A method for self-learning safety control and emergency response of an energy storage system and related equipment

Abstract

The present application belongs to the technical field of energy storage system safety control and artificial intelligence, and specifically relates to a kind of energy storage system self-learning safety control and emergency response method and related equipment, real-time acquisition energy storage system current operating state data, input strategy optimization model based on event knowledge base pre-training, output safety control strategy that adapts to working condition, and adaptive safety control is executed to energy storage system.The event knowledge base stores the whole process data of the historical safety events of energy storage system completely, providing full-scene sample support for model training.When the system detects the target safety event based on real-time operation data, the current state characteristics of the event are extracted, similarity search is carried out in the event knowledge base, the corresponding historical safety event is matched, and the targeted emergency response plan is quickly generated relying on historical disposal experience.The present application effectively improves the safety and response efficiency of energy storage system operation.

Description

Technical Field

[0001] This invention belongs to the field of energy storage system safety control and artificial intelligence technology, specifically involving a self-learning safety control and emergency response method and related equipment for energy storage systems. Background Technology

[0002] With the large-scale deployment and application of offshore wind power platform energy storage systems, intelligent safety control strategies and emergency response plans are the core keys to ensuring the long-term safe, economical, and stable operation of energy storage systems. Currently, the safety control and emergency response technologies of offshore wind power energy storage systems still have many shortcomings and are difficult to adapt to the operational needs of complex offshore conditions.

[0003] Existing energy storage safety control strategies are mostly static and fixed rules, relying on fixed thresholds for operations such as over-temperature derating and fault isolation. These strategies are unable to adapt to scenarios involving equipment aging, environmental fluctuations, and changes in operating status, leading to early false alarms and delayed warnings, resulting in insufficient control precision. Furthermore, existing emergency plans are mostly generic templates, lacking differentiated design based on fault type, severity, and operating conditions. They lack standardized and quantifiable handling criteria, heavily relying on the experience of maintenance personnel, resulting in low efficiency and a high risk of misoperation. In addition, the system does not effectively utilize historical data such as fault data and operating records, lacks a self-learning and evolution mechanism, and cannot review and optimize control strategies, making it prone to recurring similar faults. Energy storage safety control involves a multi-objective game involving safety, profitability, and equipment lifespan. Existing technologies lack a dynamic optimization framework, making them prone to overlooking certain aspects. Crucially, most existing technologies rely on land-based energy storage solutions and are ill-suited to the unique challenges of unmanned offshore platforms, remote operation and maintenance, delayed rescue efforts, and harsh environments, resulting in insufficient autonomous control capabilities and operational robustness. Therefore, there is an urgent need for a self-learning method for offshore energy storage safety control and emergency response to address the numerous shortcomings of existing technologies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a self-learning safety control and emergency response method and related equipment for energy storage systems, which addresses the shortcomings of the prior art and solves the technical problems of static and fixed safety control strategies, lack of specificity in emergency plans, and insufficient exploitation of historical data value in existing energy storage systems.

[0005] The objective of this invention is achieved through the following technical solutions: In a first aspect, the present invention provides a self-learning safety control and emergency response method for an energy storage system, characterized in that it includes: Obtain the current operating status data of the energy storage system; The current operating status data is input into a pre-trained strategy optimization model, which outputs a safety control strategy and performs adaptive safety control on the energy storage system based on the safety control strategy. The strategy optimization model is trained based on an event knowledge base, which is used to store the full-process data of historical safety events of the energy storage system. When a target safety event is detected in the energy storage system based on the current operating status data, the current status features of the target safety event are extracted, a similarity search is performed in the event knowledge base to obtain matching historical safety events, and an emergency response plan for the target safety event is generated based on the matching historical safety events.

[0006] As a further improvement of the present invention, the event knowledge base stores full-process data including event metadata, trigger condition data, system response records, operation and maintenance records, result feedback, and environmental and operating condition data. Before inputting the current running state data into the pre-trained policy optimization model, the following steps are also included: Collect full-process data corresponding to the aforementioned historical security events; A time-series database is used to store the sensor time-series data in the triggering condition data, a relational database is used to store the event metadata and the operation and maintenance records, and a graph database is used to store the causal relationships between different historical security events.

[0007] As a further improvement of the present invention, the initial training of the policy optimization model includes: Define the state space and action space of the strategy optimization model. The state space includes the energy storage device status, environmental parameters, historical characteristics and operating condition indicators. The action space includes continuous adjustment actions and discrete command actions. Construct a multi-objective reward function, which includes a weighted sum of safety rewards, economic rewards, and lifespan rewards, and deducts violation penalties; Based on the deep reinforcement learning algorithm, the entire process data in the event knowledge base is used as training samples, and the policy network is iteratively trained with the goal of maximizing the cumulative reward of the multi-objective reward function to obtain the policy optimization model.

[0008] As a further improvement of the present invention, the step of performing adaptive safety control on the energy storage system according to the safety control strategy includes a dynamic threshold adjustment step: Extract the equipment aging status parameters and current ambient temperature from the current operating status data; Based on the equipment aging status parameters and the current ambient temperature, the preset baseline thermal runaway warning threshold is weighted and corrected to obtain the dynamic thermal runaway warning threshold. When the temperature parameter in the current operating status data exceeds the dynamic thermal runaway early warning threshold, the derating action or shutdown action in the safety control strategy is executed.

