Intelligent fire extinguishing method and system for electrochemical energy storage system based on deep reinforcement learning
Patent Information
- Application Number
- CN202610976790.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-02
- Publication Date
- 2026-09-25
AI Technical Summary
然而，现有固定策略无法根据火灾的实际发展态势（火焰强度、温度场变化等）动态调整灭火剂喷射流量、喷施时长和间歇周期等，导致灭火时机或强度失配，特别是储能电池火灾易发生复燃，固定灭火策略往往会在初起火灾时将灭火剂喷放完毕，造成发生复燃后无法继续提供保护
[0044]基于控制周期实时获取火灾状态数据，基于该实时的火灾状态数据，通过灭火决策智能体对一个或多个着火区域进行灭火动作推理，得到各着火区域的灭火优先级和资源分配比例，并基于各着火区域的灭火优先级和资源分配比例，生成各着火区域的最优灭火动作序列，能够根据火灾的实际发展动态前瞻性调整灭火策略，实现动态、自适应灭火，提升灭火效能；同时，本方案能够优化多个着火区域的灭火优先级和资源分配比例，实现全局灭火资源优化与局部精细执行的协同，在同时存在多着火点的情况下，能够智能分配灭火资源，避免资源错配；本方案不断积累新的训练数据，对灭火决策模型和灭火决策智能体进行持续优化，形成“数据积累—模型迭代—策略提升”自我学习闭环机制，使灭火决策智能体能够根据不断积累的灭火数据不断优化，克服了传统固定策略无法自我改进的缺陷。
Smart Images
Figure CN122806019A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of electrochemical energy storage fire safety technology, and in particular to an intelligent fire extinguishing method and system for electrochemical energy storage systems based on deep reinforcement learning. Background Technology
[0002] With the large-scale application and development of electrochemical energy storage, fires caused by battery thermal runaway are frequent, posing a severe challenge to the fire safety of energy storage power stations. Traditional fire extinguishing systems for energy storage mostly adopt a fixed fire extinguishing strategy based on sensor alarm triggering and spraying extinguishing agents according to a pre-set discharge logic. However, this approach has significant shortcomings in the following aspects:
[0003] First, the fixed response strategy makes it difficult to adapt to the dynamic changes in the fire process. The response logic of traditional fire suppression systems is simple: upon receiving a fire alarm signal, the fire suppression procedure is initiated. However, existing fixed strategies cannot dynamically adjust the extinguishing agent spray flow rate, spraying duration, and intermittent period according to the actual development of the fire (flame intensity, temperature field changes, etc.), leading to a mismatch between the timing or intensity of fire suppression. In particular, since energy storage battery fires are prone to reignition, fixed fire suppression strategies often exhaust the extinguishing agent at the initial stage of the fire, making it impossible to continue providing protection after reignition.
[0004] Second, there is a lack of multi-objective optimization and coordination mechanisms. When multiple fire points occur simultaneously within an energy storage system, existing strategies cannot achieve global optimization of firefighting resources, resulting in low firefighting efficiency or firefighting failure.
[0005] Third, there is a lack of continuous optimization capabilities for fire suppression strategies. Once existing fire suppression strategies are solidified, they cannot be improved based on feedback from actual fire suppression tests, making continuous optimization difficult. Summary of the Invention
[0006] This application provides an intelligent fire extinguishing method and system for electrochemical energy storage systems based on deep reinforcement learning, aiming to at least partially solve one of the technical problems in related technologies. The technical solution of this application is as follows:
[0007] In a first aspect, embodiments of this application propose an intelligent fire extinguishing method for electrochemical energy storage systems based on deep reinforcement learning, comprising:
[0008] Based on the control cycle, the global fire status data of the energy storage system is obtained, and the global fire status data includes fire status data of at least one fire zone.
[0009] Based on the global fire status data, the fire extinguishing decision-making agent performs fire extinguishing action reasoning on the at least one fire area to obtain the fire extinguishing priority and resource allocation ratio of each fire area. Based on the fire extinguishing priority and resource allocation ratio of each fire area, the optimal fire extinguishing action sequence for each fire area in the at least one fire area is obtained. The optimal fire extinguishing action sequence includes the fire extinguishing actions that the fire extinguishing system needs to execute at each time step in multiple consecutive time steps.
[0010] The optimal fire extinguishing action sequence for each fire zone in the at least one fire zone is mapped to a fire command, and the fire extinguishing system is controlled to extinguish the fire based on the fire command.
[0011] In some implementations, the fire-fighting decision-making agent adopts a hierarchical reinforcement learning architecture, which includes:
[0012] A high-level strategy network is used to receive global fire status data of the energy storage system, and to perform fire extinguishing action reasoning on the at least one fire area based on the global fire status data, so as to obtain the fire extinguishing priority and resource allocation ratio of each fire area in the at least one fire area.
[0013] At least one low-level execution network, which corresponds one-to-one with the at least one fire zone, and each low-level execution network is used to obtain the optimal fire extinguishing action sequence for the corresponding fire zone based on the fire extinguishing priority and resource allocation ratio of the fire zone and the local fire status data of the fire zone.
[0014] In some implementations, after controlling the fire extinguishing system to extinguish the fire based on the fire command, the method further includes:
[0015] Based on the flame and temperature detection signals in the global fire status data of the next control cycle, determine whether the open flame has been completely extinguished;
[0016] If the open flame has been completely extinguished, the reignition risk index is obtained based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system.
