A novel energy storage system operation optimization and safety risk early warning method

Abstract

The application discloses a novel energy storage system operation optimization and safety risk early warning method, and relates to the technical field of energy storage system operation control. The specific implementation scheme is as follows: the discharge operation data of each module and the coupling connection relationship are synchronously acquired to construct a space-time graph, and a pre-trained space-time graph neural network is input to predict a state; an electrochemical abuse, thermal runaway and capacity attenuation risk index are determined and weighted into a comprehensive risk index; the comprehensive risk index is mapped into a charging and discharging power upper limit according to a preset mapping rule and a feasible region is contracted; a multi-objective optimization considering economy and safety is solved and operation is controlled in the feasible region; when the comprehensive risk index continuously exceeds a pre-warning threshold, early warning information and disposal instructions are generated and executed immediately; after disposal, the deviation between the adjusted state prediction result and the actual operation data is used as a correction amount to correct the space-time graph neural network and the preset mapping rule. Through the application, the risk can be early warned and the power can be dynamically constrained, so that closed-loop correction of early warning and operation optimization is realized.

Description

Technical Field

[0001] This application relates to the field of energy storage system operation control technology, and more specifically, to a novel method for energy storage system operation optimization and safety risk early warning. Background Technology

[0002] New energy storage power stations serve as peak-valley arbitrage, demand management, and backup power sources in industrial and commercial parks and new energy power generation facilities. In energy storage systems centered on lithium iron phosphate batteries, multiple battery modules are connected in series and parallel to form battery clusters. These modules exchange heat through shared cooling circuits and the thermal conductivity of the enclosure, and also transmit current through electrical connections, resulting in mutual influence on their operating states. During peak discharge periods, energy storage power stations need to continuously output high power according to the dispatch plan. Under discharge conditions, cell polarization intensifies, heat generation concentrates, and four safety issues—overcurrent, overvoltage, rapid temperature rise, and accelerated capacity decay—are most likely to surface during discharge. If not addressed promptly, these issues may induce thermal runaway and spread along the coupling paths between modules. Therefore, it is necessary to proactively predict and manage the safety risks of each battery module while ensuring discharge benefits.

[0003] Existing technologies mostly employ safety monitoring schemes based on battery management systems, which involve real-time acquisition of the voltage, current, and temperature of the battery clusters, comparing the acquired values ​​with fixed thresholds, and outputting alarm signals and cutting off the circuit when the limits are exceeded. Additionally, Chinese patent applications disclose economic dispatch schemes for energy storage power stations, which involve developing an overall charging and discharging plan for the battery clusters based on peak-valley time-of-use electricity prices, charging during off-peak hours and discharging during peak hours according to the plan to profit from the price difference.

[0004] However, threshold comparison-based monitoring can only trigger alarms when parameters have exceeded limits, failing to provide advance risk predictions for future periods. Scheduling methods that plan charge / discharge cycles based on the entire battery cluster treat each battery module as a homogeneous entity, ignoring risk propagation and state differences caused by thermal coupling and electrical connections between modules. Safety constraints are not dynamically adjusted based on the actual risks of each module. Furthermore, safety monitoring and economic scheduling are independent; after early warning and response actions are executed, the prediction model and constraint rules are not updated based on the actual response results, leading to repeated occurrences of the same deviations in subsequent operations. Therefore, how to predict the operating status of each battery module in advance, quantify the safety risks under the influence of inter-module coupling, transform the risk quantification results into dynamic power constraints, and achieve closed-loop correction between early warning and response actions and operational optimization are urgent technical problems to be solved in this field. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a novel method for optimizing the operation and early warning of safety risks of energy storage systems. This method can achieve both operational safety and economic efficiency while ensuring the operation of the energy storage system, predict the operating status of each battery module in advance and quantify safety risks, and realize closed-loop correction of early warning and operation optimization.

[0006] According to the first aspect of this application, a novel method for optimizing the operation and providing early warning of safety risks in an energy storage system is provided, comprising: S100: Synchronously acquire discharge operation data of each battery module in the energy storage system and the coupling connection relationship between each battery module. Construct a spatiotemporal graph with each battery module as a node and the coupling connection relationship as a connection edge. Input the spatiotemporal graph and the discharge operation data of each battery module into a pre-trained spatiotemporal graph neural network to obtain the state prediction results of each battery module within a preset time period. The state prediction results include the voltage, current, temperature and state of charge of each battery module within a preset time period; S200: Based on the state prediction results of each battery module, determine the electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index of each battery module, and then perform a weighted summation to determine the comprehensive risk index. S300: According to the preset mapping rules, the comprehensive risk index of each battery module is mapped to the upper limit of the charging and discharging power of each battery module, and the upper limit of the charging and discharging power obtained by mapping is used to narrow the feasible domain of the charging and discharging power of each battery module. S400: With operational economy and operational safety as optimization objectives, a multi-objective optimization problem is solved within the feasible domain of charge and discharge power after each battery module is shrunk, and an operational optimization strategy is obtained. The operation of the energy storage system is controlled according to the operational optimization strategy. S500: When the comprehensive risk index of any battery module exceeds the preset warning threshold for the first time in a preset number of consecutive prediction cycles, a warning message and a handling instruction are generated and the handling instruction is executed immediately. S600: After executing the disposal command, adjust the state prediction result of the corresponding battery module according to the disposal command. Use the deviation between the adjusted state prediction result and the actual operating data of each battery module as the correction amount. Correct the network parameters of the spatiotemporal graph neural network and the preset mapping rules according to the correction amount. Apply the corrected spatiotemporal graph neural network and the preset mapping rules to the next optimization cycle.

[0007] In some embodiments, the discharge operation data includes the voltage, current, temperature and state of charge of each battery module, and the coupling connection relationship includes the thermal coupling path and the electrical connection path between battery modules, wherein the thermal coupling path is the heat transfer path formed by each battery module through a shared cooling circuit and the heat conduction structure of the housing, and the electrical connection path is the series and parallel electrical connection path of each battery module within the battery cluster.

[0008] In some embodiments, the specific method for constructing a spatiotemporal graph with each battery module as a node and the coupling connection relationship as the connection edge is as follows: A thermal coupling edge is established between battery module pairs directly connected via a thermal coupling path. The edge weight of the thermal coupling edge is determined according to the thermal resistance between the battery module pairs. The thermal resistance is determined according to the length of the heat transfer path between the battery module pairs and the thermal conductivity of the heat-conducting medium. An electrical connection edge is established between battery module pairs directly connected via an electrical connection path. The edge weight of the electrical connection edge is determined according to the connection impedance between the battery module pairs. A single-period graph structure is formed by the thermal coupling edge and the electrical connection edge. The single-period graph structure with a continuously preset number of sampling periods is arranged in chronological order to form a spatiotemporal graph. The node characteristics of each battery module node in the spatiotemporal graph are the discharge operation data of the corresponding battery module in the corresponding sampling period.

[0009] In some embodiments, the pre-training method of the spatiotemporal graph neural network is as follows: acquiring historical data of the discharge operation data in the energy storage system, constructing an electro-thermal coupling simulation model of the battery module, simulating the operation process of each battery module under normal and abnormal operating conditions using the electro-thermal coupling simulation model, generating simulation operation data, and jointly training the spatiotemporal graph neural network with the historical data of the discharge operation data and the simulation operation data, wherein the spatiotemporal graph and discharge operation data corresponding to each historical sampling time are used as training inputs, and the voltage, current, temperature and state of charge of each battery module within a preset time after each historical sampling time are used as training labels; The spatiotemporal graph neural network includes a graph convolutional layer and a gated temporal layer. The graph convolutional layer aggregates the discharge operation data of each battery module node according to the connection edges of the spatiotemporal graph and extracts the spatial coupling features between each battery module. The gated temporal layer extracts the time dependence features of the operating state of each battery module according to the time sequence of the sampling period.

[0010] In some embodiments, the step of determining the electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index of each battery module based on the state prediction results of each battery module, and then performing a weighted summation to determine the comprehensive risk index, is specifically carried out by determining the electrochemical abuse risk index based on the deviation of the voltage of each battery module from the preset voltage safety boundary in the state prediction results, and the deviation of the current of each battery module from the preset current safety boundary in the state prediction results. The thermal runaway risk index is determined based on the temperature of each battery module in the state prediction results and the temperature difference between battery modules connected by thermal coupling paths. Based on the rate of change of state of charge and depth of charge / discharge of each battery module in the state prediction results, the capacity decay risk index is determined. The electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index were normalized and then weighted and summed to obtain the comprehensive risk index. The sum of the weight coefficients of the weighted summation was 1.

[0011] In some embodiments, the comprehensive risk index of each battery module is mapped to the upper limit of the charging and discharging power of each battery module according to a preset mapping rule, and the feasible domain of the charging and discharging power of each battery module is narrowed by the mapped upper limit of the charging and discharging power. Specifically, the upper boundary of the feasible domain of the charging and discharging power of each battery module is initially the rated charging and discharging power of the corresponding battery module. The preset mapping rule divides the value range of the comprehensive risk index into multiple consecutive risk intervals. Each risk interval corresponds to a charging and discharging power upper limit level. The union of each risk interval covers the entire value range of the comprehensive risk index. The upper limit of charge and discharge power corresponding to the risk range of the comprehensive risk index of each battery module is taken as the upper limit of charge and discharge power of the corresponding battery module. The upper boundary of the feasible region of charge and discharge power is replaced by the upper limit of charge and discharge power obtained by mapping, while the lower boundary of the feasible region of charge and discharge power remains unchanged, thus obtaining the contracted feasible region of charge and discharge power.

[0012] In some embodiments, the optimization objectives are operational economy and operational safety. A multi-objective optimization problem is solved within the feasible domain of the charging and discharging power after the shrinkage of each battery module to obtain an operational optimization strategy. The operation of the energy storage system is controlled according to the operational optimization strategy. Specifically, the charging and discharging revenue under the peak-valley electricity price difference within a preset time period is used as the operational economy indicator, and the weighted sum of the mean and maximum values ​​of the comprehensive risk index of all battery modules is used as the operational safety indicator. The sum of the weight coefficients of the weighted sum is 1. Within the feasible domain of the charge and discharge power of each battery module after shrinkage, a multi-objective optimization problem is solved with the optimization objectives of operational economic indicators and operational safety indicators. The charge and discharge power plan of each battery module, the balancing strategy between battery modules and the cooling control parameters are obtained as operational optimization strategies. The charging and discharging power of each battery module is controlled according to the charging and discharging power plan, the power distribution between battery modules is adjusted according to the balancing strategy, the operating status of the cooling system is adjusted according to the cooling control parameters, and the multi-objective optimization problem is solved and the operation optimization strategy is updated in a rolling manner according to the preset optimization cycle.

[0013] In some embodiments, when the comprehensive risk index of any battery module exceeds a preset warning threshold for the first time for a preset number of consecutive prediction cycles, a warning message and a handling instruction are generated and the handling instruction is executed immediately. Specifically, the preset warning threshold includes a prompt threshold, a warning threshold, and an alarm threshold, wherein the prompt threshold is lower than the warning threshold and the warning threshold is lower than the alarm threshold. When the comprehensive risk index of any battery module exceeds the prompt threshold for the first time in a preset number of prediction cycles but does not exceed the warning threshold, a prompt-level early warning message is generated, and a handling instruction is generated to maintain the current operation optimization strategy and strengthen the monitoring of the battery module's status. When the comprehensive risk index of any battery module exceeds the warning threshold for the first time in a preset number of prediction cycles but does not exceed the alarm threshold, a warning-level early warning message is generated, and a handling instruction to reduce the charge and discharge rate of the corresponding battery module is generated. When the comprehensive risk index of any battery module exceeds the alarm threshold for the first time in a preset number of prediction cycles, an alarm-level warning message is generated, and a handling instruction is generated to stop the charging and discharging of the corresponding battery module and request manual inspection.

[0014] In some embodiments, after the disposal instruction is executed, the state prediction result of the corresponding battery module is adjusted according to the disposal instruction. The deviation between the adjusted state prediction result and the actual operating data of each battery module is used as the correction amount. The network parameters of the spatiotemporal graph neural network and the preset mapping rules are corrected according to the correction amount. Specifically, the actual operating data of each battery module within a preset number of prediction cycles after the disposal instruction is executed is obtained. The actual operating data includes the actual voltage, actual current, actual temperature and actual state of charge of each battery module. The upper limit of the charge and discharge power of the corresponding battery module after disposal is determined according to the disposal instruction, and the state prediction result of the corresponding battery module is adjusted based on the upper limit of the charge and discharge power after disposal. The deviation between the actual operating data of each battery module and the adjusted state prediction result is calculated for each prediction cycle, and the average deviation of each prediction cycle is used as the correction amount. When the correction amount exceeds the preset correction threshold, the network parameters of the spatiotemporal graph neural network are incrementally updated with the correction amount, and the mapping relationship between the comprehensive risk index and the upper limit of charging and discharging power in the preset mapping rules is shifted and adjusted.