[0009] As a further improvement of the present invention, the execution of the security control strategy further includes a differentiated throttling allocation step: When an abnormal battery cluster with local anomalies is detected in the energy storage system, the total derating command power output by the strategy optimization model is obtained; Calculate the temperature deviation between the actual temperature of the abnormal battery cluster and the dynamic thermal runaway early warning threshold; The total derating command power is allocated according to the temperature deviation value, so that the rated output power of the abnormal battery cluster is reduced first, while the output power of the normal battery cluster remains unchanged or slightly changed.

[0010] As a further improvement of the present invention, the step of performing adaptive safety control on the energy storage system according to the safety control strategy further includes a self-reset evaluation step: When the current operating status data indicates that the energy storage system has a resettable fault, the historical automatic reset success rate of the resettable fault is queried in the event knowledge base. When the historical auto-reset success rate is greater than or equal to the first set threshold, an auto-reset command is output, and an upper limit on the number of auto-resets is set. When the historical automatic reset success rate is lower than the second set threshold, the automatic reset action is intercepted and a manual processing prompt message is output; the first set threshold is greater than the second set threshold.

[0011] As a further improvement of the present invention, the step of extracting the current state features of the target security event, performing a similarity search in the event knowledge base to obtain matching historical events, and generating an emergency response plan for the target security event based on the matching historical events includes: The current state features are converted into a current state feature vector. The cosine similarity between the current state feature vector and the historical state feature vector in the event knowledge base is calculated, and the top K historical events with the highest similarity are retrieved. Extract the historical handling measures and result feedback corresponding to the first K historical events, and determine the expected success probability of each historical handling measure; Recommended handling steps are generated in descending order of expected success probability, and an emergency response plan is formed that includes an event summary, expected effects, alternative solutions, and safety precautions.

[0012] As a further improvement of the present invention, it also includes acquiring first event handling data of executing the adaptive safety control, and acquiring second event handling data generated by executing the emergency response plan; updating the event knowledge base based on the first event handling data and the second event handling data; and retraining the strategy optimization model based on the updated event knowledge base when a preset model update condition is met, specifically including: When the number of newly added first event handling data and second event handling data in the event knowledge base reaches a set data volume threshold, or when the time interval from the last model update reaches a set period, it is determined that the preset model update condition is met. An incremental learning algorithm is used to calculate the temporal difference error of each sample from the newly added first event handling data and second event handling data, and samples with temporal difference errors higher than a preset error threshold are selected to fine-tune the parameters of the strategy optimization model.

[0013] Secondly, the present invention provides a self-learning safety control and emergency response system for energy storage systems, comprising: The data acquisition module is used to acquire the current operating status data of the energy storage system; An adaptive safety control module is used to input the current operating status data into a pre-trained strategy optimization model, output a safety control strategy, and perform adaptive safety control on the energy storage system according to the safety control strategy; the strategy optimization model is trained based on an event knowledge base, which is used to store the full-process data of historical safety events of the energy storage system; The intelligent emergency response plan generation module is used to extract the current state features of the target safety event when a target safety event is detected in the energy storage system based on the current operating status data, perform a similarity search in the event knowledge base to obtain matching historical safety events, and generate an emergency response plan for the target safety event based on the matching historical safety events.

[0014] Thirdly, the present invention provides a computer device, comprising: a processor and a computer-readable storage medium; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the self-learning safety control and emergency response method for energy storage systems as described above.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed as described above for the energy storage system self-learning safety control and emergency response method.