[0017] Based on the reignition risk index, the fire extinguishing decision-making agent makes a reignition warning decision according to the reignition handling strategy, and obtains the reignition warning decision result.
[0018] In some implementations, the step of making a reignition warning decision based on the reignition risk index and the reignition handling strategy by the fire extinguishing decision-making agent to obtain the reignition warning decision result includes:
[0019] If the reignition risk index is less than the first risk threshold, the fire extinguishing decision-making agent outputs a stop spraying decision.
[0020] When the reignition risk index is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, the fire extinguishing decision agent outputs a cooling maintenance decision, which is used to instruct the fire extinguishing system to perform low-flow continuous spraying or intermittent spraying.
[0021] If the reignition risk index is greater than the second risk threshold, global fire status data for the next control cycle is obtained to infer fire extinguishing actions.
[0022] In some implementations, obtaining the reignition risk index based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system includes:
[0023] The current battery surface temperature, temperature rise rate, and characteristic gas concentration are normalized to obtain the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration.
[0024] The reignition risk index is obtained by weighted summation of the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration.
[0025] In some implementations, the training method for the fire extinguishing decision-making agent includes:
[0026] Based on historical fire suppression experimental data of energy storage systems, an initial model is trained to obtain a fire suppression decision model. The fire suppression decision model is used to predict the fire state sequence and fire suppression effectiveness for the next K time steps based on the fire state sequence and fire suppression action sequence of the current N time steps.
[0027] Based on the aforementioned fire extinguishing decision-making model, an initial agent is trained using a deep reinforcement learning algorithm to obtain the fire extinguishing decision-making agent; wherein...
[0028] The initial agent learns through trial and error with the fire extinguishing decision model, and updates its parameters with the goal of maximizing the cumulative reward, so that the initial agent can learn to generate the optimal fire extinguishing action sequence under different fire scenarios.
[0029] In some implementations, the initial model is a long short-term memory network or a Transformer time series prediction model. The input of the initial model includes a two-dimensional tensor composed of input feature vectors for N consecutive time steps. The input feature vectors for each time step include fire status data, fire extinguishing actions, and time codes.
[0030] In some implementations, the method further includes:
[0031] Acquire relevant fire extinguishing data during the fire extinguishing process, and optimize the fire extinguishing decision model and the fire extinguishing decision agent based on the relevant fire extinguishing data to obtain the optimized fire extinguishing decision model and the fire extinguishing decision agent.
[0032] Secondly, embodiments of this application propose an intelligent fire extinguishing system based on a deep reinforcement learning electrochemical energy storage system, the system comprising:
[0033] The on-site perception and execution layer, deployed inside the energy storage compartment of the energy storage system, includes a data acquisition unit and an execution unit. The data acquisition unit includes a multimodal sensor network deployed inside the energy storage system, which is used to collect global fire status data in real time and upload it to the station-level host. The execution unit is used to receive and execute fire-fighting commands issued by the station-level host. After the fire-fighting command is executed, the execution unit is also used to return the execution status and preliminary fire-fighting effectiveness evaluation indicators to the station-level host. The global fire status data includes fire status data of at least one fire area.
[0034] The edge decision control layer, deployed locally at the energy storage power station as a station-level host, is connected to the data acquisition unit and the execution unit. It receives global fire status data uploaded by the data acquisition unit, performs real-time fusion processing on the global fire status data, and inputs the fused global fire status data into the fire extinguishing decision-making agent. The fire extinguishing decision-making agent then performs fire extinguishing action reasoning on the at least one fire area to obtain the fire extinguishing priority and resource allocation ratio for each fire area. Based on the fire extinguishing priority and resource allocation ratio for each fire area, it determines the optimal fire extinguishing action for each fire area within the at least one fire area. The optimal fire extinguishing action sequence includes the fire extinguishing actions required by the fire extinguishing system at each time step in multiple consecutive time steps. The optimal fire extinguishing action sequence for each fire area in the at least one fire area is mapped to a fire command and issued to the execution unit. The edge decision control layer is also used to obtain the effect feedback data after the fire command is executed, and upload the effect feedback data and the global fire status data to the cloud optimization training layer after packaging. The edge decision control layer is also used to generate the optimal fire extinguishing action sequence based on a backup fixed strategy in the event that the fire extinguishing decision agent fails to reason or communication is abnormal.
[0035] The cloud-based optimization training layer, deployed on a remote server or industrial cloud platform, is used to receive global fire status data and effect feedback data uploaded by the station-level host and add them to the historical fire extinguishing case library as historical fire extinguishing data. The cloud-based optimization training layer includes a model training module and a reinforcement learning strategy optimization module. The model training module and the reinforcement learning strategy optimization module are used to optimize the fire extinguishing decision model and the fire extinguishing decision agent based on the historical fire extinguishing data, and push the optimized fire extinguishing decision agent to the station-level host.
[0036] Thirdly, embodiments of this application propose an intelligent fire extinguishing system based on a deep reinforcement learning electrochemical energy storage system, comprising:
[0037] The data acquisition module is used to acquire global fire status data of the energy storage system based on the control cycle. The global fire status data includes fire status data of at least one fire zone.