[0015] One embodiment of the above application has the following advantages or beneficial effects: To address the urgent need for safety management in peak-valley arbitrage operations of energy storage power stations, this application's embodiments directly transform the quantified safety risks of each battery module into dynamic contraction constraints of the feasible domain of charging and discharging power. Within the contracted feasible domain, a multi-objective optimization problem is solved with operational economy and operational safety as optimization objectives. This ensures that while the energy storage system obtains revenue from peak-valley electricity price differences, the operating power of each battery module remains within a range that matches its own risk level, thus achieving overall coordinated protection of the operational economy and operational safety of the energy storage system.

[0016] Compared to existing technologies that only alarm after a threshold comparison monitoring limit is exceeded and cluster-level scheduling ignores module differences, the embodiments of this application construct a spatiotemporal graph with each battery module as a node and the coupling connection relationship between modules as the connection edge. The pre-trained spatiotemporal graph neural network is used to predict the state of each battery module within a preset time period, thus shifting safety control from post-event alarm to pre-event prediction. Furthermore, the spatiotemporal graph explicitly depicts the thermal coupling path and electrical connection path, enabling risk quantification to include the impact of inter-module coupling propagation. This allows for the identification of local risks in individual modules that cannot be reflected by cluster-level average indicators.

[0017] To address the issues of disconnect between prediction and response and recurring deviations in existing technologies, this application's embodiments, after executing a response command, use the deviation between the state prediction result adjusted according to the response command and the actual operating data of each battery module as a correction amount. Based on this, the network parameters and preset mapping rules of the spatiotemporal graph neural network are corrected and applied to the next optimization cycle. This allows the prediction model and risk constraint rules to be continuously updated with the actual response effect, forming a closed loop of early warning, response, and correction. This upgrades the traditionally fragmented safety monitoring and operation scheduling to predictive proactive safety management with self-correction capabilities.

[0018] Other effects of the above-mentioned alternative methods will be described below in conjunction with specific embodiments. Attached Figure Description

[0019] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1 This is a flowchart illustrating the novel energy storage system operation optimization and safety risk early warning method provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the change in pre-training loss of a spatiotemporal graph neural network, provided in an embodiment of this application. Figure 3 This is a schematic diagram of the upper limit mapping of charging and discharging power before and after the preset mapping rule is corrected, provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the change of the comprehensive risk index of each battery module over time, as provided in an embodiment of this application. Figure 5 This is a schematic diagram illustrating the change of the upper limit of the charging and discharging power of a battery module M7 over time, as provided in an embodiment of this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] To facilitate understanding of this application, the embodiments of this application will be briefly described below: The solution in this application is applied to a novel energy storage system with lithium iron phosphate battery modules as the core energy storage unit. Under discharge conditions, the internal electrochemical reaction rate of the battery module accelerates, polarization intensifies, and voltage, current, temperature, and state of charge are coupled together: high-current discharge causes Joule heat accumulation, and the increased temperature, in turn, accelerates side reactions and capacity decay. Abnormal terminal voltage and state of charge are direct indicators of electrochemical abuse. Simultaneously, battery modules within the same battery cluster exchange heat through a shared cooling circuit and the heat-conducting structure of the casing, and current is transmitted through series and parallel electrical connection paths. Heat generation and current disturbances in one battery module can propagate to adjacent battery modules along the aforementioned coupling paths, causing safety risks to exhibit spatial propagation and temporal accumulation.

[0022] The aforementioned physical laws dictate that the safe management of energy storage systems must simultaneously utilize the temporal evolution information of module states and the spatial coupling information between modules. Furthermore, operational optimization requires a trade-off between peak-valley electricity price differences and safety margins: higher discharge power yields greater price difference benefits, but also exacerbates module temperature rise and capacity decay. Existing methods lack the quantitative utilization of inter-module coupling relationships and the closed-loop capability to integrate prediction, constraint, optimization, handling, and correction. Therefore, this application constructs a spatiotemporal graph using each battery module as a node and coupling connections as connecting edges. It utilizes a pre-trained spatiotemporal graph neural network to predict the state of each battery module, quantifies three risks—electrochemical abuse, thermal runaway, and capacity decay—and narrows the feasible domain of charge and discharge power. Within this feasible domain, it solves a multi-objective optimization problem that balances economy and safety, providing early warnings and immediate handling for exceeding time limits. After handling, the prediction model and mapping rules are corrected based on measured deviations, thereby continuously improving operational economy while ensuring operational safety.

[0023] See Figure 1 , Figure 1 This is a flowchart illustrating the novel energy storage system operation optimization and safety risk early warning method provided in this application embodiment. Figure 1 The execution subject of the method shown can be a combination of software and / or hardware, specifically, it can be one or more of various types of terminals, hardware systems, cloud computing, etc.

[0024] Figure 1The novel energy storage system operation optimization and safety risk early warning method shown includes steps S100 to S600, as detailed below: S100: Synchronously acquire discharge operation data of each battery module in the energy storage system and the coupling connection relationship between each battery module. Construct a spatiotemporal graph with each battery module as a node and the coupling connection relationship as a connection edge. Input the spatiotemporal graph and the discharge operation data of each battery module into a pre-trained spatiotemporal graph neural network to obtain the state prediction results of each battery module within a preset time period. The state prediction results include the voltage, current, temperature and state of charge of each battery module within a preset time period.

[0025] Specifically, the discharge operation data consists of time-series data of each battery module arranged chronologically under discharge conditions, depicting the real-time operating state of each module. The coupling connections depict the heat and current transfer channels between battery modules. Since the safety risks of battery modules evolve over time and propagate along coupling paths, this step constructs a spatiotemporal graph using each battery module as a node and coupling connections as edges, organizing the dispersed module states into a data whole with both spatial and temporal structures. The pre-trained spatiotemporal graph neural network is a network pre-trained using historical operating data. It extracts spatial coupling features and temporal dependence features from the spatiotemporal graph, and then outputs the state prediction results of each battery module within a preset time period, providing a future-oriented data foundation for subsequent risk quantification. Through this step, discrete module sampling data is transformed into quantitative predictions of operating states for future periods.

[0026] S200: Based on the state prediction results of each battery module, determine the electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index of each battery module, and then perform a weighted summation to determine the comprehensive risk index.

[0027] Specifically, the voltage and current in the state prediction results reflect the degree of electrochemical abuse, the temperature and inter-module temperature difference reflect the risk of thermal runaway, and the rate of change of state of charge and depth of charge / discharge reflect the rate of capacity decay. Since the causes and forms of harm of these three risks differ, this step separately determines the electrochemical abuse risk index, the thermal runaway risk index, and the capacity decay risk index. These are then weighted and summed to form a comprehensive risk index, allowing the safety status of each battery module in the future to be expressed by a single, comparable quantitative indicator, providing a unified basis for subsequent power constraint reduction. Through this step, multi-dimensional state prediction is transformed into a single-dimensional risk quantification result.

[0028] S300: According to the preset mapping rules, the comprehensive risk index of each battery module is mapped to the upper limit of the charging and discharging power of each battery module, and the feasible domain of the charging and discharging power of each battery module is narrowed by the upper limit of the charging and discharging power obtained by mapping.

[0029] Specifically, the feasible charge / discharge power domain is the power range allowed for the battery module during operational optimization, and the upper limit of charge / discharge power is the power amplitude boundary within this range. A pre-defined mapping rule establishes a correspondence between the comprehensive risk index and the upper limit of charge / discharge power; the higher the risk, the lower the mapped upper limit of charge / discharge power. This step transforms the abstract risk quantification result into a specific power boundary according to this correspondence, and replaces the original boundary with the mapped upper limit of charge / discharge power, thus shrinking the feasible charge / discharge power domain. This ensures that battery modules with higher risks can only use lower power in subsequent optimizations, preventing further risk accumulation from a constraint perspective. Through this step, the risk quantification result is transformed into a directly executable power constraint.

[0030] S400: With operational economy and operational safety as optimization objectives, a multi-objective optimization problem is solved within the feasible domain of charge and discharge power after the shrinkage of each battery module to obtain an operational optimization strategy, and the operation of the energy storage system is controlled according to the operational optimization strategy.

[0031] Specifically, operational economics is reflected in the charging and discharging revenue under peak-valley electricity price differences, while operational safety is reflected in the overall level of the comprehensive risk index of all battery modules. Since the revenue objective tends to increase power while the safety objective tends to limit power, these two objectives are inherently conflicting. Therefore, this step treats operational economics and operational safety as parallel optimization objectives. A multi-objective optimization problem is solved within the feasible region of the reduced charging and discharging power for each battery module. The resulting operational optimization strategy provides the power arrangement for each battery module and corresponding inter-module coordination and cooling regulation methods. The energy storage system is then controlled according to the operational optimization strategy. Through this step, the revenue and safety trade-offs under power constraints are transformed into executable control commands.

[0032] S500: When the comprehensive risk index of any battery module exceeds the preset warning threshold for the first time in a preset number of consecutive prediction cycles, a warning message and a handling instruction are generated and the handling instruction is executed immediately.

[0033] Specifically, the prediction period is the time unit for the rolling output of state prediction results, and the comprehensive risk index is updated once in each prediction period. A single exceedance of the preset warning threshold may be caused by instantaneous disturbances; therefore, this step uses the first consecutive exceedance of the preset warning threshold for a preset number of prediction periods as the trigger condition. Warning information and handling instructions are only generated when the risk persists, avoiding false alarms caused by glitches. Handling instructions are executed immediately within the same period they are generated, ensuring no additional delay is introduced between control actions and risk identification. Through this step, persistent risks are transformed into immediately effective handling actions.

[0034] S600: After executing the disposal command, adjust the state prediction result of the corresponding battery module according to the disposal command. Use the deviation between the adjusted state prediction result and the actual operating data of each battery module as the correction amount. Correct the network parameters of the spatiotemporal graph neural network and the preset mapping rules according to the correction amount. Apply the corrected spatiotemporal graph neural network and the preset mapping rules to the next optimization cycle.

[0035] Specifically, the disposal command altered the power boundary of the battery module. The original state prediction results were obtained based on the power conditions before disposal, which no longer aligns with the actual operating conditions after disposal. Therefore, the state prediction results of the corresponding battery module are first adjusted according to the disposal command to align the prediction benchmark with the operating conditions after disposal. Subsequently, the deviation between the adjusted state prediction results and the actual operating data of each battery module is used as a correction amount. The magnitude of the deviation reflects the degree of mismatch between the prediction model and the mapping rules in the real disposal scenario. The network parameters of the spatiotemporal graph neural network and the preset mapping rules are corrected based on the correction amount, and the corrected spatiotemporal graph neural network and preset mapping rules are applied to the next optimization cycle, ensuring that the predictions and constraints of the next cycle are based on actual verification. Through this step, the actual results of a single disposal are transformed into continuous improvement of the model and rules, forming a closed-loop correction.

[0036] This embodiment, through the synergistic effect of the above six steps, upgrades the safety management of energy storage systems from passive alarms after exceeding limits to proactive predictive management: constructing a spatiotemporal diagram to explicitly represent the coupling and propagation relationships between modules; predicting module states to obtain a future-oriented risk perspective; mapping and shrinking the feasible region to quantify risks into power boundaries; rolling multi-objective optimization to balance benefits and safety within constraints; tiered early warning and immediate handling to prevent risk accumulation; and correcting models and rules based on measured deviations to maintain long-term effectiveness. This constitutes a closed-loop logical framework of prediction, constraint, optimization, handling, and correction. This embodiment has the ability to identify three risks in advance: electrochemical abuse, thermal runaway, and capacity decay. It can output decision results combining the charging and discharging power plans of each battery module, inter-module balancing strategies, and cooling control parameters, and automatically correct predictions and constraints after handling, realizing the integration of economic optimization and safety risk early warning for energy storage system operation.

[0037] This embodiment elaborates on the relationship between discharge operation data and coupling connections in detail, specifically as follows: The discharge operation data includes the voltage, current, temperature and state of charge of each battery module. The coupling connection relationship includes the thermal coupling path and the electrical connection path between battery modules. The thermal coupling path is the heat transfer path formed by each battery module through a shared cooling circuit and the heat conduction structure of the housing. The electrical connection path is the series and parallel electrical connection path of each battery module within the battery cluster.