[0016] The beneficial effects of this invention are as follows: This invention provides a self-learning safety control and emergency response method for energy storage systems. By acquiring the current operating status data of the energy storage system, it can accurately capture the actual operating conditions of the energy storage system in real time, providing real and effective data support for subsequent precise strategy control and fault determination, avoiding the defects of existing technologies that rely on fixed parameters and are disconnected from the actual operating status; by inputting the current operating status data into a pre-trained strategy optimization model and outputting a safety control strategy, the adaptive safety control of the energy storage system is completed. The safety control rules can be dynamically adjusted according to the real-time operating status of the energy storage system, completely solving the problem that existing static and fixed control strategies cannot adapt to scenarios such as equipment aging and environmental changes, resulting in false alarms and delays; relying on an event knowledge base that stores the entire process data of historical safety events of the energy storage system to train the strategy optimization model, it can fully explore and reuse the value of historical operation, fault and handling data, endow the control model with self-learning and evolution capabilities, and prevent the recurrence of similar faults and similar handling errors; when the target safety of the energy storage system is detected... Following an incident, current state features are extracted and similarity searches are performed in the event knowledge base to match corresponding historical safety events. Based on real historical fault handling experience, targeted emergency response plans can be generated, replacing existing generalized, one-size-fits-all plan templates. This enables differentiated and standardized handling of different faults, operating conditions, and fault severity, eliminating reliance on the personal experience of maintenance personnel and effectively improving fault handling efficiency and reducing the risk of misoperation. Through the coordinated cooperation of model adaptive control and historical event matching for emergency response, this invention can dynamically balance multiple objectives such as safe operation, power generation revenue, and equipment lifespan during the safety control of energy storage systems, avoiding the problem of sacrificing one aspect for another in the trade-off between multiple objectives. At the same time, it significantly improves the autonomy and robustness of safety control in unattended and remote operation and maintenance scenarios of offshore wind power energy storage systems, fully adapting to the harsh and special operating environment at sea, and fundamentally solving all the technical problems of insufficient intelligence, adaptability, and precision in the existing offshore energy storage safety control. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the self-learning safety control and emergency response method for an energy storage system in an embodiment of the present invention; Figure 2 This is an internal structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives and technical solutions of this invention clearer and easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Example 1 This invention provides a self-learning safety control and emergency response method for energy storage systems. It is applicable to unmanned or minimally manned scenarios such as offshore wind power platforms. Through deep reinforcement learning and a historical event knowledge base, it achieves dynamic optimization of safety control strategies and intelligent generation of emergency response plans. Figure 1 The diagram illustrates a specific implementation of a self-learning safety control and emergency response method for energy storage systems. In some embodiments of this invention, the energy storage system includes several battery clusters, a data acquisition unit, an edge controller, a site server, an alarm execution unit, and a fire suppression or linkage execution unit. The data acquisition unit is used to collect real-time data on the battery clusters' temperature, voltage, current, insulation resistance, SOC, SOH, gas concentration, ambient temperature and humidity, and operating conditions. The edge controller or site server is used to deploy a strategy optimization model and execute control actions such as derating, shutdown, ventilation, fire suppression, and self-reset based on the model output. The event knowledge base is deployed on a local server, edge device, or cloud platform, configured according to the system scale and communication conditions.

[0022] S1: Obtain the current operating status data of the energy storage system.

[0023] The current operating status data collected in this step is multi-dimensional real-time information covering the core characteristics of system operation, specifically including four categories: The first category is event metadata, including event identifier, occurrence time, equipment number, event type, duration, and personnel handling the event. Event types can include warnings, faults, fire-fighting actions, shutdowns, and recovery. This type of data is used to identify the basic attributes of the event, facilitating subsequent retrieval and statistics by time, equipment, and type. The second category is trigger condition data, including sensor time-series data within a preset time window before the event occurs. Preferably, the time window can be set to several minutes to several hours before the event occurs, and the specific duration can be configured according to the scenario. Trigger condition data includes at least the temperature, temperature rise rate, gas concentration, voltage, current, insulation resistance, SOC, and SOH at each measuring point. This type of data is used to describe the evolution process before the event occurs and is the main input source for strategy optimization model training and similarity retrieval. The third category is system response records, including the sequence of actions automatically executed by the system. The action sequence can include warning level output, derating instructions, shutdown instructions, fire-fighting spraying time, spraying points, spraying dosage, ventilation actions, and circuit breaker opening and closing status. The first category of data records the response process of the automatic control system during an event, facilitating the evaluation of the effects of different action combinations. The fourth category is operation and maintenance (O&M) handling records, including on-site inspection steps by O&M personnel, component replacement, manual reset, parameter adjustments, maintenance measures, and time consumption. This data records the manual intervention process, providing a basis for subsequent emergency plan generation and automatic reset strategy assessment. The fifth category is result feedback, including the final event outcome, recovery time, power generation loss, spare parts consumption, and subsequent recurrence. The final event outcome can include false alarms, successful suppression, thermal runaway, or equipment damage level. This data is used to establish a mapping relationship between control actions and results. The sixth category is environmental and operating condition data, including ambient temperature, humidity, salt spray concentration, and system operating conditions at the time of the event. System operating conditions include at least charge / discharge status, power level, and grid status. For offshore wind power platform energy storage systems, environmental factors such as wind speed, sea fog, and salt spray erosion levels can also be added. This data characterizes the impact of external conditions on energy storage risks.

[0024] S2: Input the current operating status data into the pre-trained strategy optimization model, output the safety control strategy, and perform adaptive safety control on the energy storage system according to the safety control strategy.

[0025] The strategy optimization model is trained based on an event knowledge base, which stores the full-process data of historical safety events of the energy storage system.

[0026] The strategy optimization model uses deep reinforcement learning algorithms as its core architecture and is pre-trained with massive amounts of historical security event data. The adaptive security control operation completely replaces the traditional coarse control mode with fixed thresholds. It can adjust control parameters in real time according to the current system operation, achieving a dynamic balance between safety margin and operational benefits.

[0027] In a further embodiment of the present invention, the event knowledge base stores full-process data including event metadata, trigger condition data, system response records, operation and maintenance records, result feedback, and environmental and operating condition data.

[0028] The event knowledge base of this invention preferably adopts a hybrid database structure. Sensor time-series data is stored in a time-series database, which facilitates playback and aggregation statistics by time window; event metadata and operation and maintenance records are stored in a relational database, which facilitates structured queries; and the relationships between events are stored in a graph database, for example, expressing causal chains such as "fault A triggers maintenance operation B, which subsequently recurs C".