[0038] The action reasoning module is used to perform fire extinguishing action reasoning on the at least one fire area based on the global fire status data through the fire extinguishing decision-making intelligent agent, to obtain the fire extinguishing priority and resource allocation ratio of each fire area, and to obtain the optimal fire extinguishing action sequence for each fire area in the at least one fire area based on the fire extinguishing priority and resource allocation ratio of each fire area. The optimal fire extinguishing action sequence includes the fire extinguishing actions that the fire extinguishing system needs to perform at each time step in multiple consecutive time steps.
[0039] The fire extinguishing execution module is used to map the optimal fire extinguishing action sequence of each fire zone in the at least one fire zone into a fire command, and control the fire extinguishing system to extinguish the fire based on the fire command.
[0040] Fourthly, embodiments of this application provide an electronic device, including: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method described in the first aspect.
[0041] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method described in the first aspect.
[0042] In a sixth aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.
[0043] This application has the following advantages and beneficial effects:
[0044] Based on real-time fire status data acquired during the control cycle, a fire-fighting decision-making agent infers fire-fighting actions for one or more fire zones, obtaining the fire-fighting priority and resource allocation ratio for each fire zone. Based on these priorities and ratios, the optimal fire-fighting action sequence for each fire zone is generated. This allows for dynamic and proactive adjustment of fire-fighting strategies according to the actual development of the fire, achieving dynamic and adaptive fire-fighting and improving fire-fighting efficiency. Simultaneously, this solution optimizes the fire-fighting priority and resource allocation ratio for multiple fire zones, achieving synergy between global fire-fighting resource optimization and local fine-grained execution. In the presence of multiple fire points, it intelligently allocates fire-fighting resources, avoiding resource misallocation. Furthermore, this solution continuously accumulates new training data, continuously optimizing the fire-fighting decision model and the fire-fighting decision-making agent, forming a self-learning closed-loop mechanism of "data accumulation—model iteration—strategy improvement." This enables the fire-fighting decision-making agent to continuously optimize based on the accumulated fire-fighting data, overcoming the shortcomings of traditional fixed strategies that cannot self-improve.
[0045] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0046] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0047] Figure 1 A schematic flowchart illustrating an intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning, provided as an embodiment of this application.
[0048] Figure 2 A block diagram of an intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage provided in this application embodiment;
[0049] Figure 3 A block diagram of an intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage provided in this application embodiment;
[0050] Figure 4 This is a block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0051] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0052] The following description, with reference to the accompanying drawings, describes an intelligent fire extinguishing method, apparatus, and device for an electrochemical energy storage system based on deep reinforcement learning, according to embodiments of this application.
[0053] Figure 1 This is a schematic flowchart illustrating an intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning, provided as an embodiment of this application.
[0054] It should be noted that the execution subject of the intelligent fire extinguishing method for electrochemical energy storage systems based on deep reinforcement learning in this application embodiment is the intelligent fire extinguishing system for electrochemical energy storage systems based on deep reinforcement learning in this application embodiment. The intelligent fire extinguishing system for electrochemical energy storage systems based on deep reinforcement learning can be configured in an electronic device so that the electronic device can perform the intelligent fire extinguishing function for electrochemical energy storage systems based on deep reinforcement learning.
[0055] like Figure 1 As shown, the intelligent fire extinguishing method for electrochemical energy storage systems based on deep reinforcement learning includes the following steps:
[0056] Step S101: Based on the control cycle, acquire the global fire status data of the energy storage system. The global fire status data includes the fire status data of at least one fire zone.
[0057] In some embodiments, fire status data includes, but is not limited to, battery and module temperatures, ambient temperature, typical gas concentrations such as CO and H2, infrared temperature field, flame detector data, and thermal radiation flux; fire status data are collected through a multimodal sensor network deployed inside the energy storage system; if there is one ignition point, fire status data for one ignition area is obtained; if there are multiple ignition points, fire status data for multiple ignition areas are obtained as global fire status data.
[0058] Step S102: Based on global fire status data, the fire extinguishing decision-making agent performs fire extinguishing action reasoning on at least one fire area to obtain the fire extinguishing priority and resource allocation ratio of each fire area. Based on the fire extinguishing priority and resource allocation ratio of each fire area, the optimal fire extinguishing action sequence for each fire area in at least one fire area is obtained. The optimal fire extinguishing action sequence includes the fire extinguishing actions that the fire extinguishing system needs to execute at each time step in multiple consecutive time steps.
[0059] In some embodiments, the fire extinguishing decision agent adopts a hierarchical reinforcement learning architecture, which includes: a high-level policy network and at least one low-level execution network. The high-level policy network is used to receive global fire status data of the energy storage system and to perform fire extinguishing action reasoning on at least one fire area based on the global fire status data to obtain the fire extinguishing priority and resource allocation ratio of each fire area in the at least one fire area. The at least one low-level execution network corresponds one-to-one with at least one fire area. Each low-level execution network is used to obtain the optimal fire extinguishing action sequence for the corresponding fire area based on the fire extinguishing priority and resource allocation ratio of the corresponding fire area and the local fire status data of the fire area.