[0038] Specifically, the four components of the discharge operation data each carry different physical information: voltage is the terminal voltage of the battery module, reflecting the external characteristics of cell polarization and remaining capacity; current is the discharge current of the battery module, the direct source of Joule heating; temperature is the body temperature of the battery module, the earliest observable precursor to thermal runaway risk; and state of charge (SCC) is the percentage of remaining capacity of the battery module, a fundamental quantity for scheduling charge and discharge operations and assessing capacity degradation. These four components together constitute the entirety of the subsequent spatiotemporal graph node features, providing a data foundation for state prediction and risk quantification.

[0039] Specifically, the discharge operation data of each battery module is sampled synchronously according to the same sampling period, which can be set to, for example, 10 seconds. The sampling times of each battery module are timestamped and aligned using a unified clock source, with an alignment tolerance of, for example, 1 ms. That is, any two battery modules whose sampling times differ by no more than 1 ms in the same frame are considered to be in the same sampling period. The resulting discharge operation data is time-series data arranged in chronological order. The sampled value of the k-th sampling period corresponds to the same physical moment, ensuring that when a spatiotemporal diagram is constructed with a subsequent preset number of sampling periods, each node is strictly aligned on the same layer of the diagram structure, avoiding the erroneous superposition of unrelated sequential changes at the same moment due to asynchronous sampling. When data packet loss occurs in a certain sampling period, it is supplemented by linear interpolation of the sampled values ​​of the two adjacent frames before and after that sampling period, and the supplemented data carries an interpolation mark for identification by subsequent correction stages.

[0040] This embodiment limits data acquisition and state prediction to discharge conditions because discharge is the condition where the risk of electrochemical abuse and thermal runaway is most concentrated: during discharge, cell polarization intensifies, Joule heating and electrochemical reaction heat overlap, and overcurrent, overvoltage, and abnormal temperature rise first appear in the discharge operation data; while the feasible domain of charge and discharge power and operation optimization retain both charging and discharging because the economic goal of peak-valley arbitrage naturally includes both off-peak charging and peak-peak discharging, and charging capacity must be retained in the optimization variables. If the same prediction link is established for charging conditions, the distribution of conditions in the training data will be significantly wider, and the prediction accuracy of the high-risk scenario of discharge conditions will be diluted. Therefore, this embodiment adopts a hierarchical arrangement of discharge-side prediction and charge-discharge dual-side optimization.

[0041] Specifically, the thermal coupling path is the heat transfer path formed by the shared cooling circuit and the heat-conducting structure of the housing for each battery module: each battery module in the same battery cluster shares a cooling circuit, and the coolant flows sequentially through the liquid cooling plates of each battery module, transferring heat from the upstream module to the downstream module; simultaneously, each battery module conducts heat to the housing wall through the metal end plates and thermal pads within the housing. The electrical connection path is the series-parallel electrical connection path of each battery module within the battery cluster: in this embodiment, the twelve battery modules are connected in series within the battery cluster, and adjacent battery modules are electrically connected through connecting bars and contact points. An abnormal increase in connection impedance is both a heat source and a precursor to a fault. By introducing the above thermal coupling path and electrical connection path, the two physical channels through which risk propagates between modules can be explicitly depicted, avoiding the problem of omitting the coupling propagation effect by treating each battery module as an isolated individual, thus providing a physical basis for the construction of the connection edges in the spatiotemporal diagram in S100.

[0042] This embodiment elaborates on the construction method of the spatiotemporal graph, specifically as follows: The specific method for constructing a spatiotemporal graph using each battery module as a node and coupling connection relationships as connection edges is as follows: Thermal coupling edges are established between battery module pairs directly connected via thermal coupling paths. The edge weight of the thermal coupling edge is determined by the thermal resistance between the battery module pairs. The thermal resistance is determined by the length of the heat transfer path between the battery module pairs and the thermal conductivity of the heat-conducting medium. Electrical connection edges are established between battery module pairs directly connected via electrical connection paths. The edge weight of the electrical connection edge is determined by the connection impedance between the battery module pairs. A single-period graph structure is formed by the thermal coupling edges and the electrical connection edges. The single-period graph structure with a continuously preset number of sampling periods is arranged in chronological order to form a spatiotemporal graph. The node characteristics of each battery module node in the spatiotemporal graph are the discharge operation data of the corresponding battery module in the corresponding sampling period.

[0043] Specifically, the spatiotemporal graph organizes the operating state of the energy storage system into graph-structured data: each battery module is a node, and the edges connecting nodes represent coupling channels between modules. The reason for using a graph structure instead of simply concatenating the features of each module into a vector is that the graph structure can explicitly encode the topological information—which two modules have heat and current transfer channels—into the network input, strictly limiting the aggregation range of graph convolution to physically connected module pairs. The construction details of thermal coupling edges, electrical connection edges, and temporal stacking are explained below.

[0044] Specifically, thermal coupling edges are established between battery module pairs directly connected via thermal coupling paths, and the edge weights are determined based on the thermal resistance between the battery module pairs. Let the length of the heat transfer path between battery module i and battery module j be... The thermal conductivity of the heat-conducting medium is The thermal conductivity cross-sectional area is The thermal resistance between battery modules is expressed as:

[0045] in, The thermal resistance between battery module i and battery module j is expressed in Kelvin per watt. This represents the length of the heat transfer path, in meters. The thermal conductivity of the heat-conducting medium is expressed in watts per meter (Kelvin). This refers to the thermal conductivity cross-sectional area, expressed in square meters. A lower thermal resistance results in stronger heat transfer and tighter coupling between the two modules.

[0046] The edge weight of a thermally coupled edge is represented as:

[0047] in, The edge weight of the thermally coupled edge is expressed in watts per Kelvin. To prevent the introduction of extremely small constants due to a zero denominator, a value of 0.001 Kelvin per watt can be used, for example. This embodiment uses the reciprocal of the thermal resistance rather than the thermal resistance itself as the edge weight because thermal resistance is inversely related to coupling strength. Taking the reciprocal results in the edge weight being in the same direction as the coupling strength. When graph convolution aggregates according to edge weights, the influence of adjacent modules monotonically increases with the tightness of coupling. If thermal resistance is directly used as the edge weight, a larger aggregation weight corresponds to weaker heat transfer, contrary to physical facts. For example, adjacent battery modules M5 and M6 transfer heat through an aluminum alloy heat-conducting plate inside the casing. The heat transfer path length is 0.05m, the thermal conductivity of the aluminum alloy is 160 Kelvin per meter, and the thermal cross-sectional area is 0.02 square meters. Substituting into the above formula, we get a thermal resistance of 0.0156 Kelvin per watt, corresponding to a thermal coupling edge weight of 60.2 Kelvin per watt.

[0048] Specifically, electrical connection edges are established between battery module pairs directly connected via electrical connection paths, and the edge weights are determined based on the connection impedance between the battery module pairs. Let the resistivity of the connection bus be... The length of the connecting row is The cross-sectional area of ​​the connecting row is Contact resistance is The connection impedance is then expressed as:

[0049] in, The connection impedance between battery module i and battery module j is expressed in ohms. To determine the resistivity of the busbar, the copper busbar is selected as... Omme; The length of the connecting row is in meters; The cross-sectional area is for connecting rows, and the unit is square meters; The contact resistance between the connector and the module terminal is expressed in ohms. The edge weight of the electrical connection edge is represented as follows:

[0050] in, Let be the edge weight of the electrically connected edge; To prevent the introduction of extremely small constants due to a zero denominator, for example, it can be set as follows: Ohms. For example, the copper busbar between adjacent battery modules M5 and M6 has a length of 0.1m and a cross-sectional area of... Square meters, contact resistance 10 microohms, substituting these values ​​gives a connection impedance of 15.7 microohms, corresponding to the weight of the electrical connection edge. The connection impedance increases as the connector becomes loose and the contact surface oxidizes, and the edge weight decreases accordingly, allowing the diagram structure to reflect the degradation of the electrical connection status.

[0051] Specifically, a single-period graph structure is constructed using all thermally coupled edges and electrically connected edges within the same sampling period. These single-period graph structures, spanning a predetermined number of consecutive sampling periods, are arranged chronologically to form a spatiotemporal graph. The node features of each battery module node in the spatiotemporal graph are the discharge operation data of the corresponding battery module within the corresponding sampling period, i.e., a four-dimensional vector composed of four components: voltage, current, temperature, and state of charge. The predetermined number of consecutive sampling periods can be set, for example, to 12 sampling periods, i.e., a 120-second time window with a 10-second sampling period. If this number is too small, the time context is insufficient to characterize the inertial process of thermal accumulation; if it is too large, the long-term information within the window becomes outdated, and the network input dimension is too high. This embodiment uses 12 sampling periods because the thermal time constant of the lithium iron phosphate module is on the order of minutes, and the 120-second window precisely covers the observable starting segment of the temperature rise trend. Through this construction method, the spatiotemporal graph simultaneously carries the coupling topology between modules and the timing of the operating state, providing both structural and data inputs for feature extraction from the spatiotemporal graph neural network, avoiding the problem of unextractable coupling features due to the loss of topological information.

[0052] This embodiment elaborates on the pre-training method and network structure of the spatiotemporal graph neural network, specifically as follows: The pre-training method of the spatiotemporal graph neural network is as follows: historical data of the discharge operation data in the energy storage system is acquired, an electro-thermal coupling simulation model of the battery module is constructed, the electro-thermal coupling simulation model is used to simulate the operation process of each battery module under normal and abnormal operating conditions, and simulation operation data is generated. The spatiotemporal graph neural network is jointly trained with the historical data of the discharge operation data and the simulation operation data. The spatiotemporal graph and discharge operation data corresponding to each historical sampling time are used as training inputs, and the voltage, current, temperature and state of charge of each battery module within a preset time after each historical sampling time are used as training labels. The spatiotemporal graph neural network includes a graph convolutional layer and a gated temporal layer. The graph convolutional layer aggregates the discharge operation data of each battery module node according to the connection edges of the spatiotemporal graph and extracts the spatial coupling features between each battery module. The gated temporal layer extracts the time dependence features of the operating state of each battery module according to the time order of the sampling period.

[0053] Specifically, the electro-thermal coupled simulation model consists of a circuit model and a lumped-parameter thermal model coupled together. The circuit model describes the electrical behavior of the battery module using an ideal voltage source, ohmic internal resistance, and a polarized RC network. The lumped-parameter thermal model describes the heat generation and dissipation behavior of the module using heat generation rate, thermal capacity, and thermal resistance. Joule heating couples the two models together. The electro-thermal coupled simulation model simulates the operation under normal and abnormal conditions. Abnormal conditions include three typical scenarios: cooling circuit blockage, loose connectors, and internal short circuits within the battery cell. Simulation operation data covering both normal and abnormal conditions is generated. This embodiment uses a combination of historical data and simulation operation data for training because there are very few abnormal condition samples in the actual operation of energy storage power stations. Training solely based on historical data would make the network insensitive to abnormal precursors. Simulation data is generated according to the physical model, which can supplement abnormal condition samples in a controllable manner, allowing the network to learn the coupling evolution law of each module state under abnormal conditions during the pre-training stage.

[0054] Specifically, the training samples are organized using a sliding time window: a preset number of consecutive sampling periods before each historical sampling moment serve as the input window, and the measured values ​​within a preset time period after that historical sampling moment serve as the label window. The input window and the label window are connected end-to-end on the time axis without overlapping. For example, the input window can be set to 12 consecutive sampling periods, with a preset time period of 60 seconds (6 sampling periods). Then, the spatiotemporal diagrams and discharge operation data of the 1st to 12th sampling periods are used as training inputs, and the measured voltage, current, temperature, and state of charge of each battery module within the 13th to 18th sampling periods are used as training labels. Subsequently, the time window slides along the time axis by one sampling period at a time, predicting the 14th to 19th sampling periods using the 2nd to 13th sampling periods, and so on, traversing all historical data. Thus, each prediction of the network during the training phase is constrained to extrapolate the future from the past. After training, the output state prediction result when deployed online is also the future state within a preset time period after the current moment, rather than a restatement of the currently measured state.