[0029] In some embodiments of the present invention, the construction of the event knowledge base further includes data cleaning, field standardization, and tag generation. Data cleaning is used to remove outliers, duplicate records, and severely missing data; field standardization is used to unify field naming across different devices, manufacturers, and time periods; tag generation is used to label historical events into categories such as warnings, faults, thermal runaway risks, resettable faults, or non-resettable faults, for subsequent model training and similar event retrieval. This step aims to improve the data quality of the event knowledge base and reduce noise interference with model training.

[0030] In some embodiments of the present invention, before inputting the current operating status data into the pre-trained strategy optimization model, the method further includes: collecting full-process data corresponding to historical security events; using a time-series database to store sensor time-series data in the trigger condition data; using a relational database to store event metadata and operation and maintenance records; and using a graph database to store the causal relationships between different historical security events.

[0031] In some embodiments of the present invention, a deep reinforcement learning algorithm is employed, using an event knowledge base as the training data source, to construct an optimal mapping strategy from the state space to the action space. The core task of the strategy optimization model is to output control actions that balance safety, economy, and lifespan given the current operating state, thereby achieving self-learning safety control of the energy storage system.

[0032] The initial training of the strategy optimization model includes: defining the state space and action space of the strategy optimization model. The state space includes the energy storage device state, environmental parameters, historical features, and operating condition indicators. The action space includes continuous adjustment actions and discrete command actions. A multi-objective reward function is constructed, which includes a weighted sum of safety rewards, economic rewards, and lifetime rewards, and deducts violation penalties. Based on a deep reinforcement learning algorithm, using full-process data from an event knowledge base as training samples, the strategy network is iteratively trained with the goal of maximizing the cumulative reward of the multi-objective reward function to obtain the strategy optimization model.

[0033] Specifically, the state space encompasses key characteristics of the energy storage system's current operating state, including equipment status, environmental parameters, historical features, and operating condition identifiers. Equipment status includes at least the SOC, SOH, highest temperature, lowest temperature, average temperature, maximum temperature rise rate, and voltage range for each battery cluster. Environmental parameters include at least ambient temperature, humidity, salt spray concentration, and wind speed. Historical features include at least the number of warnings in the past hour, temperature fluctuation variance, and gas concentration integral value. Operating condition identifiers include at least the current charge / discharge state, power level, grid frequency, and dispatch commands. In some embodiments of this invention, the state vector can be concatenated in a fixed order or input into the neural network after normalization. For data with different dimensions, such as temperature, voltage, SOC, and gas concentration, normalization or standardization methods can be used to avoid the impact of numerical scale differences on training stability. Discrete operating condition identifiers can be represented using one-hot encoding or embedded encoding.

[0034] The action space encompasses the executable decision-making actions of the safety control system, including continuous and discrete actions. Continuous actions include at least derating power percentage, exhaust air volume percentage, and cooling power adjustment. Discrete actions include at least warning level output, shutdown command, fire sprinkler activation, self-reset attempt, and sprinkler point selection. The purpose of the action space is to enable the model to output truly executable control commands, rather than merely risk scores. Continuous actions are used to adjust control intensity, while discrete actions are used to determine whether to execute critical safety actions. By unifying continuous and discrete control variables into the action space, the model can formulate more granular control strategies in complex scenarios.

[0035] In some embodiments of the present invention, the action space can be further expanded into a hierarchical set of actions, such as dividing derating actions into multiple levels with different percentages, and dividing fire sprinkler actions into different locations and different dosage schemes. The advantage of this is that the model output more closely resembles the executable logic on-site, which is beneficial for the direct implementation of edge controllers.

[0036] The reward function R comprehensively considers three dimensions: safety, economy, and lifespan. It is defined as the weighted sum of rewards for each objective, and the specific formula is as follows:

[0037] in, R safety For safety rewards, successfully avoiding thermal runaway is a positive reward (+10), while thermal runaway or fire is a negative reward (-10). R economy As an economic incentive, the availability value of energy storage is taken into account, and downtime losses are deducted. R lifetime For lifespan rewards, gentle charging and discharging is a positive reward, while overcharging, over-discharging, and high-rate charging and discharging (>1C) are negative rewards. R penalty Penalties for violations, including penalties for violating safety constraints (such as exceeding temperature or voltage limits); ω 1-4 These are configurable weighting coefficients that can be adjusted according to operational strategy preferences.

[0038] In some embodiments of the present invention, the weighting coefficients can be configured according to the scenario. For high-risk operating conditions, the safety reward weight can be increased; for normal operating conditions, the economic reward and lifespan reward weights can be appropriately increased. This design gives the model strong scenario adaptability.

[0039] In some embodiments of the present invention, a deep deterministic policy gradient algorithm is used for model training. Historical data from an event knowledge base is used as training samples, and an initial policy network is obtained through offline training. During training, the model selects actions based on the current state and updates policy parameters according to the reward function, thereby gradually increasing the cumulative reward.