[0060] The fire suppression decision-making agent in this embodiment adopts a hierarchical reinforcement learning architecture. The high-level policy network receives global fire status data of the energy storage system (such as high-dimensional state feature vectors composed of normalized parameters such as temperature, gas concentration, and flame signal). It extracts abstract features related to fire development and fire suppression needs from the high-dimensional state feature vectors, eliminating the need for manual division of fire zones or structuring of the situation. It outputs the fire suppression priority and resource allocation ratio for each fire zone (e.g., 60% of the total fire extinguishing agent is allocated to fire zone 1, and 40% to fire zone 2). Each fire zone corresponds to a low-level execution network, which, under the resource and priority constraints allocated by the high-level policy network, outputs refined fire suppression actions such as nozzle flow rate and intermittent period based on the local fire status data.
[0061] It should be noted that the fire extinguishing action sequence refers to a series of fire extinguishing actions corresponding to fire commands executed by the fire extinguishing system within multiple consecutive time steps. These actions may include, but are not limited to, the opening degree of the selector valves for each zone, the spray flow rate, spray angle, and spray mode of each sprinkler head. For example, in time step 1: Sprinkler A is opened with a flow rate of 1.0 L / s and a spray angle of 45 degrees; in time step 2: if the flame does not significantly weaken, the flow rate is increased to 1.8 L / s, and sprinkler B is added at an angle of 60 degrees; in time step 3: as the flame area shrinks, the flow rate is reduced to 0.5 L / s, and the system switches to intermittent mode (spraying for 1 second, stopping for 2 seconds) for cooling.
[0062] In some embodiments, the training method for the fire extinguishing decision agent includes: training an initial model based on historical fire extinguishing experimental data of the energy storage system to obtain a fire extinguishing decision model. The fire extinguishing decision model is used to predict the fire state sequence and fire extinguishing efficiency for the next K time steps based on the fire state sequence and fire extinguishing action sequence of the current N time steps; training the initial agent using a deep reinforcement learning algorithm based on the fire extinguishing decision model to obtain the fire extinguishing decision agent; wherein the initial agent learns through trial and error interaction with the fire extinguishing decision model, and updates its parameters with the goal of maximizing the cumulative reward, so that the initial agent can learn to generate the optimal fire extinguishing action sequence under different fire scenarios; wherein N and K are both positive integers. For example, during training, the optimal fire extinguishing effect can be used as the reward, that is, achieving the fastest fire extinguishing capability with the minimum fire extinguishing agent consumption and preventing reignition.
[0063] In this embodiment, the trained fire extinguishing decision model is used as an environmental proxy model (which can also be regarded as a simulator) to replace the real physical environment, allowing the fire extinguishing decision agent to learn through millions of trial and error cycles, without the need for actual ignition.
[0064] In some embodiments, the initial model is a Long Short-Term Memory (LSTM) network or a Transformer time series prediction model. The input of the initial model includes a two-dimensional tensor composed of input feature vectors for N consecutive time steps. The input feature vectors for each time step include fire status data, fire extinguishing actions, and time codes.
[0065] As an example, historical data from fire suppression experiments of energy storage systems is collected. This historical data includes time-series monitoring data (battery and module temperatures, ambient temperature, typical gas concentrations such as CO and H2, flame detector data, thermal radiation flux, extinguishing agent injection time, extinguishing time, and other fire suppression effectiveness evaluation indicators) under different fire scenarios (different battery types, different fire ignition locations, etc.). The fire suppression effectiveness can be evaluated through the test results of these indicators. Based on this historical data, a Transformer time-series prediction model is trained using supervised learning or generative adversarial networks. This model can predict the fire state sequence (fire state evolution) and fire suppression effectiveness for the next K time steps based on the current N time-step fire state sequence and fire suppression action sequence. Using the Transformer time-series prediction model as input, the fire state sequence and fire suppression action sequence of the past N time steps are input, and the output is the fire state sequence for the next K time steps, used to predict the complete trajectory of the fire suppression process in one go. The input of a time series prediction model is generally a two-dimensional tensor with the shape of [sequence length, feature dimension]. The input of this example is a joint time series sequence of fire status and fire extinguishing action. For example, assuming the input sequence length M=10, with a time step every 0.5 seconds, the input feature vector corresponding to each time step includes: (1) fire status data (e.g., temperature, gas concentration, flame area, etc., a total of 30 dimensions), the executed fire extinguishing action (e.g., flow rate of each nozzle, angle, etc., a total of 8 dimensions), (3) time encoding (e.g., the position of the time step in a day, used to distinguish the difference between day and night temperature rise), which is optional; therefore, the dimension of the input feature vector corresponding to a time step = 30 (fire status) + 8 (fire extinguishing action) + 1 (time encoding) = 39 dimensions; the shape of the input tensor is [10, 39], representing the complete record of the past 10 time steps (i.e., the past 5 seconds); where the generator in the generative adversarial network is used to generate virtual but physically consistent data to expand the training samples.
[0066] The trained fire-fighting decision-making agent is deployed in the edge controller or station-level host of the energy storage system. When a real fire occurs, the system collects fire status data in real time through sensors. This data is then input into the fire-fighting decision-making agent, which obtains the optimal fire-fighting action sequence through forward reasoning.
[0067] Step S103: Map the optimal fire extinguishing action sequence of each fire zone in at least one fire zone into a fire command, and control the fire extinguishing system to extinguish the fire based on the fire command.
[0068] The optimal fire extinguishing action sequence obtained through reasoning is mapped into fire-fighting commands or control commands and executed; in the next control cycle, new fire status data is received and the above process is repeated.