[0055] Specifically, the spatiotemporal graph neural network includes graph convolutional layers and gated temporal layers. The graph convolutional layers aggregate the discharge operation data of each battery module node according to the connection edges of the spatiotemporal graph, and extract the spatial coupling features between each battery module. In this embodiment, two graph convolutional layers are used, and the feature representation of node i in the (l+1)th layer is as follows:

[0056] in, Let i be the feature vector of node i in the (l+1)th layer; For non-linear activation functions, this embodiment uses the ReLU function; Let i be the set of adjacent nodes of node i in the spatiotemporal graph; Let be the edge weight of the connection between node i and its neighboring node j. For hot-coupled edges, take... Electrical connection side ; and These are the adjacency feature transformation matrix and the self-feature transformation matrix of the l-th layer, respectively; Let be the feature vector of the neighboring node j in the l-th layer; This is the bias vector for layer l. The gated temporal layer uses gated recurrent units to extract time-dependent features from the spatial coupling feature sequence of each node according to the time order of the sampling period. These features are then mapped by the output layer to voltage, current, temperature, and state of charge for each sampling period within a preset time. This embodiment adopts a serial structure of spatial aggregation followed by temporal extraction because the coupling propagation between modules affects the state of adjacent nodes before the temporal evolution of a single module. First, graph convolution gathers the adjacent influences to each node, and then the gated temporal layer characterizes the temporal dependence of each node itself. The extraction order of the two features is consistent with the physical propagation order of the risk.

[0057] Specifically, the state prediction results include the voltage, current, temperature, and state of charge of each battery module within a preset time period, i.e., a four-component prediction sequence for each battery module within a preset time period, sampled period by sample. During online operation, the spatiotemporal graph neural network outputs the state prediction results once every prediction period, with the prediction period consistent with the preset time period, for example, 60 seconds. This ensures that a complete prediction covering the next 60-second window is obtained at the beginning of each prediction period, providing rhythmically updated input for determining the risk index of S200 and judging the threshold of S500 within each prediction period. Through the above pre-training method and network structure, the spatiotemporal graph neural network possesses both spatial coupling perception and temporal extrapolation capabilities, avoiding the problem of failing to characterize coupling propagation by modeling only the time sequence of a single module, thus providing a reliable prediction basis for the quantification of the three risk indices.

[0058] To verify the convergence of the pre-training process of the spatiotemporal graph neural network, see [reference needed]. Figure 2 , Figure 2 This is a schematic diagram of the pre-training loss variation of a spatiotemporal graph neural network provided in this embodiment. The horizontal axis represents the training rounds, and the vertical axis represents the loss value. The training samples are composed of historical data of discharge operation data and simulation operation data generated by the electro-thermal coupling simulation model. The loss value is the mean square error of the spatiotemporal graph neural network on the training samples.

[0059] Depend on Figure 2 It can be seen that during the 0 to 20 rounds, the training loss rapidly decreased from 0.120 to 0.031, and the validation loss decreased from 0.128 to 0.044. The spatial coupling features extracted by the graph convolutional layer and the temporal dependency features extracted by the gated temporal layer enable the spatiotemporal graph neural network to quickly fit the operating rules under normal and abnormal conditions. During the 20 to 60 rounds, the training loss decreased from 0.031 to 0.005, and the validation loss decreased from 0.044 to 0.011, with the decrease slowing down. During the 60 to 100 rounds, the training loss stabilized at around 0.005, and the validation loss gradually decreased from 0.011 to 0.009. After convergence, the difference between the two was about 0.004 and no longer increased, and there was no overfitting phenomenon of validation loss rebound. The above convergence process shows that the method of jointly training with historical data of discharge operation data and simulation operation data can enable the spatiotemporal graph neural network to obtain the generalization prediction ability of normal and abnormal operating conditions under the condition of limited measured abnormal samples. It provides a reliable model basis for the online rolling output of the state prediction results of each battery module within a preset time period, and verifies the trainability and prediction reliability of the pre-trained spatiotemporal graph neural network in S100.

[0060] This embodiment elaborates on the determination methods of the three risk indices and the weighted summation of the comprehensive risk index, specifically as follows: The process involves determining the electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index for each battery module based on the state prediction results, and then weighting and summing them to determine the comprehensive risk index. Specifically, the electrochemical abuse risk index is determined based on the deviations of the voltage and preset voltage safety boundary of each battery module in the state prediction results, and the deviations of the current and preset current safety boundary of each battery module in the state prediction results. The thermal runaway risk index is determined based on the temperature of each battery module in the state prediction results, and the temperature difference between battery modules connected via thermal coupling paths. The capacity decay risk index is determined based on the rate of change of state of charge and depth of charge / discharge of each battery module in the state prediction results. The electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index are normalized and then weighted and summed to obtain the comprehensive risk index, with the sum of the weighting coefficients for the weighted summation being 1.

[0061] Specifically, the three risk indices characterize three different failure paths. All indices use the predicted sequence from the state prediction results, with each sample period serving as input. The maximum value of each deviation in the predicted sequence is used in the calculation, ensuring the index reflects the most unfavorable moment within the preset time period. Since the three risk indices have different physical dimensions, they must be unified to the same dimensionless interval and have the same direction (higher values ​​indicate higher risk) using their respective normalization benchmarks before being weighted. The specific normalization benchmarks are given below.

[0062] Specifically, the electrochemical abuse risk index is synthesized from voltage deviation and current deviation. The preset voltage safety boundary includes a lower limit and an upper limit, and the preset current safety boundary is the upper limit of the allowable charge / discharge current amplitude; these safety boundaries serve as the normalization benchmark. Let the predicted voltage of battery module i be... The upper limit of the preset voltage safety boundary is The lower limit is The overvoltage deviation and undervoltage deviation are expressed as follows:

[0063]

[0064] in, This is an overpressure deviation, dimensionless. This is a dimensionless measurement of undervoltage deviation. The predicted voltage of battery module i in the state prediction results is expressed in volts. and The upper and lower limits of the preset voltage safety boundary are defined in volts. The min operation truncates the deviation to within 1 to ensure the exponent is bounded. Let the predicted current of battery module i be... The preset current safety boundary is The current deviation and electrochemical abuse risk index are expressed as:

[0065]

[0066] in, This refers to the current deviation, which is dimensionless. The predicted current of battery module i in the state prediction results is positive for discharging and negative for charging, and the unit is amperes. This is a preset current safety boundary, measured in amperes. This is the electrochemical abuse risk index for battery module i, dimensionless, and ranging from 0 to 1. For example, it can be set to... =73V =56V =140A. This set of values ​​is calculated based on the safe operating range of 2.8V to 3.65V for a single lithium iron phosphate module cell, with 20 cells in series and a margin for adjustment. The current safety boundary is adjusted at a rate of 0.5C.

[0067] Specifically, the thermal runaway risk index is synthesized from temperature deviation and inter-module temperature difference deviation, using a preset temperature safety boundary and the critical thermal runaway temperature as normalization benchmarks. Let the predicted temperature of battery module i be... The preset temperature safety boundary is The critical temperature for thermal runaway is The maximum predicted temperature difference between battery modules connected via thermal coupling paths is The preset allowable temperature difference is The temperature deviation, temperature difference deviation, and thermal runaway risk index are respectively expressed as:

[0068]

[0069]

[0070] in, Temperature deviation, dimensionless; This is a dimensionless temperature difference deviation. The predicted temperature of battery module i in the state prediction results, in degrees Celsius; To preset a safe temperature boundary, for example, it can be set to 55 degrees Celsius; The critical temperature for thermal runaway can be set, for example, to 130 degrees Celsius. The maximum predicted temperature difference between battery module i and each battery module directly connected via a thermal coupling path, in degrees Celsius. The preset allowable temperature difference can be set, for example, to 10 degrees Celsius; The thermal runaway risk index for battery module i is dimensionless and ranges from 0 to 1. Temperature deviation is normalized to the range from the safety boundary to the critical temperature, meaning that the temperature deviation is 1 when the predicted temperature reaches the thermal runaway critical temperature, with a clear physical meaning. The inter-module temperature difference term is introduced because even if the temperature of a single module does not exceed the limit, abnormal temperature differences between modules characterize heat transfer imbalances along the thermal coupling path, and are precursors to the spread of thermal runaway within the cluster.

[0071] Specifically, the capacity degradation risk index is synthesized from the deviation of the rate of change of state of charge (SOC) and the deviation of the depth of charge / discharge, using the upper limit of allowable degradation as a normalization benchmark. The upper limit of allowable degradation includes both the upper limit of the allowable rate of change of SOC and the upper limit of the allowable depth of charge / discharge. Let the predicted rate of change of SOC for battery module i be... The upper limit of the allowable rate of change of state of charge is The predicted depth of charge and discharge is The maximum allowable depth of charge / discharge is The two deviation components and the capacity decay risk index are expressed as follows:

[0072]

[0073]

[0074]

[0075] in, The rate of change of the state of charge of battery module i is obtained from the difference in the predicted state of charge between adjacent sampling times in the state prediction results; The predicted state of charge at sampling time t; The interval between adjacent sampling times is consistent with the sampling period; To allow an upper limit on the rate of change of state of charge, for example, it can be set to 0.5C, or 50% per hour; To predict the depth of charge and discharge, it is determined by the difference between the maximum and minimum values ​​of the predicted state of charge within a preset time period; To allow an upper limit on the depth of charge and discharge, for example, it can be set to 90%; and These are the velocity deviation and depth deviation, respectively, and are dimensionless. , is the capacity degradation risk index for battery module i, dimensionless, and ranges from 0 to 1. The faster the rate of change of state of charge and the deeper the charge / discharge depth, the more intense the cell side reactions, and the faster the capacity degradation. Therefore, the greater the deviation between the two, the higher the capacity degradation risk index.

[0076] Specifically, the comprehensive risk index is expressed as:

[0077] in, The comprehensive risk index of battery module i is dimensionless and ranges from 0 to 1. , , These are the weighting coefficients for the electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index, respectively. The weighting coefficients are constructed as follows: Since the consequences of thermal runaway are the most severe once it occurs, the weighting coefficient for the thermal runaway risk index is preset to 0.4; the sum of the weighting coefficients for the other two items is the difference between 1 and 0.4, which is 0.6, and this sum is divided equally among the remaining two indices, each at 0.3. =0.3、 =0.4、 =0.3, the sum of the three weighting coefficients is always equal to 1, so the value range of the comprehensive risk index remains within the range of 0 to 1.

[0078] This embodiment uses the maximum value of each deviation component instead of the mean value to synthesize each risk index because the safety risk is determined by the weakest deviation component. Mean value synthesis will dilute the risk of a single serious over-limit: if mean value synthesis is used, the electrochemical abuse risk index will be lowered when the voltage is normal but the current is seriously over-limit, resulting in an underestimation of the risk. For example, the predicted results for a certain battery module are: predicted voltage 73.5V, predicted current 148A, predicted temperature 63 degrees Celsius, predicted temperature difference with the thermally coupled adjacent module 16 degrees Celsius, predicted rate of change of state of charge 0.6C, and predicted depth of charge / discharge 92%. Substituting these into the above formula, we get the overvoltage deviation 0.007, current deviation 0.057, electrochemical abuse risk index 0.057, temperature deviation 0.107, temperature difference deviation 0.6, thermal runaway risk index 0.6, rate deviation 0.2, depth of charge / discharge deviation 0.022, and capacity decay risk index 0.2. The comprehensive risk index is the sum of 0.3×0.057, 0.4×0.6, and 0.3×0.2, which is 0.317. By synthesizing based on the maximum deviation and normalizing based on the safety boundary, the three risk indices are dimensionless, directional, and bounded, providing a directly comparable input for the S300 to determine the upper limit of charge / discharge power according to the preset mapping rules.

[0079] This embodiment elaborates on the method of narrowing the feasible region of charge and discharge power, specifically as follows: According to a preset mapping rule, the comprehensive risk index of each battery module is mapped to the upper limit of the charge and discharge power of each battery module. The feasible domain of charge and discharge power of each battery module is then narrowed using the mapped upper limit of charge and discharge power. Specifically, the upper boundary of the feasible domain of charge and discharge power of each battery module is initially the rated charge and discharge power of the corresponding battery module. The preset mapping rule divides the value range of the comprehensive risk index into multiple consecutive risk intervals. Each risk interval corresponds to a charge and discharge power upper limit level. The union of the risk intervals covers the entire value range of the comprehensive risk index. The upper limit of charge and discharge power corresponding to the risk interval to which the comprehensive risk index of each battery module belongs is taken as the upper limit of charge and discharge power of the corresponding battery module. The upper boundary of the feasible domain of charge and discharge power is replaced with the mapped upper limit of charge and discharge power, while the lower boundary of the feasible domain of charge and discharge power remains unchanged, resulting in the narrowed feasible domain of charge and discharge power.

