously, at the physical hardware level, it applies liquid cooling circulation with extreme duty cycles and the highest frequency active equalization to force underlying physical suppression, thus constructing a highly reliable defense barrier to cope with the risk of data blind spots in digital twins;

[0207] Meanwhile, the solution introduces a dynamic allocation weight based on temperature rise margin and charge capacity margin on the undisturbed second type of energy storage cluster, performing highly accurate power compensation. This series of multi-field coupling technologies effectively overcomes the problem of global shutdown caused by local failures leading to the paralysis of the entire energy storage system due to the anomaly of a single physical node in complex power grid interactions. It significantly improves the equipment fault tolerance rate throughout the entire life cycle of the energy storage system and its continuous collaborative support capability to the external power grid under extreme conditions such as local communication paralysis, demonstrating strong adaptability to industrial implementation environments and outstanding technical application value.

[0208] Example 4:

[0209] like Figure 7 As shown, this embodiment provides a digital twin safety assessment and prediction system for the entire lifecycle of an energy storage system. As is known from existing technologies, current energy storage safety assessment systems typically treat energy storage power stations as independent physical entities for monitoring, relying primarily on single-dimensional hardware operation data and employing fixed static safety thresholds for risk warnings. When energy storage systems are deeply involved in high-frequency coordinated scheduling tasks such as grid frequency regulation and backup support, this static assessment architecture can lead to conflicts between operational safety boundaries and coordinated response efficiency. If conservative thresholds are used, the system will experience frequent unplanned shutdowns; if the thresholds are relaxed, there is a lack of quantitative methods to understand the cross-domain coupling patterns between underlying physical characteristics and grid dispatch commands, which can easily induce serious accidents such as thermal runaway.

[0210] To address the aforementioned issues, the system provided in this embodiment relies on a physical hardware architecture network consisting of an edge computing layer, a cloud server layer, and a grid-side dispatch center to achieve a dynamic balance between the safety and efficiency of the energy storage system.

[0211] Specifically, the digital twin safety assessment and prediction system for the entire life cycle of an energy storage system described in this embodiment is characterized by comprising: a data acquisition and fusion module, a hybrid model construction module, a risk assessment and prediction module, a collaborative decision-making and early warning module, and a degradation control and feedback closed-loop module.

[0212] The data acquisition and fusion module acquires full lifecycle data from the energy storage side and collaborative data from the grid side, performs three-dimensional quality grading to obtain data quality levels, and weights and fuses the full lifecycle data from the energy storage side and the collaborative data from the grid side based on cross-domain coupling coefficients to generate a five-dimensional risk collaborative profile. In actual hardware deployment, this module mainly relies on the underlying hardware of the Supervisory Control and Data Acquisition (SCADA) system and the Battery Management System (BMS) deployed inside the energy storage container. The acquisition of full lifecycle data from the energy storage side depends on NTC thermistors, high-precision analog-to-digital converters (ADCs), and Hall current sensors installed on the battery modules. These hardware components transmit real-time physical quantities to the local edge computing gateway via a CAN bus. The acquisition of collaborative data from the grid side relies on remote communication units (RTUs), which receive Automatic Generation Control (AGC) and Automatic Voltage Control (AVC) commands from the grid dispatch center via IEC 61850 or Modbus TCP industrial Ethernet protocols. The data cleaning microprocessor built into the edge computing gateway executes the three-dimensional quality grading algorithm to identify and rate sensor signals with communication packet loss or glitches caused by electromagnetic interference. Based on the cross-domain coupling coefficients obtained from the computing unit, this module imports data streams distributed in different physical spaces and different sampling frequency dimensions into the system's high-speed cache for tensor splicing, thereby generating a standardized five-dimensional risk collaborative profile matrix.