[0069] Step S104: Based on the flame and temperature detection signals in the global fire status data of the next control cycle, determine whether the open flame has been completely extinguished. If the open flame has been completely extinguished, enter the reignition warning mode.
[0070] In some embodiments, entering a reignition alert mode includes: obtaining a reignition risk index based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system; and making a reignition alert decision based on the reignition risk index by a fire extinguishing decision agent according to a reignition handling strategy, thereby obtaining a reignition alert decision result.
[0071] Therefore, in this embodiment, after the fire command is executed, the fire extinguishing decision-making agent continuously receives real-time global fire status data, and determines whether the open flame has been completely extinguished based on the flame and temperature detection signals (i.e., flame detector data and temperature data) in the global fire status data of the next control cycle. When it is determined that the open flame has been extinguished, the system automatically switches to the reignition warning mode without immediately terminating the fire extinguishing action. In the reignition warning mode, the fire extinguishing decision-making agent analyzes the reignition risk based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration, and makes a decision based on the degree of reignition risk.
[0072] In some embodiments, a method for obtaining a reignition warning decision result by having a fire extinguishing decision-making agent make a reignition warning decision based on a reignition handling strategy, based on a reignition risk index, includes: when the reignition risk index is less than a first risk threshold, outputting a stop spraying decision through the fire extinguishing decision-making agent; when the reignition risk index is greater than or equal to the first risk threshold and less than or equal to a second risk threshold, outputting a cooling maintenance decision through the fire extinguishing decision-making agent, the cooling maintenance decision being used to instruct the fire extinguishing system to perform low-flow continuous spraying or intermittent spraying; when the reignition risk index is greater than the second risk threshold, acquiring global fire status data for the next control cycle to perform fire extinguishing action reasoning. As an example, the first risk threshold is 0.25, the second risk threshold is 0.60, and when the reignition risk index... When the concentration is below 0.25, the risk of reignition is determined to be extremely low, and the fire suppression decision-making agent outputs a stop spraying command, exiting the reignition alert mode; when... When the value is between 0.25 and 0.60, a risk of reignition is determined, and the fire suppression decision-making agent outputs a cooling maintenance action, either by continuous low-flow injection or intermittent injection, while maintaining temperature and gas concentration monitoring; when When the value is above 0.6, the risk of reignition is determined to be high. The fire extinguishing decision-making agent automatically reverts to the active fire extinguishing mode, calls steps S101-S102 again, generates a fire extinguishing action sequence to extinguish the fire, and then switches to the reignition warning mode after the fire is suppressed again.
[0073] In some embodiments, a method for obtaining a reignition risk index based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of an energy storage system includes: normalizing the current battery surface temperature, temperature rise rate, and characteristic gas concentration to obtain normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration; and performing a weighted summation of the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration to obtain the reignition risk index. As an example, the reignition risk index... The calculation formula is:
[0074] ;
[0075] in, , , These are the normalized values of battery surface temperature, temperature rise rate, and CO concentration, respectively. The maximum value used for normalization is the relevant parameter value of the battery when thermal runaway is triggered by each sensor before the fire extinguishing action is triggered. , , The weights for battery surface temperature, temperature rise rate, and CO concentration are respectively, such as 0.35, 0.35, and 0.3.
[0076] In some embodiments, this application further includes: acquiring relevant fire extinguishing data during the fire extinguishing process, optimizing the fire extinguishing decision model and the fire extinguishing decision agent based on the relevant fire extinguishing data, and obtaining the optimized fire extinguishing decision model and the fire extinguishing decision agent. As an example, in actual fire extinguishing tests, all time-series data, such as the evolution of the fire state (fire state data sequence), the sequence of fire extinguishing actions, and fire extinguishing effectiveness evaluation indicators, are fully recorded. The fire extinguishing effectiveness evaluation indicators include, but are not limited to, fire extinguishing time, extinguishing agent action time, extinguishing agent dosage, whether reignition occurs, maximum temperature, and cooling rate at measuring points. The above-mentioned new data is used as additional training samples to periodically perform incremental training or fine-tuning of the fire extinguishing decision model, improving the prediction accuracy of the environmental model. Based on the updated fire extinguishing decision model, an offline reinforcement learning algorithm is used to perform a new round of strategy optimization on the fire extinguishing decision agent. After simulation verification, the optimized fire extinguishing decision agent is updated and deployed online in the actual system.
[0077] This application continuously accumulates new training data through each real fire extinguishing test or actual fire extinguishing operation in engineering, and continuously optimizes the fire extinguishing decision model and fire extinguishing decision agent, forming a self-learning closed-loop mechanism of "data accumulation - model iteration - strategy improvement". This enables the fire extinguishing decision agent to continuously optimize based on the continuously accumulated fire extinguishing data, overcoming the shortcomings of traditional fixed strategies that cannot improve themselves.
[0078] The intelligent fire suppression method for electrochemical energy storage systems based on deep reinforcement learning in this application embodiment acquires fire status data in real time based on the control cycle. Based on this real-time fire status data, a fire suppression decision-making agent infers fire suppression actions for one or more fire zones, obtaining the fire suppression priority and resource allocation ratio for each fire zone. Based on the fire suppression priority and resource allocation ratio for each fire zone, the optimal fire suppression action sequence for each fire zone is generated. This method can dynamically and proactively adjust the fire suppression strategy according to the actual development of the fire, achieving dynamic and adaptive fire suppression and improving fire suppression efficiency. At the same time, this solution can optimize the fire suppression priority and resource allocation ratio for multiple fire zones, achieving synergy between global fire suppression resource optimization and local fine-grained execution. In the case of multiple fire points, it can intelligently allocate fire suppression resources and avoid resource misallocation.