[0040] The purpose of this step is to transform historical handling experience into a learnable policy network. Since energy storage safety control typically involves both continuous and discrete actions, deep reinforcement learning is beneficial for learning near-optimal policies in complex state spaces. The DDPG algorithm is suitable for continuous action control, and therefore can handle continuous decision-making variables such as derating power, exhaust airflow, and cooling power effectively.

[0041] In some embodiments of the present invention, the training process may employ an experience replay mechanism, inputting historical samples into the neural network in batches; alternatively, a target network smooth update method may be used to improve training stability. For samples under different operating conditions, a priority sampling mechanism may be introduced to give higher training weights to high-risk events and high-error samples.

[0042] After training, the strategy is backtested in a simulation environment to verify its effectiveness. The performance of the new strategy and historical strategies on the same events is compared, including indicators such as warning accuracy, false alarm rate, net operational benefits, and equipment lifespan reduction. This step confirms that the strategy optimization effect is quantifiable and verifiable.

[0043] In some embodiments of the present invention, the trained policy model is deployed on an edge controller or a field server to receive sensor data in real time and dynamically adjust safety control parameters and thresholds. This step is the core link in implementing the policy optimization model into control actions.

[0044] The thermal runaway early warning threshold is dynamically adjusted based on the equipment's aging status and ambient temperature.

[0045] in, Threshold dynamic This is the dynamic thermal runaway early warning threshold; Threshold base This serves as the baseline for thermal runaway early warning thresholds; f This is a multi-factor coupled correction function, which integrates correction coefficients for factors such as equipment aging, ambient temperature, and cycle life. It is obtained by machine learning algorithms through offline training or online optimization using historical operating data; SOH represents the battery health state. T amb Ambient temperature; CycleCount This represents the number of energy storage cycles.

[0046] In some embodiments of the present invention, the correction function f can be obtained offline by a machine learning algorithm based on historical operating data, or it can be updated online during system operation. In this way, the threshold not only considers fixed empirical rules, but can also gradually adapt to the current system state as operating data accumulates.

[0047] In some embodiments of the present invention, adaptive safety control is performed on the energy storage system according to the safety control strategy, including a dynamic threshold adjustment step: extracting equipment aging status parameters and current ambient temperature from the current operating status data; performing weight correction on a preset benchmark thermal runaway warning threshold based on the equipment aging status parameters and the current ambient temperature to obtain a dynamic thermal runaway warning threshold; when the temperature parameter in the current operating status data exceeds the dynamic thermal runaway warning threshold, executing a derating action or a shutdown action in the safety control strategy.

[0048] The execution of the safety control strategy also includes a differentiated derating allocation step: when an abnormal battery cluster with local anomalies is detected in the energy storage system, the total derating command power output by the strategy optimization model is obtained; the temperature deviation between the actual temperature of the abnormal battery cluster and the dynamic thermal runaway warning threshold is calculated; the total derating command power is allocated according to the temperature deviation value, so that the rated output power of the abnormal battery cluster is reduced first, while the output power of the normal battery cluster remains unchanged or changes little. Specifically, the derating allocation formula is:

[0049] in, Pi,derate For the first i Actual derating output power of cluster batteries; P i,rated For the first i Cluster battery rated output power; Δ T i For the first i Cluster battery temperature safety deviation value; α This is the temperature deviation sensitivity coefficient; P totole,derate This is the total derated command power of the system.

[0050] Based on the prediction of thermal runaway development trends, shutdown or isolation operations are performed in advance before thermal runaway becomes inevitable. The decision logic is as follows: when the predicted thermal runaway time is less than a set threshold (e.g., 60 seconds) and the current power is higher than the safe power limit, the safe shutdown process of "first reduce load, then shut down, and finally isolate" is automatically executed.

[0051] Adaptive safety control of the energy storage system based on the safety control strategy also includes a self-reset evaluation step: When a resettable fault occurs in the energy storage system, the current operating status data indicates that the historical automatic reset success rate of the resettable fault is queried from the event knowledge base. When the historical automatic reset success rate is greater than or equal to a first set threshold, a self-reset command is output, and an upper limit for the number of automatic resets is set. When the historical automatic reset success rate is lower than a second set threshold, the automatic reset action is intercepted, and a manual handling prompt is output. The first set threshold is greater than the second set threshold. Specifically, for resettable faults (such as momentary communication anomalies or brief voltage fluctuations), the model determines whether to automatically reset and the upper limit for the number of resets based on the historical success rate. The decision rule is: if the historical automatic reset success rate of this type of fault is ≥70%, then two automatic resets are allowed; if the success rate is between 30% and 70%, one automatic reset is allowed; if the success rate is <30%, then manual handling is directly reported, and automatic reset is not performed.

[0052] S3: When a target safety event is detected in the energy storage system based on the current operating status data, the current status features of the target safety event are extracted, a similarity search is performed in the event knowledge base to obtain matching historical safety events, and an emergency response plan for the target safety event is generated based on the matching historical safety events.

[0053] In this step, the current status feature is a core feature vector extracted from the current operating status data that can characterize the event attributes, severity, and development trend. The similarity retrieval operation quickly locates similar events that have occurred by quantifying the feature matching degree between the current event and historical events. Based on the mature handling experience accumulated in historical events, a customized emergency response plan is generated to avoid the problem of insufficient targeting of traditional general plans and to provide standardized and quantifiable handling guidance for operation and maintenance personnel.