[0213] A hybrid model construction module is used to construct a five-field coupled physical mechanism model of heat, electricity, aging, insulation, and equilibrium. The cross-domain coupling coefficient is introduced as a model parameter correction term to quantify the impact of coordinated actions on risk. A data-driven prediction model is generated through reinforcement learning training using preset virtual accident samples. The five-field coupled physical mechanism model and the data-driven prediction model are then fused to form a hybrid model. In practical applications, because the five-field coupled physical mechanism model involves the real-time solution of a large number of nonlinear ordinary differential equations, and the data-driven prediction model involves matrix multiplication and addition operations of deep neural networks, this module is mainly deployed in a cloud-based high-performance computing cluster equipped with a graphics processing unit (GPU) or tensor processor (TPU) array. The parameterized thermal, electrical, and aging mechanism models characterize the heat transfer coefficient of the liquid-cooled pipes, the ohmic internal resistance decay rate of the cell separator, and the ground impedance decay curve of the insulation monitoring instrument in the actual energy storage cabinet. By introducing the cross-domain coupling coefficient, the hybrid model construction module directly maps the megawatt-level power fluctuation commands issued by the power grid dispatching platform to the boundary conditions of the internal heating power of the battery. In the offline training zone of the cloud server, the server uses a pre-set virtual accident sample dataset to update the gradient of the reinforcement learning algorithm; in the online running zone, the trained and solidified network weights are distributed to the AI ​​acceleration cards of the edge computing nodes to form a hybrid inference architecture in which physical equations and neural networks run concurrently.

[0214] The risk assessment and prediction module calculates the risk coupling coefficient between core risks using mutual information entropy. Based on this risk coupling coefficient, a three-dimensional mapping rule composed of the data quality level and the collaborative demand level, it dynamically adjusts the early warning threshold and outputs multi-dimensional risk classification results and risk evolution prediction results through the hybrid model. This module, as the core logic operation unit of the system, runs its software process in the main control cabinet of the energy management system (EMS) of the energy storage power station. In physical systems, risks often exhibit a chain-like transmission effect (e.g., a loose connection in a single-unit voltage sampling harness leads to an imbalance, which in turn triggers local overcharging and thermal runaway). This module uses parallel threads of the central processing unit to calculate the mutual information entropy in real time, representing the correlation between the time series of different physical quantities in the hardware. When the calculation shows that the current local data quality level of the energy storage power station is poor (e.g., some sensor communication is interrupted), and the grid-side collaborative demand level is in an emergency state (e.g., the grid faces frequency drops and requires mandatory support), the logic controller of this module will issue dynamically modified threshold parameter instructions to the system's alarm relays and software early warning interfaces according to the three-dimensional mapping rule, avoiding frequent false triggering trip signals generated by traditional static comparator circuits during data fluctuations.

[0215] The collaborative decision-making and early warning module is used to establish a preset three-dimensional cost model based on the multi-dimensional risk classification results. With cost minimization as the optimization objective, it solves the three-dimensional cost model to obtain a combined decision scheme and outputs a quantitatively interpretable early warning report containing the grid collaborative traceability link. At the system device level, this module directly interfaces with the power plant's operation and maintenance dispatch workstation and monitoring screen. Various economic parameters required for the three-dimensional cost model (including electricity settlement price, equipment depreciation rate, default penalty rate, etc.) are stored in a relational database. This module uses a heuristic optimization algorithm solver to find the Pareto optimal solution for comprehensive economic and safety costs within the computation node. The generated combined decision scheme is transformed into specific equipment control words, such as generating an active power setpoint for a specific energy storage converter (PCS) or a speed setpoint for the liquid-cooled unit compressor. Simultaneously, the generated quantitatively interpretable early warning report is pushed to the mobile terminals of operation and maintenance personnel and the monitoring interface in the central control room via a web server. The report uses graphs to illustrate how grid commands drive the physical traceability link of increased polarization within the battery, providing an intuitive basis for manual intervention.