[0079] To clearly illustrate the above embodiments, specific examples are provided below. The intelligent fire extinguishing method for electrochemical energy storage systems based on deep reinforcement learning of this application is applied to an intelligent fire extinguishing system for electrochemical energy storage systems based on deep reinforcement learning, such as... Figure 2 As shown, this intelligent fire extinguishing system based on deep reinforcement learning and electrochemical energy storage system includes:
[0080] The on-site perception and execution layer, deployed inside the energy storage compartment of the energy storage system, includes a data acquisition unit and an execution unit. The data acquisition unit includes a multimodal sensor network deployed inside the energy storage system, which is used to collect global fire status data in real time and upload it to the station-level host. The execution unit is used to receive and execute fire-fighting commands issued by the station-level host. After the fire-fighting command is executed, the execution unit is also used to return the execution status and preliminary fire-fighting effectiveness evaluation indicators to the station-level host. The global fire status data includes fire status data of at least one fire area.
[0081] The edge decision control layer, deployed locally at the station-level host of the energy storage power station, connects to the data acquisition unit and execution unit. It receives global fire status data uploaded by the data acquisition unit, performs real-time fusion processing on the global fire status data, and inputs the fused global fire status data into the fire suppression decision-making agent. The fire suppression decision-making agent then performs fire suppression action reasoning on at least one fire area, obtaining the fire suppression priority and resource allocation ratio for each fire area. Based on the fire suppression priority and resource allocation ratio for each fire area, it obtains the optimal fire suppression action sequence for each fire area within at least one fire area. The optimal fire suppression action sequence includes the fire suppression system's actions in multiple consecutive fire areas. The fire extinguishing actions to be performed at each time step in the time step map the optimal fire extinguishing action sequence for each fire zone in at least one fire zone into fire commands and issue them to the execution unit. The edge decision control layer is also used to obtain the effect feedback data after the fire commands are executed, and upload the effect feedback data to the cloud optimization training layer after associating and packaging it with the global fire status data. The edge decision control layer is also used to generate the optimal fire extinguishing action sequence based on the backup fixed strategy in the event of reasoning failure or communication abnormality of the fire extinguishing decision agent. The effect feedback data includes the status feedback of the execution unit (such as whether the specified valve is open, pipeline flow, etc.) and the feedback of changes in fire status.
[0082] The cloud-based optimization training layer, deployed on a remote server or industrial cloud platform, receives global fire status data and effect feedback data uploaded by the station-level host and adds it to the historical fire suppression case library as historical fire suppression data. The cloud-based optimization training layer includes a model training module and a reinforcement learning strategy optimization module. The model training module and the reinforcement learning strategy optimization module are used to optimize the fire suppression decision model and fire suppression decision agent based on historical fire suppression data, and push the optimized fire suppression decision agent to the station-level host to realize the continuous evolution of the fire suppression strategy.
[0083] Therefore, the intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage in this example adopts a three-level collaborative architecture of "end-edge-cloud" to realize real-time on-site fire perception, edge intelligent decision-making and continuous optimization in the cloud.
[0084] To achieve the above embodiments, this application also proposes an intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage systems. Figure 3 This is a schematic diagram of the structure of an intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage, provided as an embodiment of this application. Figure 3 As shown, this intelligent fire extinguishing system based on deep reinforcement learning and electrochemical energy storage system may include:
[0085] The data acquisition module 301 is used to acquire global fire status data of the energy storage system based on the control cycle. The global fire status data includes fire status data of at least one fire zone.
[0086] The action reasoning module 302 is used to perform fire extinguishing action reasoning on at least one fire area based on global fire status data through a fire extinguishing decision-making intelligent agent, to obtain the fire extinguishing priority and resource allocation ratio of each fire area, and to obtain the optimal fire extinguishing action sequence for each fire area based on the fire extinguishing priority and resource allocation ratio of each fire area. The optimal fire extinguishing action sequence includes the fire extinguishing actions that the fire extinguishing system needs to execute at each time step in multiple consecutive time steps.
[0087] The fire extinguishing execution module 303 is used to map the optimal fire extinguishing action sequence of each fire zone in at least one fire zone into fire-fighting instructions, and control the fire extinguishing system to extinguish the fire based on the fire-fighting instructions.
[0088] Furthermore, in one possible implementation of this application embodiment, the fire extinguishing decision-making agent adopts a hierarchical reinforcement learning architecture, which includes:
[0089] The high-level strategy network is used to receive global fire status data of the energy storage system, and to infer fire extinguishing actions for at least one fire area based on the global fire status data, so as to obtain the fire extinguishing priority and resource allocation ratio of each fire area in at least one fire area.
[0090] At least one low-level execution network is provided, and each low-level execution network corresponds one-to-one with at least one fire zone. Each low-level execution network is used to obtain the optimal fire extinguishing action sequence for the corresponding fire zone based on the fire extinguishing priority and resource allocation ratio of the fire zone and the local fire status data of the fire zone.