[0080] Specifically, the feasible domain of charging and discharging power for each battery module is defined by power amplitude: the upper boundary is initially the rated charging and discharging power of the corresponding battery module, for example, it can be set to 9kW, and the lower boundary is zero power; contraction only replaces the upper boundary with the mapped upper limit of charging and discharging power, while the lower boundary remains unchanged at zero power. The power in the charging condition is represented by a negative value, and the power amplitude in the charging direction is constrained by the same upper limit of charging and discharging power. That is, the feasible domain in the charging direction contracts mirrored with the same upper limit in the discharging direction, thereby preserving optimization space in both charging and discharging directions.

[0081] Specifically, the preset mapping rule divides the comprehensive risk index range of 0 to 1 into multiple consecutive risk intervals. Each risk interval corresponds to a charging / discharging power limit level. The union of all risk intervals covers all values ​​from 0 to 1, and adjacent intervals are contiguous and do not overlap. The mapping rule used in this embodiment is as follows: when the comprehensive risk index falls between 0 and 0.15 (excluding 0.15), the level coefficient is 1.0, and the charging / discharging power limit is the rated charging / discharging power; when it falls between 0.15 and 0.3 (excluding 0.3), the level coefficient is 0.8; when it falls between 0.3 and 0.45 (excluding 0.45), the level coefficient is 0.6; when it falls between 0.45 and 0.6 (excluding 0.6), the level coefficient is 0.3; and when it falls between 0.6 and 1, the level coefficient is 0.1. The mapped charging / discharging power limit is expressed as:

[0082] in, This represents the upper limit of the charging and discharging power obtained by mapping battery module i, in kilowatts. The rated charge and discharge power of battery module i is expressed in kilowatts. Comprehensive risk index The tier coefficient corresponding to the risk range is selected from one of the following ranges: 1.0, 0.8, 0.6, 0.3, and 0.1. This mapping rule is adjusted based on the derating curve given in the module manufacturer's safety manual and historical operating data of the energy storage power station: no derating is performed in the low-risk range below 0.15 to avoid excessive constraints that could harm economics; the higher the risk, the more the tier coefficient is compressed, thus accelerating the tightening of the power margin in the high-risk range. For example, when the comprehensive risk index of a battery module with a rated charge / discharge power of 9kW is 0.317, it falls in the range of 0.3 to 0.45, with a tier coefficient of 0.6, and the mapped upper limit of charge / discharge power is 5.4kW.

[0083] This embodiment employs a stepped-level mapping instead of a continuous linear mapping because the energy storage converter adjusts the module power with finite precision. The stepped levels match the actuator's adjustment granularity, ensuring that the upper power limit remains constant even when the overall risk index fluctuates slightly near the interval boundary. This avoids the problem of frequent fluctuations in the upper power limit due to index jitter under continuous mapping. If continuous linear mapping were used, numerical noise near the boundary would be amplified into power command jitter, affecting actuator lifespan. By narrowing the feasible region module by module according to the risk interval, the power constraint is precisely matched with the risk level of each battery module, providing a feasible region boundary that dynamically adjusts with risk for the multi-objective optimization of the S400.

[0084] This embodiment elaborates on the construction and solution of multi-objective optimization problems in detail, specifically as follows: The process involves optimizing both operational economy and operational safety. Within the feasible domain of the charge / discharge power after the reduction of each battery module, a multi-objective optimization problem is solved to obtain an operational optimization strategy. The energy storage system is then controlled according to this strategy. Specifically, the operational economy index is the charge / discharge revenue under the peak-valley electricity price difference within a preset time period, and the operational safety index is the weighted sum of the mean and maximum values ​​of the comprehensive risk index of all battery modules, with the weight coefficients of the weighted sum being 1. Within the feasible domain of the charge / discharge power after the reduction of each battery module, a multi-objective optimization problem is solved with both the operational economy and operational safety indices as optimization objectives. This yields the charge / discharge power plan for each battery module, the balancing strategy among battery modules, and the cooling control parameters, which serve as the operational optimization strategy. The charge / discharge power of each battery module is controlled according to the charge / discharge power plan. The power distribution among battery modules is adjusted according to the balancing strategy. The operating state of the cooling system is adjusted according to the cooling control parameters. The multi-objective optimization problem is solved continuously over a preset optimization cycle, and the operational optimization strategy is updated accordingly.

[0085] Specifically, the operational economic indicators are measured by the charging and discharging revenue under the peak-valley electricity price difference within a preset time period, and are expressed as follows:

[0086] in, The charging and discharging revenue within a preset time period is expressed in yuan. The number of sampling periods included in the preset duration; when the preset duration is 60s, it is set to 6. and These are the discharge power and charging power of the energy storage system in the t-th sampling period, respectively, in kilowatts. Both are obtained by summing the charging and discharging power of each battery module, and at most one of them is non-zero at any given time. and These are the peak and off-peak electricity prices for the t-th sampling period, respectively, in yuan per kilowatt-hour. For example, the peak electricity price is 1.2 yuan per kilowatt-hour and the off-peak electricity price is 0.35 yuan per kilowatt-hour. The sampling period is in seconds; divide by 3600 to convert seconds to hours. Discharging is recorded as positive revenue based on peak-hour electricity prices, while charging is recorded as negative cost based on off-peak electricity prices. The difference between peak and off-peak electricity prices is the source of revenue.

[0087] Specifically, the operational safety index is measured by the weighted sum of the mean and maximum values ​​of the comprehensive risk index of all battery modules, and is expressed as follows:

[0088] in, This is a dimensionless indicator for operational safety; the smaller the value, the safer the system. The weighting factor for the mean is preset to 0.6; The weighting coefficient for the maximum value is taken as the difference between 1 and 0.6, which is 0.4. The sum of the two weighting coefficients is always equal to 1. This represents the total number of battery modules. This is the comprehensive risk index for battery module i. The mean term characterizes the overall risk level of the battery cluster, while the maximum term characterizes the individual module with the highest risk. The weighted average of the two terms allows the index to reflect both the overall situation and the local situation.

[0089] The following numerical calculation example illustrates the necessity of introducing the maximum value term. Assume a battery cluster contains 12 battery modules, with 11 modules having a comprehensive risk index of 0.25, and battery module M7 having a comprehensive risk index of 0.90. First, the average comprehensive risk index of all battery modules is (11 × 0.25 + 0.90) / 12 = 0.304; subsequently, the operational safety index is 0.6 × 0.304 + 0.4 × 0.90 = 0.542. In contrast, if only the average is used as the operational safety index, the index value is 0.304, only slightly higher than the warning threshold of 0.3. Multi-objective optimization hardly sacrifices benefits to improve safety, and the optimization result is safe in an average sense. However, battery module M7, with a comprehensive risk index of 0.90, is already at extremely high risk; the average masking of local dangers. As can be seen, the operational safety index after introducing the maximum value term is 0.542, which is 1.78 times that of the pure mean index. The safety target significantly tightens the power allocation during optimization. The local high risks of individual modules are still identifiable in the system-level index. The optimization results will actively reduce the power plan of battery module M7 and strengthen the corresponding cooling, avoiding the problem of the mean masking local risks.

[0090] Specifically, the multi-objective optimization problem uses the charging and discharging power of each battery module in each sampling period within a preset time as the decision variable, and the upper limit of the charging and discharging power mapped from each battery module as the constraint boundary. The two optimization objectives are transformed into a single objective using a linear weighting method, as follows:

[0091] in, For the weighted single-objective function, the optimization direction is to find the minimum value; The weighting coefficient for the economic objective can be set, for example, to 0.5; This is the maximum charge / discharge benefit that can be achieved at full charge / discharge power within a preset time period, used to make the benefit dimensionless. The weighting coefficient for the safety objective is taken as the difference between 1 and 0.5, which is 0.5. Negative values ​​for the reward ensure that a larger reward results in a smaller target value, aligning with the safety objective. This embodiment uses the interior-point method to solve this constrained optimization problem, and the optimal solution provides the charging and discharging power plan for each battery module. Based on this, a balancing strategy allocates discharging tasks according to the state of charge (SOC) of each battery module, with modules with higher SOC undertaking more discharge power, thus balancing the charge levels among modules. Cooling control parameters include the pump speed and flow distribution of the cooling circuit, with battery modules having higher overall risk indices receiving larger cooling flow rates in their corresponding branches. These three outputs together constitute the operational optimization strategy.

[0092] Specifically, the optimization strategy is updated on a rolling basis according to a preset optimization cycle, which can be set to, for example, 15 minutes. At the beginning of each optimization cycle, the feasible domain of charge and discharge power for each battery module is narrowed down again using the latest comprehensive risk index. The multi-objective optimization problem is then solved again, and the optimization strategy is updated to ensure that control commands are always based on the latest risk predictions. By embedding risk contraction constraints into the feasible domain of multi-objective optimization, benefits and safety are balanced within the same solution framework. This avoids the frequent overturning of plans caused by making plans first and then passively limiting power, thereby outputting executable control commands that balance economy and safety.

[0093] This embodiment elaborates on the generation and execution of tiered early warning and response instructions, specifically as follows: When the comprehensive risk index of any battery module exceeds a preset warning threshold for the first time in a preset number of consecutive prediction cycles, a warning message and a handling instruction are generated and executed immediately. Specifically, the preset warning thresholds include a prompt threshold, a warning threshold, and an alarm threshold, with the prompt threshold being lower than the warning threshold and the warning threshold being lower than the alarm threshold. When the comprehensive risk index of any battery module exceeds the prompt threshold but not the warning threshold for the first time in a preset number of consecutive prediction cycles, a prompt-level warning message is generated. When the comprehensive risk index of any battery module exceeds the warning threshold but not the alarm threshold for the first time in a preset number of consecutive prediction cycles, a warning-level warning message is generated, along with a handling instruction to reduce the charge / discharge rate of the corresponding battery module. When the comprehensive risk index of any battery module exceeds the alarm threshold for the first time in a preset number of consecutive prediction cycles, an alarm-level warning message is generated, along with a handling instruction to stop the charging / discharging of the corresponding battery module and request manual inspection.

[0094] Specifically, the preset early warning thresholds include three levels: a prompt threshold, a warning threshold, and an alarm threshold. For example, these can be set to 0.3, 0.45, and 0.6 respectively. The prompt threshold is lower than the warning threshold, and the warning threshold is lower than the alarm threshold. The three threshold levels are adjusted according to the risk interval boundaries of the mapping rules, maintaining consistency with the interval division. The criterion for the first consecutive preset number of prediction cycles exceeding the threshold is as follows: the preset number can be set to, for example, three prediction cycles. The comprehensive risk index is updated on a rolling basis according to the prediction cycle. A single-point crossing of the threshold may be caused by instantaneous disturbances or sampling glitches. Therefore, it is required that the threshold be exceeded for three consecutive prediction cycles before triggering, with the first consecutive exceedance counted as the trigger time. This ensures that the early warning targets persistent, real risks rather than instantaneous interference, avoiding false alarms that could waste resources.

[0095] Specifically, when the comprehensive risk index of any battery module exceeds the alert threshold of 0.3 but does not exceed the warning threshold of 0.45 for the first time in three consecutive prediction cycles, an alert-level warning message is generated. The alert-level warning message includes the battery module number, the comprehensive risk index, and the duration. The corresponding action instruction is to maintain the current operation optimization strategy unchanged and strengthen the status monitoring of the battery module. When the comprehensive risk index of any battery module exceeds the warning threshold of 0.45 but does not exceed the alarm threshold of 0.6 for the first time in three consecutive prediction cycles, a warning-level warning message is generated, and an action instruction to reduce the charge / discharge rate of the corresponding battery module is generated. For example, if the charge / discharge rate is reduced from 0.5C to 0.15C, the upper limit of the charge / discharge power after the action is determined to be 2.7kW according to the power corresponding to the rate reduction. The action instruction is executed immediately within the same prediction cycle, and the upper boundary of the feasible domain of the module's charge / discharge power is simultaneously tightened to 2.7kW. When the comprehensive risk index of any battery module exceeds the alarm threshold of 0.6 for the first time for three consecutive prediction cycles, an alarm-level warning message is generated, and an action instruction to stop the charging and discharging of the corresponding battery module and request manual inspection is generated. After the action, the upper limit of the charge / discharge power is zero, and the module stops charging and discharging within the same prediction cycle of the action execution. The three branches cover three situations: exceeding the prompt threshold, exceeding the warning threshold, and exceeding the alarm threshold. The branches are mutually exclusive and exhaustive according to the threshold levels, and each branch has definite warning content and action results.