[0216] The degradation control and feedback closed-loop module is used to trigger a degradation mode to dynamically correct the warning threshold and system power boundary when the detected data quality confidence level is lower than a preset confidence threshold or the model prediction uncertainty is higher than a preset uncertainty threshold. It then executes the combined decision-making scheme based on a three-level decision-making authority allocation system at the edge layer, cloud, and grid side, and receives feedback instructions to form a decision closed loop. This module is a key hub connecting digital spatial prediction results with physical spatial hardware execution. In severe operating conditions such as local sensor data anomalies or model confidence collapse, this module intervenes in the operation of underlying devices through a programmable logic controller (PLC) or microcontroller unit (MCU). When executing edge layer decisions, the local controller directly controls the opening of the electronic proportional control valve on the liquid cooling circuit, or activates the bidirectional active balancing circuit of the battery management system through an isolated driver chip to perform physical heat dissipation and charge transfer at the highest frequency. When executing cloud-based decisions, the system sends a stepped-down power output limit command to the digital signal processor (DSP) of the energy storage converter via the industrial communication network. This reduces the charging and discharging current at the physical level by changing the duty cycle of the drive pulse width modulation (PWM) of the insulated gate bipolar transistor (IGBT). After receiving the degraded energy storage availability status, the grid-side dispatch system corrects the global grid dispatch strategy and sends the updated background data back to the system's cloud server, thus constructing a complete control and feedback physical closed loop spanning the equipment layer, station control layer, and grid dispatch layer.

[0217] In summary, the digital twin safety assessment and prediction system for the entire lifecycle of energy storage systems provided in this embodiment breaks down information silos between the energy storage physical control domain and the grid dispatch domain through data acquisition and cross-domain fusion from multiple hardware devices. Relying on a hybrid model deployed in a cloud-edge collaborative architecture and a three-dimensional cost optimization algorithm, the system effectively solves the technical problem of traditional energy storage power stations being unable to simultaneously achieve both safety and response efficiency in complex interactive environments.

[0218] Especially under severe operating conditions where data distortion or loss occurs in the underlying sensor hardware, this system avoids the crude control method of triggering a global forced disconnection of the system-level circuit breaker, as is common in existing technologies. Instead, it executes a precise local degradation and physical forced compensation mechanism through a three-level linkage architecture. This allows the system to maximize the preservation of the operational capabilities of healthy hardware units while strictly adhering to the underlying core electrochemical and thermodynamic safety boundaries. This ensures the continuous and coordinated support of energy storage assets to the external power grid, significantly improving the operational economic benefits and system-level risk resistance of large-scale energy storage power stations.

[0219] Example 5:

[0220] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0221] like Figure 8 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0222] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0223] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0224] The memory 103 stores a computer program corresponding to the digital twin safety assessment and prediction method for the entire lifecycle of an energy storage system according to the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0225] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 8 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0226] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for digital twin-based safety assessment and prediction of an energy storage system throughout its entire lifecycle, characterized in that, include: Acquire full lifecycle data of the energy storage side and collaborative data of the grid side, perform three-dimensional quality grading to obtain data quality level, and perform weighted fusion of the full lifecycle data of the energy storage side and the collaborative data of the grid side based on cross-domain coupling coefficient to generate a five-dimensional risk collaborative profile; A five-field coupled physical mechanism model of heat-electricity-aging-insulation-equilibrium is constructed. The cross-domain coupling coefficient is introduced as a model parameter correction term to quantify the impact of coordinated actions on risk. A data-driven prediction model is generated by reinforcement learning training combined with preset virtual accident samples. The five-field coupled physical mechanism model and the data-driven prediction model are then fused to form a hybrid model. The risk coupling coefficient between core risks is calculated by mutual information entropy. The early warning threshold is dynamically adjusted based on the three-dimensional mapping rule composed of the risk coupling coefficient, the data quality level and the collaborative requirement level. The hybrid model outputs multi-dimensional risk classification results and risk evolution prediction results. Based on the multi-dimensional risk classification results, a preset three-dimensional cost model is established. The three-dimensional cost model is solved with the goal of minimizing costs to obtain a combined decision scheme and output a quantitatively interpretable early warning report that includes the power grid collaborative traceability link. When the data quality confidence level is detected to be lower than the preset confidence threshold or the model prediction uncertainty is higher than the preset uncertainty threshold, the degraded mode is triggered to dynamically correct the warning threshold and the system power boundary. The combined decision-making scheme is executed based on the three-level decision-making authority allocation system of edge layer, cloud and grid side, and feedback instructions are received to form a decision closed loop.