[0054] Specifically, the current state features are converted into a current state feature vector, the cosine similarity between the current state feature vector and the historical state feature vectors in the event knowledge base is calculated, and the top K historical events with the highest similarity are retrieved; the historical handling measures and result feedback corresponding to the top K historical events are extracted, and the expected success probability corresponding to each historical handling measure is determined; recommended handling steps are generated in descending order of expected success probability, and an emergency response plan containing an event summary, expected effects, alternative solutions, and safety precautions is formed.

[0055] In some embodiments of the present invention, when a new event occurs, the current state feature vector is extracted, and the K most similar historical events are retrieved from the event knowledge base. Cosine similarity is used for similarity calculation.

[0056] In the formula, This is the feature vector of the current state; This represents the feature vector of the historical state.

[0057] The purpose of similar event retrieval is to find the closest case in the state space to the current event from historical events, so that the generation of contingency plans no longer depends on fixed templates, but on real historical experience. Cosine similarity is suitable for measuring the directional consistency between high-dimensional vectors, and therefore applicable to matching comprehensive features such as temperature, gas concentration, voltage, and operating conditions.

[0058] In some embodiments of the present invention, the current state feature vector can be obtained by concatenating device status, environmental parameters, historical features, and operating condition identifiers. Historical state feature vectors are extracted from labeled events in an event knowledge base. By calculating cosine similarity, historical cases most closely related to the current event can be selected.

[0059] Based on the retrieved similar events, their handling measures and final results are statistically analyzed to predict the success probability of the current event under different handling plans. This step aims to transform the feedback from similar historical events into a basis for decision-making in the current event, thereby improving the relevance and feasibility of the contingency plan.

[0060] In some embodiments of the present invention, the success rate, recovery time, degree of loss, and recurrence rate after taking a certain action can be statistically analyzed for the previous K historical events, and then combined to form a candidate treatment plan score for the current event. The higher the score, the higher the priority for inclusion in the plan.

[0061] Optionally, this embodiment also includes an online learning and model update mechanism. The system continuously collects first event handling data generated by the execution of adaptive safety control and second event handling data generated by the execution of emergency response plans, and adds these two types of new data to the event knowledge base. When the number of newly added first and second event handling data in the event knowledge base reaches a set data volume threshold, or when the time interval since the last model update reaches a set period, it is determined that the preset model update conditions are met. Subsequently, an incremental learning algorithm is used to calculate the temporal difference error of each sample from the newly added first and second event handling data. High-value samples with temporal difference errors higher than the preset error threshold are selected first to fine-tune the parameters of the strategy optimization model, avoiding low-value duplicate samples from occupying training resources. The new strategy model is first tested in a simulation environment or on a small number of units. After comparing the performance indicators of the new and old strategies to confirm the optimization effect, it is gradually promoted to the entire field. At the same time, version management of the strategy model is implemented, and historical versions are retained. When the actual performance of the new model is not good, it supports one-click rollback to the previous stable version. This online self-evolution mechanism allows the system to continuously accumulate new fault handling experience during long-term operation, constantly optimize control strategies, and achieve the effect of the system becoming smarter with use. During long-term operation, the system's early warning accuracy can continuously improve with the accumulation of events, adapting to the aging characteristics of equipment at different stages.

[0062] Specifically, as an embodiment applicable to offshore wind power platforms, the first step is to establish an event knowledge base. During operation, the system automatically collects and records the entire process data of six types of historical safety events: event metadata, trigger condition data, system response records, operation and maintenance records, result feedback, and environmental and operating condition data. Event metadata includes event ID, occurrence time, equipment number, event type, duration, and personnel involved. Trigger condition data includes sensor time-series data N minutes before the event, such as temperature, temperature rise rate, gas concentration, voltage, current, insulation resistance, state of charge, and health status at each measuring point. System response records include the sequence of actions automatically executed by the system, such as warning level, derating command, shutdown command, fire sprinkler parameters, ventilation actions, and circuit breaker opening and closing status. Operation and maintenance records include on-site handling operations by operation and maintenance personnel, such as inspection steps, component replacement, manual reset, parameter adjustment, maintenance measures, and time consumption. Result feedback includes the final result of the event, recovery time, power generation loss, spare parts consumption, and subsequent recurrence. Environmental and operating condition data includes the ambient temperature, humidity, salt spray concentration, and system operating conditions at the time of the event. The storage phase employs a time-series database to store sensor time-series data from the trigger condition data, a relational database to store event metadata and operation and maintenance records, and a graph database to store the causal relationships between different historical safety events. It supports rapid retrieval by multiple dimensions such as time, device, and event type, and allows exporting datasets for model training. This hybrid storage approach fully leverages the advantages of different database types: the time-series database adapts to the high-frequency, continuous read / write requirements of sensor time-series data, the relational database ensures the consistency of structured event information queries, and the graph database supports causal reasoning and root cause analysis between events.