[0096] Specifically, the conversion of disposal instructions into upper limits of charge and discharge power is determined according to the disposal type: For derating disposal, the upper limit of charge and discharge power after disposal is the power corresponding to the derating rate, i.e., the upper limit of charge and discharge power after disposal equals the product of the derating rate, module capacity, and module nominal voltage. For example, with a 0.15C rate, a capacity of 280Ah, and a nominal voltage of 64V, the upper limit of charge and discharge power after disposal is 2.7kW; for shutdown disposal, the upper limit of charge and discharge power after disposal is zero. The rule for adjusting the state prediction results according to the upper limit of charge and discharge power after disposal is as follows: the portion of the predicted current amplitude exceeding the upper limit of charge and discharge power after disposal is clamped to the current corresponding to that upper limit; the predicted state of charge is recalculated by integrating the clamped current over time; the predicted temperature is recalculated by the heat generation rate corresponding to the clamped current through the thermal model part of the electro-thermal coupling simulation model; and the predicted voltage is recalculated by the voltage-current relationship of the circuit model based on the recalculated state of charge and the clamped current, so that the adjusted state prediction results are consistent with the operating conditions after disposal, providing an aligned prediction benchmark for the deviation calculation of S600. By combining tiered thresholds with immediate action, the intensity of risk control is matched to the level of risk at each level, avoiding the problems of over-handling small risks and losing benefits, and under-handling large risks and missing opportunities. This provides a clear predictive adjustment benchmark for closed-loop correction.

[0097] This embodiment elaborates on the closed-loop correction method in detail, specifically as follows: After executing the disposal command, the state prediction results of the corresponding battery module are adjusted according to the disposal command. The deviation between the adjusted state prediction results and the actual operating data of each battery module is used as the correction amount. The network parameters of the spatiotemporal graph neural network and the preset mapping rules are corrected according to the correction amount. Specifically, the actual operating data of each battery module within a preset number of prediction cycles after the disposal command is executed are obtained. The actual operating data includes the actual voltage, actual current, actual temperature and actual state of charge of each battery module. The upper limit of the charge and discharge power of the corresponding battery module after disposal is determined according to the disposal command. The state prediction results of the corresponding battery module are adjusted according to the upper limit of the charge and discharge power after disposal. The deviation between the actual operating data of each battery module and the adjusted state prediction results is calculated for each prediction cycle. The average deviation of each prediction cycle is used as the correction amount. When the correction amount exceeds the preset correction threshold, the network parameters of the spatiotemporal graph neural network are incrementally updated with the correction amount, and the mapping relationship between the comprehensive risk index and the upper limit of charge and discharge power in the preset mapping rules is shifted and adjusted.

[0098] Specifically, the actual operating data includes the actual voltage, actual current, actual temperature, and actual state of charge of each battery module within a preset number (e.g., 3) prediction cycles after the disposal command is executed. Actual operating data and S100 discharge operating data are two different terms: discharge operating data only covers the discharge condition, while actual operating data covers all measured states after disposal. After the shutdown disposal, the battery module is in a static state with no discharge current, and the actual current is zero. At this time, the deviation is calculated based on the actual temperature and actual state of charge during the static period and the adjusted state prediction results. The deviation consists only of two components: temperature and state of charge. The correction chain continues completely and uninterrupted under the shutdown branch.

[0099] Specifically, the deviations are synthesized after component normalization, with each component based on its corresponding safe range: the voltage range is set at the difference of 17V between the upper and lower limits of the preset voltage safe boundary; the current range is set at the preset current safe boundary of 140A; the temperature range is set at the difference of 75 degrees Celsius between the thermal runaway critical temperature and the preset temperature safe boundary; and the state of charge range is set at the upper limit of the allowable depth of charge and discharge of 90%. Let the absolute deviations between the adjusted state prediction result and the actual operating data in the four components of voltage, current, temperature, and state of charge in the k-th prediction cycle be as follows: , , , Then the deviation and correction amount for the k-th prediction period are expressed as follows:

[0100]

[0101] in, The deviation for the k-th prediction period is dimensionless. , , , The safe ranges for voltage, current, temperature, and state of charge are respectively set to 17V, 140A, 75 degrees Celsius, and 90%; This is a correction quantity, dimensionless. The number of prediction cycles involved in the correction is, for example, 3. Under the shutdown branch, the current component is zero, and the deviation consists of temperature and state-of-charge components; the formula naturally degenerates without requiring an additional branch.

[0102] Specifically, when the correction amount exceeds a preset correction threshold (e.g., 0.02), the network parameters of the spatiotemporal graph neural network are incrementally updated using the correction amount, and the mapping relationship between the comprehensive risk index and the upper limit of charging and discharging power in the preset mapping rules is shifted and adjusted. The incremental update of the network parameters is expressed as follows:

[0103] in, and These are the network parameters of the spatiotemporal graph neural network after and before the incremental update, respectively. The learning rate for incremental updates can be set to, for example, 0.01. The gradient is obtained by applying the actual operating data and adjusted state prediction results of the corrected samples to the loss function of the spatiotemporal graph neural network using a preset number of consecutive prediction cycles after the treatment. The translation adjustment of the mapping relationship is as follows: the boundary values ​​of each risk interval in the preset mapping rule are uniformly increased by a translation amount of -δ, that is, the boundary of each risk interval is shifted to the left by δ. The same comprehensive risk level corresponds to a lower upper limit of charging and discharging power, so that the mapping rule tightens as the prediction deviation is exposed. When the correction amount does not exceed the preset correction threshold, the network parameters of the spatiotemporal graph neural network and the preset mapping rule are not updated, and they are directly used to the next optimization cycle to avoid amplifying random fluctuations within the normal range into rule drift. This embodiment uses incremental updates with thresholds instead of full retraining after each treatment because full retraining is time-consuming and easily biased by single deviations. Incremental updates with thresholds are only triggered when the deviation is large enough, and the update magnitude is proportional to the correction amount, which takes into account both response speed and stability. The modified spatiotemporal graph neural network and the preset mapping rules are applied to the next optimization cycle, so that the state prediction and feasible region shrinkage of the next cycle are based on actual verification, thereby completing a complete closed loop from prediction, optimization, early warning, handling to correction.

[0104] To visually demonstrate the gear mapping relationship and translation adjustment effect of the preset mapping rules, please refer to... Figure 3 , Figure 3 This embodiment provides a schematic diagram of the upper limit of charge and discharge power before and after the preset mapping rule is corrected. In the figure, the horizontal axis is the comprehensive risk index, the vertical axis is the upper limit of charge and discharge power, the solid line is the mapping rule before correction, the dotted line is the mapping rule after correction obtained by uniformly shifting the boundary of each risk interval to the left by 0.022, and the round dots are the mapping points of the battery module with a rated charge and discharge power of 9kW and a comprehensive risk index of 0.317 in the aforementioned example.

[0105] Depend on Figure 3It can be seen that in the mapping rule before the revision, when the comprehensive risk index falls in the range of 0 to 0.15, the upper limit of charging and discharging power remains at the rated charging and discharging power of 9kW. When it falls in the ranges of 0.15 to 0.3, 0.3 to 0.45, 0.45 to 0.6, and 0.6 to 1, the corresponding upper limits of charging and discharging power are 7.2kW, 5.4kW, 2.7kW, and 0.9kW, respectively. The upper limit of charging and discharging power tightens step by step with the risk range, and the upper limit of charging and discharging power remains constant within the same risk range. Small fluctuations of the comprehensive risk index near the range boundary will not cause fluctuations in the upper limit of charging and discharging power. After the revision of the mapping rule, the boundaries of each risk range are uniformly shifted to the left by 0 from 0.15, 0.3, 0.45, and 0.6. From 0.22 to 0.128, 0.278, 0.428, and 0.578, the upper limit of charging and discharging power for each gear level remains unchanged. The same comprehensive risk index falls into a higher risk range after the correction. For example, a comprehensive risk index of 0.13 falls in the range of 0 to 0.15 before the correction, corresponding to an upper limit of charging and discharging power of 9kW. After the correction, it falls in the range of 0.128 to 0.278, corresponding to an upper limit of charging and discharging power of 7.2kW. The mapping rules are tightened overall. A comprehensive risk index of 0.317 falls in the range corresponding to a gear level coefficient of 0.6 before and after the correction. The upper limit of charging and discharging power obtained by mapping is 5.4kW. This shows that the translation adjustment only changes the boundary of the range and does not change the gear level structure. Figure 3 This indicates that the stepped level mapping, while adapting to the adjustment granularity of the energy storage converter, can transform the prediction deviation exposed after treatment into a more prudent power constraint through the translation adjustment with a threshold, verifying the construction method of the preset mapping rule in S300 and the technical logic of translating the preset mapping rule in S600.

[0106] To more clearly illustrate the technical solution and effects of this application, the following detailed explanation will be based on the actual operation scenario of an energy storage power station in an industrial and commercial park.

[0107] The energy storage power station is equipped with a 215kWh energy storage cabinet, which contains a battery cluster. The battery cluster consists of twelve battery modules M1 to M12 connected in series. Each battery module consists of 20 strings of lithium iron phosphate cells with a nominal voltage of 64V, a capacity of 280Ah, and a rated charge and discharge power of 9kW. The key parameters in the scenario are as follows: sampling period 10s, each battery module is sampled synchronously and timestamped; the spatiotemporal graph is composed of a single-period graph structure of 12 consecutive sampling periods; preset duration 60s, i.e., 6 sampling periods; prediction period 60s, with a consecutive preset number of 3 prediction periods; preset optimization period 15min; prompt threshold, warning threshold, and alarm threshold are 0.3, 0.45, and 0.6, respectively; preset correction threshold 0.02, incremental update learning rate 0.01; preset mapping rule level coefficients are set to 1.0, 0.8, 0.6, 0.3, and 0.1 respectively for five risk ranges: 0 to 0.15, 0.15 to 0.3, 0.3 to 0.45, 0.45 to 0.6, and 0.6 to 1; peak electricity price is 1.2 yuan per kilowatt-hour, and off-peak electricity price is 0.35 yuan per kilowatt-hour. The initial time was 10:00 AM, during the peak electricity price period. The state of charge of each battery module was between 55% and 57%, and the energy storage system operated in peak discharge mode.

[0108] First, according to S100, each battery module is sampled synchronously at a sampling period of 10 seconds. At 10 o'clock, the discharge current of each battery module is 80A to 95A, the terminal voltage is 63.8V to 64.5V, the temperature is 31 degrees Celsius to 34 degrees Celsius, and the state of charge is 55% to 57%. A spatiotemporal diagram is constructed with each battery module as a node and the coupling connection relationship as the connection edge. The thermal resistance between adjacent modules through the aluminum alloy heat-conducting plate is 0.0156 Kelvin per watt, and the corresponding thermal coupling edge weight is 60.2 watts per Kelvin. The connection impedance through the copper busbar is 15.7 microohms, and the corresponding electrical connection edge weight is... The state prediction results output by the pre-trained spatiotemporal graph neural network show that the state of each battery module is stable within the following 60 seconds.

[0109] Subsequently, according to S200 and S300, the comprehensive risk index of each battery module during normal periods is between 0.08 and 0.14, all falling within the risk range of 0 to 0.15, with a level coefficient of 1.0. The upper limit of the charging and discharging power of each battery module remains at the rated charging and discharging power of 9kW. According to S400, multi-objective optimization aims to maximize peak discharge benefits and minimize operational safety indicators, outputting a charging and discharging power plan for each battery module at a discharge power not exceeding 9kW, along with corresponding balancing strategies and cooling control parameters. The energy storage system operates according to the plan.

[0110] Then, at 10:20 AM, a partial blockage occurred in the cooling branch of battery module M7. According to S100, the spatiotemporal neural network, based on the latest spatiotemporal graph, predicts: the temperature of M7 will rise to 63 degrees Celsius within the next preset time period; the predicted temperature difference with thermally coupled adjacent modules M6 and M8 will reach 16 degrees Celsius; the predicted current is 148A; the predicted rate of change of state of charge is 0.6C; and the predicted depth of charge / discharge is 92%. According to S200, the following are calculated item by item: the overvoltage deviation of 0.007 and the current deviation of 0.057 combine to form an electrochemical abuse risk index of 0.057; the temperature deviation of 0.107 and the temperature difference deviation of 0.6 combine to form a thermal runaway risk index of 0.6; the rate deviation of 0.2 and the depth deviation of 0.022 combine to form a capacity decay risk index of 0.2. The comprehensive risk index is 0.3 × 0.057 + 0.4 × 0.6 + 0.3 × 0.2 = 0.317. According to S300, 0.317 falls within the risk range of 0.3 to 0.45, with a level coefficient of 0.6. The upper limit of M7's charging and discharging power is reduced to 5.4kW, while the remaining battery modules maintain 9kW. According to S400, the average comprehensive risk index of all battery modules is 0.136, and the maximum value is 0.317. The operational safety index is 0.6×0.136+0.4×0.317=0.208. Multi-objective optimization outputs a charging and discharging power plan of 5.2kW for M7 within the reduced feasible domain, implementing a balanced strategy that allows M7 to undertake fewer discharging tasks and increases the flow rate of the cooling branch where M7 is located.