2. The method according to claim 1, characterized in that, The steps of performing three-dimensional quality grading to obtain data quality levels and weighted fusion based on cross-domain coupling coefficients include: A first evaluation model is constructed for the energy storage side's full life cycle data, which includes completeness, accuracy, and consistency. A second evaluation model is constructed for the grid side's collaborative data, which includes real-time performance, accuracy, and continuity. The corresponding data quality levels are output respectively. Calculate the mutual information entropy between the grid-side collaborative data and the energy storage risk characteristics, and use the mutual information entropy as the cross-domain coupling coefficient for the weighted fusion.

3. The method according to claim 1, characterized in that, The introduction of the cross-domain coupling coefficient as a model parameter correction term to quantify the impact of collaborative actions on risk includes: The power grid power command and frequency deviation parameters are incorporated into the heat transfer equation and equilibrium energy transfer model of the five-field coupled physical mechanism model; The temperature rise increment caused by grid coordination is calculated, and the temperature rise increment is equal to the continuous product of the preset proportional coefficient, the energy storage power adjustment amount and the cross-domain coupling coefficient.

4. The method according to claim 1, characterized in that, The step of combining preset virtual accident samples with reinforcement learning training to generate a data-driven prediction model includes: The preset boundary condition parameters are input into the five-field coupled physical mechanism model to simulate risk evolution and generate the virtual accident sample containing the time series of power grid parameters, risk characteristic trajectory and cross-domain coupling coefficient changes; After validating the virtual accident samples, the virtual accident samples are concatenated with real collaborative scenario accident samples to construct a training set. A few-shot reinforcement learning algorithm is used to train the neural network to obtain the data-driven prediction model that meets the preset collaborative requirements.

5. The method according to claim 1, characterized in that, The dynamic adjustment of the early warning threshold based on the three-dimensional mapping rule composed of the risk coupling coefficient, the data quality level, and the collaboration requirement level includes: The coupling strength coefficients between thermal runaway, insulation breakdown, overcharging and over-discharging, imbalance anomaly, inter-cluster inconsistency and fire core risk of energy storage system are calculated by mutual information entropy, and are used as the risk coupling coefficients. Under the condition that the risk coupling coefficient is greater than or equal to the first preset threshold and the data quality level is the preset optimal quality level, if the collaboration requirement level is the emergency collaboration level, then the sensitivity of the early warning threshold is increased. If the coordination requirement level is a fault coordination level, then the sensitivity of the early warning threshold is lowered.

6. The method according to claim 1, characterized in that, The multi-dimensional risk classification results include a degree of synergistic correlation, and the quantification steps for the degree of synergistic correlation include: The probability value of the current risk being triggered by the coordinated action of the power grid is calculated using conditional probability. The probability value is the degree of coordination, and its value is equal to the quotient obtained by dividing the risk increment caused by the coordinated action of the power grid by the total risk increment. Based on the magnitude of the probability value, the degree of synergistic association is classified into three levels: strong association risk, medium association risk, and weak association risk.