[0063] Furthermore, the initial training of the strategy optimization model was completed. First, the state space and action space of the strategy optimization model were defined. The state space fully covers four core categories of features: energy storage device status, environmental parameters, historical characteristics, and operating condition indicators. The action space includes continuous adjustment actions and discrete command actions. Continuous adjustment actions include derating power percentage, exhaust air volume percentage, and cooling power adjustment. Discrete command actions include warning level output, shutdown command, fire sprinkler system, sprinkler point selection, and self-reset attempt. Subsequently, a multi-objective reward function was constructed. This function integrates the objectives of safety, economy, and lifespan, and is defined as a weighted sum of the rewards for each objective.

[0064] The system includes a safety reward (Rsafety) of 10 for successfully preventing thermal runaway and a negative reward of 10 for thermal runaway or fire. The economic reward (Reconomy) considers the availability of energy storage while deducting downtime losses. The lifespan reward (Rlifetime) is positive for moderate charging and discharging and negative for overcharging, over-discharging, or charging / discharging at rates greater than 1C. The penalty (Rpenalty) is for violating safety constraints such as exceeding temperature or voltage limits. Weighting coefficients ω1 to ω4 are configurable parameters that can be flexibly adjusted according to operational strategy preferences. A deep deterministic policy gradient algorithm is then used for model training, employing historical data from an event knowledge base as training samples. An initial policy network is obtained through offline training, during which the model learns to select the optimal action under different states to maximize accumulated rewards. After training, the strategy is backtested in a simulation environment, comparing the performance of the new strategy with historical strategies on the same events. Verification metrics include warning accuracy, false alarm rate, net operational benefits, and equipment lifespan reduction, ensuring that the strategy optimization effect is quantifiable and verifiable. This multi-objective optimization training mode can solve the shortcomings of traditional control strategies that are incomplete in some aspects. Through the flexible configuration of weight coefficients, it can adapt to the preferences and needs of different operation and maintenance scenarios. For example, for newly built energy storage systems, the safety reward weight can be increased to further improve the safety margin. For old energy storage systems that have been in service for more than 5 years, the economic reward weight can be increased to improve the power generation revenue while ensuring basic safety.

[0065] Example 2 This embodiment provides a self-learning safety control and emergency response system for an energy storage system. The system mainly includes: The data acquisition module is used to acquire the current operating status data of the energy storage system; An adaptive safety control module is used to input the current operating status data into a pre-trained strategy optimization model, output a safety control strategy, and perform adaptive safety control on the energy storage system according to the safety control strategy; the strategy optimization model is trained based on an event knowledge base, which is used to store the full-process data of historical safety events of the energy storage system; The intelligent emergency response plan generation module is used to extract the current state features of the target safety event when a target safety event is detected in the energy storage system based on the current operating status data, perform a similarity search in the event knowledge base to obtain matching historical safety events, and generate an emergency response plan for the target safety event based on the matching historical safety events.

[0066] Specific limitations regarding the self-learning safety control and emergency response system for energy storage systems can be found in the limitations of the self-learning safety control and emergency response method for energy storage systems described above. The corresponding technical effects are equivalent and will not be repeated here. Each module in the aforementioned self-learning safety control and emergency response system for energy storage systems can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0067] Example 3 Figure 2 An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 2 As shown, the computer device includes a processor, memory, network interface, display, camera, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a self-learning safety control and emergency response method for the energy storage system. The display screen can be an LCD screen or an e-ink display screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0068] As will be understood by those skilled in the art, computer equipment Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computing device may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0069] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0070] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0071] In summary, the self-learning safety control and emergency response method, system, computer equipment, and storage medium provided in this application embodiment construct a high-quality event knowledge base by recording the entire process data of historical safety events. It trains a strategy optimization model adapted to multi-objective optimization requirements using deep reinforcement learning algorithms, dynamically adjusts safety control parameters based on model output to achieve all-dimensional adaptive safety control, generates customized emergency response plans by combining similar event retrieval, and implements an incremental online learning mechanism to achieve continuous self-evolution of the system. This fundamentally solves the shortcomings of existing energy storage safety control strategies being statically fixed, emergency plans lacking specificity, and the value of historical data not being fully exploited. It is particularly suitable for special application scenarios such as unmanned operation, remote maintenance, and harsh environments of offshore wind power platforms, comprehensively improving the safety, economy, and intelligence level of long-term operation of energy storage systems.

[0072] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0073] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A self-learning safety control and emergency response method for an energy storage system, characterized in that, include: Obtain the current operating status data of the energy storage system; The current operating status data is input into a pre-trained strategy optimization model, which outputs a safety control strategy and performs adaptive safety control on the energy storage system based on the safety control strategy. The strategy optimization model is trained based on an event knowledge base, which is used to store the full-process data of historical safety events of the energy storage system. When a target safety event is detected in the energy storage system based on the current operating status data, the current status features of the target safety event are extracted, a similarity search is performed in the event knowledge base to obtain matching historical safety events, and an emergency response plan for the target safety event is generated based on the matching historical safety events.