[0111] To verify the shrinkage effect of the feasible domain of charging and discharging power of each battery module according to the preset mapping rules and the power allocation results of multi-objective optimization, the comprehensive risk index, level coefficient, upper limit of charging and discharging power and charging and discharging power plan of each battery module at 10:20 are shown in Table 1.

[0112] Table 1: Feasible Domain Contraction and Charge / Discharge Power Plans for Each Battery Module at 10:20 AM

[0113] As shown in Table 1, at 10:20 AM, the comprehensive risk index of the eleven battery modules (excluding module M7) ranged from 0.112 to 0.128, all falling within the risk range of 0 to 0.15, with a level coefficient of 1.0, and the upper limit of charging and discharging power remained at the rated charging and discharging power of 9.0 kW. The comprehensive risk index of battery module M7 was 0.317, falling within the risk range of 0.3 to 0.45, with a level coefficient of 0.6, and the upper limit of charging and discharging power contracted to 5.4 kW, forming a differentiated constraint for each module compared to the other battery modules. The average comprehensive risk index of all battery modules was 0.136, with a maximum value of 0.317. The operational safety index was the sum of 0.6 × 0.136 and 0.4 × 0.317, which is 0.208. In the multi-objective optimization, within the contracted feasible region of charge and discharge power, the planned charge and discharge power for M7 is 5.2kW, which is within the upper boundary of the contracted feasible region of 5.4kW. The planned charge and discharge power for the other battery modules is between 8.5kW and 8.9kW, all within their respective upper limits of 9.0kW. Table 1 shows that the contraction of the feasible region accurately transforms the risk level of the battery modules into power boundaries. The charge and discharge power plans output by the multi-objective optimization are strictly constrained by the contracted feasible region, verifying the technical logic of contracting the feasible region of charge and discharge power according to the comprehensive risk index for each module in S300 and solving the multi-objective optimization problem within the contracted feasible region in S400.

[0114] According to S500, at 10:20, 10:21 and 10:22, the comprehensive risk index of M7 exceeded the alert threshold of 0.3 for three consecutive prediction periods but did not exceed the warning threshold of 0.45. At 10:22, an alert-level warning message was generated, and the handling instruction was to maintain the current operation optimization strategy and strengthen the monitoring of M7. Starting at 10:33, the forecast indicated that M7's temperature would rise to 72 degrees Celsius, the predicted temperature difference would be 19 degrees Celsius, the predicted current would be 155A, the predicted rate of change of state of charge would be 0.7C, the thermal runaway risk index would rise to 0.9, the electrochemical abuse risk index would be 0.107, and the capacity decay risk index would be 0.4. The comprehensive risk index was 0.3×0.107+0.4×0.9+0.3×0.4=0.512. From 10:33 to 10:35, the forecast exceeded the warning threshold of 0.45 for three consecutive forecast cycles but did not exceed the alarm threshold of 0.6. At 10:35, a warning-level early warning message was generated, and a disposal instruction was generated to reduce the charge / discharge rate of M7 from 0.5C to 0.15C. After disposal, the upper limit of charge / discharge power was 2.7kW. The action was executed immediately within the same forecast cycle, and the state prediction result of M7 was clamped and recalculated according to the current corresponding to 2.7kW.

[0115] The blockage remains unresolved. From 10:48 AM, predictions indicate that M7 temperature will rise to 92 degrees Celsius, with a predicted temperature difference of 26 degrees Celsius, a predicted current of 165A, and a predicted rate of change of state of charge of 0.9°C. The temperature deviation is (92-55) / (130-55) = 0.493, the calculated temperature difference deviation is 1.6, truncated to 1 in min calculations, resulting in a thermal runaway risk index of 1. The current deviation is (165-140) / 140 = 0.179, resulting in an electrochemical abuse risk index of 0.179. The rate of change of state of charge is... The deviation is 0.8, and the capacity decay risk index is 0.8; the comprehensive risk index is 0.3×0.179+0.4×1+0.3×0.8=0.694. From 10:48 to 10:50, the alarm threshold of 0.6 was exceeded for three consecutive prediction cycles. At 10:50, an alarm-level warning was generated, and a handling instruction to stop M7 charging and discharging and request manual inspection was generated. After the handling, the upper limit of M7's charging and discharging power was zero. M7 stopped charging and discharging and entered a static state within the same prediction cycle of the handling execution.

[0116] To illustrate the evolution of the comprehensive risk index and the triggering process of graded early warnings throughout the entire process of a partial blockage event in the cooling branch, see [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram showing the change of the comprehensive risk index of each battery module over time, as provided in this embodiment. The horizontal axis represents time, the vertical axis represents the comprehensive risk index, the dotted solid line represents the comprehensive risk index of battery module M7, the solid line represents the average of the comprehensive risk indices of the other battery modules, and the dotted line, dashed line, and dotted line represent the prompt threshold of 0.3, the warning threshold of 0.45, and the alarm threshold of 0.6, respectively.

[0117] Depend on Figure 4It can be seen that between 10:00 and 10:19, the comprehensive risk index of M7 fluctuated slightly between 0.108 and 0.125, basically overlapping with the average of 0.106 to 0.120 for the other battery modules, indicating that all battery modules were at a low risk level. After a partial blockage occurred in the cooling branch of M7 at 10:20, the state prediction results output by the spatiotemporal neural network immediately reflected the temperature rise trend. The comprehensive risk index of M7 jumped to 0.317 at 10:20, exceeding [a certain threshold] for the first time. The warning threshold was set at 0.3, and for three consecutive prediction cycles from 10:20 to 10:22, it remained above the warning threshold but below the alert threshold of 0.45. A warning-level alert was generated at 10:22. Subsequently, the congestion continued to worsen, with the comprehensive risk index of M7 climbing from 0.398 at 10:24 and 0.424 at 10:28 to 0.441 at 10:32, never prematurely reaching the alert threshold. After a prediction update at 10:33, it rose to 0.512. From 10:33 AM to 10:35 AM, the prediction rate exceeded the warning threshold for three consecutive prediction periods but did not exceed the alarm threshold of 0.6. At 10:35 AM, a warning-level alert was generated, and a reduction in the multiplier was immediately implemented. After the action, the comprehensive risk index of M7 gradually increased from 0.571 at 10:38 AM to 0.592 at 10:46 AM, never exceeding the alarm threshold. After the prediction was updated at 10:48 AM, it rose to 0.694. From 10:48 AM to 10:50 AM, the prediction rate exceeded the warning threshold for three consecutive prediction periods. The alarm threshold was triggered, and an alarm-level warning was generated at 10:50 AM, prompting an immediate shutdown. After the shutdown, the comprehensive risk index of M7 continued to decline from 0.580 at 10:52 AM. After the blockage was cleared at 11:05 AM, it dropped to 0.130 at 11:06 AM. Throughout the process, the average value of the other battery modules only gradually increased from 0.113 at 10:18 AM to 0.144 at 10:48 AM before falling back to around 0.112, remaining below the warning threshold. The risk evolution was concentrated in M7. Figure 4 This indicates that the comprehensive risk index undergoes a complete evolution from normal, abnormal, response, to recovery with the cooling branch blockage event. The criterion that the first consecutive preset number of prediction cycles exceeds the preset warning threshold enables the three-level warning to be triggered step by step when the risk persists, without being falsely triggered due to instantaneous disturbances. This verifies the risk quantification sensitivity of the weighted summation of the three risk indices in S200 and the reliability of the graded warning criteria in S500.

[0118] Figure 4 The forecast data and risk index calculation results for each key forecast period are shown in Table 2. Table 2 shows the forecast data and risk index of battery module M7 in the key forecast period.

[0119] Table 2: Forecast data and risk index of battery module M7 during key forecast periods

[0120] Table 2 shows that at 10:20, the predicted temperature of M7 was 63.0 degrees Celsius, the predicted temperature difference was 16.0 degrees Celsius, and the predicted current was 148A. Substituting these values ​​into the respective deviation formulas yielded a current deviation of 0.057, a temperature deviation of 0.107, a temperature difference deviation of 0.6, and a rate deviation of 0.2. The electrochemical abuse risk index (0.057), thermal runaway risk index (0.600), and capacity decay risk index (0.200) were synthesized based on the maximum values ​​of each deviation component. The weighted summation yielded a comprehensive risk index of 0.317. At 10:21 and 10:22, the predicted temperature rose to 64.0 and 65.0 degrees Celsius, respectively, with comprehensive risk indices of 0.353 and 0.384, exceeding the warning threshold for three consecutive prediction cycles. From 10:33 to 10:35, the predicted temperature rose to 72.0 to 73.5 degrees Celsius, with a predicted temperature difference of 19.0 to 20.0 degrees Celsius. When the temperature difference deviation rises to 0.9 to 1.0, the thermal runaway risk index rises to 0.900 to 1.000, and the comprehensive risk index is 0.512, 0.546 and 0.574 respectively, exceeding the warning threshold for three consecutive prediction cycles. From 10:48 to 10:50, the predicted temperature reaches 92.0 to 96.0 degrees Celsius, the predicted temperature difference reaches 26.0 to 28.0 degrees Celsius, the calculated value of the temperature difference deviation reaches 1.6 or above, and is truncated to 1 by the min operation. The thermal runaway risk index remains at 1.000. The predicted current of 165 to 168A causes the electrochemical abuse risk index to rise to 0.179 to 0.200. The capacity decay risk index rises to 0.800 to 0.900 with the increase of the predicted rate of change of state of charge. The comprehensive risk index is 0.694, 0.708 and 0.730 respectively, exceeding the alarm threshold for three consecutive prediction cycles. The comprehensive risk index in each row of Table 2 can be back-calculated and verified by the deviation formula and the weighted summation formula. The response direction of the three risk indices to each deviation component is consistent with the physical cause, which verifies the risk quantification logic of combining the maximum deviation and normalized weighted summation in S200 and the triggering process of the criterion for three consecutive prediction periods in S500.

[0121] Figure 4 This corroborates the findings in Table 2, which uses the same set of values ​​from the same time period. Figure 4 The continuous curve presents the entire process of the comprehensive risk index from normal fluctuations to a surge, climb, and fall, explaining why tiered early warnings are triggered step by step. Table 2 uses forecast data from key forecast periods and individual risk indices to show how the comprehensive risk index is calculated from each deviation component. Both provide a complete picture of the entire chain mechanism from local blockage of cooling branch to risk quantification, threshold determination, and tiered handling.

[0122] Subsequently, following S600, after M7 was shut down and left to stand still, there was no discharge current. The actual operating data of each battery module within three consecutive prediction cycles after the execution of the disposal command were obtained. The deviation between the actual temperature and the actual state of charge and the adjusted state prediction results of M7 during the standing period was calculated: In the first prediction cycle, the adjusted predicted temperature was 71 degrees Celsius, while the actual measured temperature was 75 degrees Celsius, a temperature deviation of 4 degrees Celsius. After normalization with the voltage deviation of 0.1V and the state of charge deviation of 0.2%, the deviation for this cycle was 0.015. In the second prediction cycle, the predicted temperature was 66 degrees Celsius, while the actual measured temperature was 72 degrees Celsius, with a deviation of 0.022. In the third prediction cycle, the predicted temperature was 62 degrees Celsius, while the actual measured temperature was 70 degrees Celsius, with a deviation of 0.029. The average deviation of the three prediction cycles, i.e., the correction amount, was 0.022, which exceeded the preset correction threshold of 0.02, indicating that the spatiotemporal neural network systematically underestimated the temperature rise under the condition of blocked cooling branches. Accordingly, the network parameters of the spatiotemporal graph neural network are incrementally updated with a learning rate of 0.01, and the boundary values ​​of each risk interval in the preset mapping rule are uniformly shifted to the left by 0.022, that is, 0.15, 0.3, 0.45, and 0.6 are shifted to 0.128, 0.278, 0.428, and 0.578 respectively. The corrected spatiotemporal graph neural network and the preset mapping rule are applied to the next optimization cycle.