7. The method according to claim 1, characterized in that, The process of establishing a preset three-dimensional cost model based on the multi-dimensional risk classification results, and solving the three-dimensional cost model with cost minimization as the optimization objective to obtain a combined decision scheme includes: Construct the three-dimensional cost model that includes the cost of missed reporting, the cost of false reporting, the penalty cost for not meeting the grid coordination requirements, the operating depreciation cost, and the cost of tracing and repairing the source, and dynamically adjust the weight coefficients of the above costs according to the preset normal coordination level, emergency coordination level, or fault coordination level. The three-dimensional cost model is solved using a non-dominated sorting genetic algorithm with collaborative constraints and a Bayesian optimization algorithm to generate the combined decision scheme that includes the early warning threshold, liquid cooling configuration parameters, balance control strategy, and power grid collaborative adjustment suggestions.

8. The method according to claim 1, characterized in that, When the detected data quality confidence level is lower than a preset confidence threshold or the model prediction uncertainty is higher than a preset uncertainty threshold, a degradeable mode is triggered to dynamically correct the warning threshold and the system power boundary, including: When the data quality confidence level is detected to be lower than the second preset threshold, the time series of the full life cycle data is interrupted, or the model prediction uncertainty is higher than the third preset threshold, the degradeable mode is triggered. In normal collaborative scenarios, the thermal runaway safety threshold in the aforementioned warning threshold is lowered; In emergency collaborative scenarios or fault collaborative scenarios, the thermal runaway safety threshold and the upper limit of energy storage power output in the system power boundary are quantitatively reduced in a stepwise manner.

9. The method according to claim 8, characterized in that, The combined decision-making scheme is executed based on a three-level decision-making authority allocation system at the edge layer, cloud, and power grid side, and feedback instructions are received to form a decision-making closed loop, including: The edge layer is responsible for executing liquid cooling parameter and equalization loop control decisions in conventional collaborative scenarios, as well as local degradeable optimization execution when data quality is abnormal; The cloud is responsible for executing power limiting decisions in emergency or fault-based collaborative scenarios, and for executing overall coordination decisions when the cross-domain coupling coefficient is greater than or equal to the fourth preset threshold. After receiving the quantified interpretable early warning report and the power grid coordinated adjustment suggestions, the power grid side sends dispatch instructions to the cloud to iteratively update the hybrid model.

10. A digital twin safety assessment and prediction system for the entire lifecycle of an energy storage system, characterized in that, include: The data acquisition and fusion module is used to acquire full life cycle data of the energy storage side and collaborative data of the grid side, perform three-dimensional quality classification to obtain data quality level, and perform weighted fusion of the full life cycle data of the energy storage side and the collaborative data of the grid side based on the cross-domain coupling coefficient to generate a five-dimensional risk collaborative profile. The hybrid model construction module is used to construct a five-field coupled physical mechanism model of heat-electricity-aging-insulation-equilibrium. The cross-domain coupling coefficient is introduced as a model parameter correction term to quantify the impact of coordinated actions on risk. The module combines preset virtual accident samples to perform reinforcement learning training to generate a data-driven prediction model. The five-field coupled physical mechanism model and the data-driven prediction model are then fused to form a hybrid model. The risk assessment and prediction module is used to calculate the risk coupling coefficient between core risks through mutual information entropy, dynamically adjust the early warning threshold based on the three-dimensional mapping rule composed of the risk coupling coefficient, the data quality level and the collaborative requirement level, and output multi-dimensional risk classification results and risk evolution prediction results through the hybrid model. The collaborative decision-making and early warning module is used to establish a preset three-dimensional cost model based on the multi-dimensional risk classification results, solve the three-dimensional cost model with cost minimization as the optimization objective to obtain a combined decision scheme, and output a quantitatively interpretable early warning report containing the power grid collaborative traceability link. The degradation control and feedback closed-loop module is used to trigger a degradation mode to dynamically correct the warning threshold and system power boundary when the data quality confidence is detected to be lower than the preset confidence threshold or the model prediction uncertainty is higher than the preset uncertainty threshold. It also executes the combined decision-making scheme based on the three-level decision-making authority allocation system of edge layer, cloud and grid side, and receives feedback instructions to form a decision closed loop.

Images