2. The self-learning safety control and emergency response method for energy storage systems according to claim 1, characterized in that, The event knowledge base stores full-process data including event metadata, trigger condition data, system response records, operation and maintenance records, result feedback, and environmental and operating condition data. Before inputting the current running state data into the pre-trained policy optimization model, the following steps are also included: Collect full-process data corresponding to the aforementioned historical security events; A time-series database is used to store the sensor time-series data in the triggering condition data, a relational database is used to store the event metadata and the operation and maintenance records, and a graph database is used to store the causal relationships between different historical security events.

3. The self-learning safety control and emergency response method for energy storage systems according to claim 1, characterized in that, The initial training of the policy optimization model includes: Define the state space and action space of the strategy optimization model. The state space includes the energy storage device status, environmental parameters, historical characteristics and operating condition indicators. The action space includes continuous adjustment actions and discrete command actions. Construct a multi-objective reward function, which includes a weighted sum of safety rewards, economic rewards, and lifespan rewards, and deducts violation penalties; Based on the deep reinforcement learning algorithm, the entire process data in the event knowledge base is used as training samples, and the policy network is iteratively trained with the goal of maximizing the cumulative reward of the multi-objective reward function to obtain the policy optimization model.

4. The self-learning safety control and emergency response method for energy storage systems according to claim 1, characterized in that, The adaptive safety control of the energy storage system based on the safety control strategy includes a dynamic threshold adjustment step: Extract the equipment aging status parameters and current ambient temperature from the current operating status data; Based on the equipment aging status parameters and the current ambient temperature, the preset baseline thermal runaway warning threshold is weighted and corrected to obtain the dynamic thermal runaway warning threshold. When the temperature parameter in the current operating status data exceeds the dynamic thermal runaway early warning threshold, the derating action or shutdown action in the safety control strategy is executed.

5. The self-learning safety control and emergency response method for energy storage systems according to claim 4, characterized in that, The execution of the security control strategy also includes a differentiated reduction allocation step: When an abnormal battery cluster with local anomalies is detected in the energy storage system, the total derating command power output by the strategy optimization model is obtained; Calculate the temperature deviation between the actual temperature of the abnormal battery cluster and the dynamic thermal runaway early warning threshold; The total derating command power is allocated according to the temperature deviation value, so that the rated output power of the abnormal battery cluster is reduced first, while the output power of the normal battery cluster remains unchanged or slightly changed.

6. The self-learning safety control and emergency response method for energy storage systems according to claim 4, characterized in that, The step of performing adaptive safety control on the energy storage system according to the safety control strategy also includes a self-reset evaluation step: When the current operating status data indicates that the energy storage system has a resettable fault, the historical automatic reset success rate of the resettable fault is queried in the event knowledge base. When the historical auto-reset success rate is greater than or equal to the first set threshold, an auto-reset command is output, and an upper limit on the number of auto-resets is set. When the historical automatic reset success rate is lower than the second set threshold, the automatic reset action is intercepted and a manual processing prompt message is output; the first set threshold is greater than the second set threshold.

7. The self-learning safety control and emergency response method for energy storage systems according to claim 1, characterized in that, The process of extracting the current state features of the target security event, performing a similarity search in the event knowledge base to obtain matching historical events, and generating an emergency response plan for the target security event based on the matching historical events includes: The current state features are converted into a current state feature vector. The cosine similarity between the current state feature vector and the historical state feature vector in the event knowledge base is calculated, and the top K historical events with the highest similarity are retrieved. Extract the historical handling measures and result feedback corresponding to the first K historical events, and determine the expected success probability of each historical handling measure; Recommended handling steps are generated in descending order of expected success probability, and an emergency response plan is formed that includes an event summary, expected effects, alternative solutions, and safety precautions.

8. The self-learning safety control and emergency response method for energy storage systems according to any one of claims 1 to 7, characterized in that, It also includes acquiring first event handling data from the execution of the adaptive safety control, and acquiring second event handling data generated from the execution of the emergency response plan; updating the event knowledge base based on the first and second event handling data; and retraining the strategy optimization model based on the updated event knowledge base when preset model update conditions are met. Specifically, this includes: When the number of newly added first event handling data and second event handling data in the event knowledge base reaches a set data volume threshold, or when the time interval from the last model update reaches a set period, it is determined that the preset model update condition is met. An incremental learning algorithm is used to calculate the temporal difference error of each sample from the newly added first event handling data and second event handling data, and samples with temporal difference errors higher than a preset error threshold are selected to fine-tune the parameters of the strategy optimization model.

9. A self-learning safety control and emergency response system for an energy storage system, characterized in that, include: The data acquisition module is used to acquire the current operating status data of the energy storage system; An adaptive safety control module is used to input the current operating status data into a pre-trained strategy optimization model, output a safety control strategy, and perform adaptive safety control on the energy storage system according to the safety control strategy. The strategy optimization model is trained based on an event knowledge base, which is used to store the full-process data of historical safety events of the energy storage system. The intelligent emergency response plan generation module is used to extract the current state features of the target safety event when a target safety event is detected in the energy storage system based on the current operating status data, perform a similarity search in the event knowledge base to obtain matching historical safety events, and generate an emergency response plan for the target safety event based on the matching historical safety events.

10. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the self-learning safety control and emergency response method for an energy storage system as described in any one of claims 1 to 8.

Images