[0123] At 11:05, maintenance personnel arrived and cleared the blockage in the M7 cooling branch, and the M7 resumed charging and discharging. In the next optimization cycle, the M7's overall risk index dropped to 0.13: if the previous mapping rule were used, 0.13 would fall within the 0 to 0.15 range, with a level coefficient of 1.0, and the maximum charging and discharging power would be restored to 9kW; according to the revised mapping rule, 0.13 would fall within the 0.128 to 0.278 range, with a level coefficient of 0.8, and the maximum charging and discharging power would be 7.2kW. This more prudent constraint allows the M7, which had just experienced a thermal anomaly, to return to normal operation with a smoother power return.

[0124] To illustrate the dynamic adjustment of the upper limit of charging and discharging power throughout the event, see [link / reference]. Figure 5 , Figure 5 This is a schematic diagram showing the change of the upper limit of the charging and discharging power of battery module M7 over time, as provided in this embodiment. The horizontal axis represents time, the vertical axis represents the upper limit of the charging and discharging power, the solid line with dots represents battery module M7, and the dashed lines represent other battery modules.

[0125] Depend on Figure 5It can be seen that between 10:00 and 10:19, the upper limit of M7's charging and discharging power was consistent with the other battery modules, maintaining a rated charging and discharging power of 9.0kW; at 10:20, the comprehensive risk index of M7 rose to 0.317 and fell within the risk range of 0.3 to 0.45, and the upper limit of charging and discharging power was reduced to 5.4kW by a level coefficient of 0.6, while the upper limit of charging and discharging power of the other battery modules remained at 9.0kW; at 10:35, a warning-level handling order reduced the charging and discharging rate of M7 from 0.5C to 0.15C. After the incident, the upper limit of charging and discharging power was further tightened to 2.7kW and took effect immediately within the same prediction cycle; at 10:50, the alarm-level disposal command stopped M7 charging and discharging, the upper limit of charging and discharging power was reduced to zero, and M7 was put into a static state after stopping charging and discharging; after the blockage was cleared at 11:05, the comprehensive risk index of M7 dropped to 0.13. According to the revised mapping rule, 0.13 falls in the range of 0.128 to 0.278, the gear coefficient is 0.8, and the upper limit of charging and discharging power was restored to 7.2kW instead of the rated charging and discharging power of 9.0kW. Figure 5 This indicates that the upper limit of charging and discharging power gradually shrinks with the risk level, tightens immediately with the disposal instructions, and is cautiously restored as the risk is eliminated. The entire process corresponds to the evolution of the comprehensive risk index at each moment, verifying the closed-loop control logic composed of the feasible domain shrinkage in S300, the immediate execution of disposal instructions in S500, and the shifting and adjustment of mapping rules in S600.

[0126] The operation of the above specific scenarios demonstrates that the technical solution of this application can provide risk prediction 60 seconds in advance before the M7 temperature exceeds the limit. By shrinking the feasible domain and optimizing multiple objectives, the power is first brought under control within the safety boundary. Then, the handling is upgraded step by step according to three levels: prompt, warning, and alarm. After handling, the prediction model and mapping rules are corrected based on the measured deviation, so that the same congestion conditions are constrained earlier and more strictly in subsequent operations. The scenario covers the complete chain of normal, prompt, warning, alarm, handling, correction and recovery. The values ​​of each stage are consistent, which proves the engineering value of predictive proactive safety control in terms of response lead time and control precision compared to passive alarm after exceeding the limit.

[0127] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0128] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0129] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A novel method for optimizing the operation and providing early warning of safety risks in energy storage systems, characterized in that, include: S100: Synchronously acquire discharge operation data of each battery module in the energy storage system and the coupling connection relationship between each battery module. Construct a spatiotemporal graph with each battery module as a node and the coupling connection relationship as a connection edge. Input the spatiotemporal graph and the discharge operation data of each battery module into a pre-trained spatiotemporal graph neural network to obtain the state prediction results of each battery module within a preset time period. The state prediction results include the voltage, current, temperature and state of charge of each battery module within a preset time period; S200: Based on the state prediction results of each battery module, determine the electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index of each battery module, and then perform a weighted summation to determine the comprehensive risk index. S300: According to the preset mapping rules, the comprehensive risk index of each battery module is mapped to the upper limit of the charging and discharging power of each battery module, and the upper limit of the charging and discharging power obtained by mapping is used to narrow the feasible domain of the charging and discharging power of each battery module. S400: With operational economy and operational safety as optimization objectives, a multi-objective optimization problem is solved within the feasible domain of charge and discharge power after each battery module is shrunk, and an operational optimization strategy is obtained. The operation of the energy storage system is controlled according to the operational optimization strategy. S500: When the comprehensive risk index of any battery module exceeds the preset warning threshold for the first time in a preset number of consecutive prediction cycles, a warning message and a handling instruction are generated and the handling instruction is executed immediately. S600: After executing the disposal command, adjust the state prediction result of the corresponding battery module according to the disposal command. Use the deviation between the adjusted state prediction result and the actual operating data of each battery module as the correction amount. Correct the network parameters of the spatiotemporal graph neural network and the preset mapping rules according to the correction amount. Apply the corrected spatiotemporal graph neural network and the preset mapping rules to the next optimization cycle.

2. The method according to claim 1, characterized in that, The discharge operation data includes the voltage, current, temperature and state of charge of each battery module. The coupling connection relationship includes the thermal coupling path and the electrical connection path between battery modules. The thermal coupling path is the heat transfer path formed by each battery module through a shared cooling circuit and the heat conduction structure of the housing. The electrical connection path is the series and parallel electrical connection path of each battery module within the battery cluster.

3. The method according to claim 2, characterized in that, The specific method for constructing a spatiotemporal graph using each battery module as a node and coupling connection relationships as connection edges is as follows: Thermal coupling edges are established between battery module pairs directly connected via thermal coupling paths. The edge weight of the thermal coupling edge is determined by the thermal resistance between the battery module pairs. The thermal resistance is determined by the length of the heat transfer path between the battery module pairs and the thermal conductivity of the heat-conducting medium. Electrical connection edges are established between battery module pairs directly connected via electrical connection paths. The edge weight of the electrical connection edge is determined by the connection impedance between the battery module pairs. A single-period graph structure is formed by the thermal coupling edges and the electrical connection edges. The single-period graph structure with a continuously preset number of sampling periods is arranged in chronological order to form a spatiotemporal graph. The node characteristics of each battery module node in the spatiotemporal graph are the discharge operation data of the corresponding battery module in the corresponding sampling period.

4. The method according to claim 2, characterized in that, The pre-training method of the spatiotemporal graph neural network is as follows: acquire historical data of the discharge operation data in the energy storage system, construct an electro-thermal coupling simulation model of the battery module, simulate the operation process of each battery module under normal and abnormal conditions using the electro-thermal coupling simulation model, generate simulation operation data, and jointly train the spatiotemporal graph neural network with the historical data of the discharge operation data and the simulation operation data. The spatiotemporal graph and discharge operation data corresponding to each historical sampling time are used as training inputs, and the voltage, current, temperature and state of charge of each battery module within a preset time after each historical sampling time are used as training labels. The spatiotemporal graph neural network includes a graph convolutional layer and a gated temporal layer. The graph convolutional layer aggregates the discharge operation data of each battery module node according to the connection edges of the spatiotemporal graph and extracts the spatial coupling features between each battery module. The gated temporal layer extracts the time dependence features of the operating state of each battery module according to the time sequence of the sampling period.

5. The method according to claim 4, characterized in that, The electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index of each battery module are determined based on the state prediction results of each battery module, and a weighted sum is performed to determine the comprehensive risk index. Specifically, the electrochemical abuse risk index is determined based on the deviation of the voltage of each battery module from the preset voltage safety boundary in the state prediction results, and the deviation of the current of each battery module from the preset current safety boundary in the state prediction results. The thermal runaway risk index is determined based on the temperature of each battery module in the state prediction results and the temperature difference between battery modules connected by thermal coupling paths. Based on the rate of change of state of charge and depth of charge / discharge of each battery module in the state prediction results, the capacity decay risk index is determined. The electrochemical abuse risk index, thermal runaway risk index, and capacity decay risk index were normalized and then weighted and summed to obtain the comprehensive risk index. The sum of the weight coefficients of the weighted summation was 1.

6. The method according to claim 1, characterized in that, According to the preset mapping rules, the comprehensive risk index of each battery module is mapped to the upper limit of the charging and discharging power of each battery module. The upper limit of the charging and discharging power obtained by mapping is used to narrow the feasible domain of the charging and discharging power of each battery module. Specifically, the upper boundary of the feasible domain of the charging and discharging power of each battery module is initially the rated charging and discharging power of the corresponding battery module. The preset mapping rules divide the value range of the comprehensive risk index into multiple consecutive risk intervals. Each risk interval corresponds to a charging and discharging power upper limit level. The union of each risk interval covers the entire value range of the comprehensive risk index. The upper limit of charge and discharge power corresponding to the risk range of the comprehensive risk index of each battery module is taken as the upper limit of charge and discharge power of the corresponding battery module. The upper boundary of the feasible region of charge and discharge power is replaced by the upper limit of charge and discharge power obtained by mapping, while the lower boundary of the feasible region of charge and discharge power remains unchanged, thus obtaining the contracted feasible region of charge and discharge power.

7. The method according to claim 1, characterized in that, The optimization objectives are operational economy and operational safety. A multi-objective optimization problem is solved within the feasible domain of the charging and discharging power after the shrinkage of each battery module to obtain an operational optimization strategy. The operation of the energy storage system is controlled according to the operational optimization strategy. Specifically, the charging and discharging revenue under the peak-valley electricity price difference within a preset time period is used as the operational economy indicator, and the weighted sum of the mean and maximum values ​​of the comprehensive risk index of all battery modules is used as the operational safety indicator. The sum of the weight coefficients of the weighted sum is 1. Within the feasible domain of the charge and discharge power of each battery module after shrinkage, a multi-objective optimization problem is solved with the optimization objectives of operational economic indicators and operational safety indicators. The charge and discharge power plan of each battery module, the balancing strategy between battery modules and the cooling control parameters are obtained as operational optimization strategies. The charging and discharging power of each battery module is controlled according to the charging and discharging power plan, the power distribution between battery modules is adjusted according to the balancing strategy, the operating status of the cooling system is adjusted according to the cooling control parameters, and the multi-objective optimization problem is solved and the operation optimization strategy is updated in a rolling manner according to the preset optimization cycle.

8. The method according to claim 1, characterized in that, When the comprehensive risk index of any battery module exceeds the preset warning threshold for the first time in a preset number of consecutive prediction cycles, a warning message and a handling instruction are generated and the handling instruction is executed immediately. Specifically, the preset warning threshold includes a prompt threshold, a warning threshold and an alarm threshold. The prompt threshold is lower than the warning threshold, and the warning threshold is lower than the alarm threshold. When the comprehensive risk index of any battery module exceeds the prompt threshold for the first time in a preset number of prediction cycles but does not exceed the warning threshold, a prompt-level early warning message is generated, and a handling instruction is generated to maintain the current operation optimization strategy and strengthen the monitoring of the battery module's status. When the comprehensive risk index of any battery module exceeds the warning threshold for the first time in a preset number of prediction cycles but does not exceed the alarm threshold, a warning-level early warning message is generated, and a handling instruction to reduce the charge and discharge rate of the corresponding battery module is generated. When the comprehensive risk index of any battery module exceeds the alarm threshold for the first time in a preset number of prediction cycles, an alarm-level warning message is generated, and a handling instruction is generated to stop the charging and discharging of the corresponding battery module and request manual inspection.

9. The method according to claim 1, characterized in that, After the execution of the disposal command, the state prediction result of the corresponding battery module is adjusted according to the disposal command. The deviation between the adjusted state prediction result and the actual operating data of each battery module is used as the correction amount. The network parameters of the spatiotemporal graph neural network and the preset mapping rules are corrected according to the correction amount. Specifically, the actual operating data of each battery module within a preset number of prediction cycles after the execution of the disposal command are obtained. The actual operating data includes the actual voltage, actual current, actual temperature and actual state of charge of each battery module. The upper limit of the charge and discharge power of the corresponding battery module after disposal is determined according to the disposal instruction, and the state prediction result of the corresponding battery module is adjusted based on the upper limit of the charge and discharge power after disposal. The deviation between the actual operating data of each battery module and the adjusted state prediction result is calculated for each prediction cycle, and the average deviation of each prediction cycle is used as the correction amount. When the correction amount exceeds the preset correction threshold, the network parameters of the spatiotemporal graph neural network are incrementally updated with the correction amount, and the mapping relationship between the comprehensive risk index and the upper limit of charging and discharging power in the preset mapping rules is shifted and adjusted.

Images