[0091] Furthermore, in one possible implementation of this application embodiment, the system further includes a reignition warning module 304, used for:
[0092] Based on the flame and temperature detection signals in the global fire status data of the next control cycle, determine whether the open flame has been completely extinguished;
[0093] If the open flame has been completely extinguished, the reignition risk index is obtained based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system.
[0094] Based on the reignition risk index, the fire extinguishing decision-making agent makes a reignition warning decision according to the reignition handling strategy, and obtains the reignition warning decision result.
[0095] Furthermore, in one possible implementation of this application embodiment, when the reignition warning module 304 makes a reignition warning decision based on the reignition risk index and the reignition handling strategy by the fire extinguishing decision-making agent, and obtains the reignition warning decision result, it is used for:
[0096] If the reignition risk index is less than the first risk threshold, the fire extinguishing decision-making agent outputs a decision to stop spraying.
[0097] When the reignition risk index is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, the fire extinguishing decision agent outputs a cooling maintenance decision, which is used to instruct the fire extinguishing system to perform low-flow continuous spraying or intermittent spraying.
[0098] If the reignition risk index is greater than the second risk threshold, obtain the global fire status data for the next control cycle to infer fire extinguishing actions.
[0099] Furthermore, in one possible implementation of this application embodiment, when the reignition warning module 304 obtains the reignition risk index based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system, it is used to:
[0100] The current battery surface temperature, temperature rise rate, and characteristic gas concentration are normalized to obtain the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration.
[0101] The reignition risk index is obtained by weighted summation of the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration.
[0102] Furthermore, in one possible implementation of this application embodiment, the system further includes a model training module 305, used for:
[0103] Based on historical fire suppression experimental data of energy storage systems, an initial model is trained to obtain a fire suppression decision model. The fire suppression decision model is used to predict the fire state sequence and fire suppression effectiveness for the next K time steps based on the fire state sequence and fire suppression action sequence of the current N time steps.
[0104] Based on the fire extinguishing decision-making model, a deep reinforcement learning algorithm is used to train the initial agent, resulting in a fire extinguishing decision-making agent; among which...
[0105] The initial agent learns through trial and error with the fire extinguishing decision model, and updates its parameters with the goal of maximizing the cumulative reward, so that the initial agent can learn to generate the optimal fire extinguishing action sequence under different fire scenarios.
[0106] Furthermore, in one possible implementation of this application embodiment, the initial model is a long short-term memory network or a Transformer time series prediction model. The input of the initial model includes a two-dimensional tensor composed of input feature vectors of N consecutive time steps. The input feature vector of each time step includes fire status data, fire extinguishing actions, and time codes.
[0107] Furthermore, in one possible implementation of this application embodiment, the model training module 305 is also used for:
[0108] Acquire relevant fire extinguishing data during the fire extinguishing process, and optimize the fire extinguishing decision model and fire extinguishing decision agent based on the relevant fire extinguishing data to obtain the optimized fire extinguishing decision model and fire extinguishing decision agent.
[0109] It should be noted that the foregoing explanation of the embodiment of the intelligent fire extinguishing method for electrochemical energy storage system based on deep reinforcement learning also applies to the intelligent fire extinguishing system for electrochemical energy storage system based on deep reinforcement learning in this embodiment, and will not be repeated here.
[0110] To implement the above embodiments, this application also proposes an electronic device. Please see [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 4 As shown, the electronic device 400 includes: a processor 401, and a memory 402 communicatively connected to the processor 401; the memory 402 stores computer execution instructions; the processor 401 executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.
[0111] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.
[0112] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.
[0113] In the foregoing descriptions of the embodiments, the terms "some embodiments," "examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0114] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0115] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. A smart fire extinguishing method for electrochemical energy storage systems based on deep reinforcement learning, characterized in that, include: Based on the control cycle, the global fire status data of the energy storage system is obtained, and the global fire status data includes fire status data of at least one fire zone. Based on the global fire status data, the fire extinguishing decision-making agent performs fire extinguishing action reasoning on the at least one fire area to obtain the fire extinguishing priority and resource allocation ratio of each fire area. Based on the fire extinguishing priority and resource allocation ratio of each fire area, the optimal fire extinguishing action sequence for each fire area in the at least one fire area is obtained. The optimal fire extinguishing action sequence includes the fire extinguishing actions that the fire extinguishing system needs to execute at each time step in multiple consecutive time steps. The optimal fire extinguishing action sequence for each fire zone in the at least one fire zone is mapped to a fire command, and the fire extinguishing system is controlled to extinguish the fire based on the fire command.
2. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 1, characterized in that, The fire extinguishing decision-making agent adopts a hierarchical reinforcement learning architecture, which includes: A high-level strategy network is used to receive global fire status data of the energy storage system, and to perform fire extinguishing action reasoning on the at least one fire area based on the global fire status data, so as to obtain the fire extinguishing priority and resource allocation ratio of each fire area in the at least one fire area. At least one low-level execution network, which corresponds one-to-one with the at least one fire zone, and each low-level execution network is used to obtain the optimal fire extinguishing action sequence for the corresponding fire zone based on the fire extinguishing priority and resource allocation ratio of the fire zone and the local fire status data of the fire zone.
3. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 1, characterized in that, After controlling the fire extinguishing system to extinguish the fire based on the fire command, the method further includes: Based on the flame and temperature detection signals in the global fire status data of the next control cycle, determine whether the open flame has been completely extinguished; If the open flame has been completely extinguished, the reignition risk index is obtained based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system. Based on the reignition risk index, the fire extinguishing decision-making agent makes a reignition warning decision according to the reignition handling strategy, and obtains the reignition warning decision result.
4. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 3, characterized in that, Based on the reignition risk index, the fire extinguishing decision-making agent makes a reignition warning decision according to the reignition handling strategy, and obtains the reignition warning decision result, including: If the reignition risk index is less than the first risk threshold, the fire extinguishing decision-making agent outputs a stop spraying decision. When the reignition risk index is greater than or equal to the first risk threshold and less than or equal to the second risk threshold, the fire extinguishing decision agent outputs a cooling maintenance decision, which is used to instruct the fire extinguishing system to perform low-flow continuous spraying or intermittent spraying. If the reignition risk index is greater than the second risk threshold, global fire status data for the next control cycle is obtained to infer fire extinguishing actions.
5. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 3, characterized in that, The method for obtaining a reignition risk index based on the current battery surface temperature, temperature rise rate, and characteristic gas concentration of the energy storage system includes: The current battery surface temperature, temperature rise rate, and characteristic gas concentration are normalized to obtain the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration. The reignition risk index is obtained by weighted summation of the normalized current battery surface temperature, temperature rise rate, and characteristic gas concentration.
6. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 1, characterized in that, The training method for the fire extinguishing decision-making agent includes: Based on historical fire suppression experimental data of energy storage systems, an initial model is trained to obtain a fire suppression decision model. The fire suppression decision model is used to predict the fire state sequence and fire suppression effectiveness for the next K time steps based on the fire state sequence and fire suppression action sequence of the current N time steps. Based on the aforementioned fire extinguishing decision-making model, an initial agent is trained using a deep reinforcement learning algorithm to obtain the fire extinguishing decision-making agent; wherein... The initial agent learns through trial and error with the fire extinguishing decision model, and updates its parameters with the goal of maximizing the cumulative reward, so that the initial agent can learn to generate the optimal fire extinguishing action sequence under different fire scenarios.
7. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 6, characterized in that, The initial model is a long short-term memory network or a Transformer time series prediction model. The input of the initial model includes a two-dimensional tensor composed of input feature vectors of N consecutive time steps. The input feature vector of each time step includes fire status data, fire extinguishing actions, and time codes.
8. The intelligent fire extinguishing method for an electrochemical energy storage system based on deep reinforcement learning according to claim 1, characterized in that, The method further includes: Acquire relevant fire extinguishing data during the fire extinguishing process, and optimize the fire extinguishing decision model and the fire extinguishing decision agent based on the relevant fire extinguishing data to obtain the optimized fire extinguishing decision model and the fire extinguishing decision agent.
9. An intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage, characterized in that, The system includes: The on-site perception and execution layer, deployed inside the energy storage compartment of the energy storage system, includes a data acquisition unit and an execution unit. The data acquisition unit includes a multimodal sensor network deployed inside the energy storage system, which is used to collect global fire status data in real time and upload it to the station-level host. The execution unit is used to receive and execute fire-fighting commands issued by the station-level host. After the fire-fighting command is executed, the execution unit is also used to return the execution status and preliminary fire-fighting effectiveness evaluation indicators to the station-level host. The global fire status data includes fire status data of at least one fire area. The edge decision control layer, deployed locally at the energy storage power station as a station-level host, is connected to the data acquisition unit and the execution unit. It receives global fire status data uploaded by the data acquisition unit, performs real-time fusion processing on the global fire status data, and inputs the fused global fire status data into the fire extinguishing decision-making agent. The fire extinguishing decision-making agent then performs fire extinguishing action reasoning on the at least one fire area to obtain the fire extinguishing priority and resource allocation ratio for each fire area. Based on the fire extinguishing priority and resource allocation ratio for each fire area, it determines the optimal fire extinguishing action for each fire area within the at least one fire area. The optimal fire extinguishing action sequence includes the fire extinguishing actions required by the fire extinguishing system at each time step in multiple consecutive time steps. The optimal fire extinguishing action sequence for each fire area in the at least one fire area is mapped to a fire command and issued to the execution unit. The edge decision control layer is also used to obtain the effect feedback data after the fire command is executed, and upload the effect feedback data and the global fire status data to the cloud optimization training layer after packaging. The edge decision control layer is also used to generate the optimal fire extinguishing action sequence based on a backup fixed strategy in the event that the fire extinguishing decision agent fails to reason or communication is abnormal. The cloud-based optimization training layer, deployed on a remote server or industrial cloud platform, is used to receive global fire status data and effect feedback data uploaded by the station-level host and add them to the historical fire extinguishing case library as historical fire extinguishing data. The cloud-based optimization training layer includes a model training module and a reinforcement learning strategy optimization module. The model training module and the reinforcement learning strategy optimization module are used to optimize the fire extinguishing decision model and the fire extinguishing decision agent based on the historical fire extinguishing data, and push the optimized fire extinguishing decision agent to the station-level host.
10. An intelligent fire extinguishing system based on deep reinforcement learning for electrochemical energy storage, characterized in that, include: The data acquisition module is used to acquire global fire status data of the energy storage system based on the control cycle. The global fire status data includes fire status data of at least one fire zone. The action reasoning module is used to perform fire extinguishing action reasoning on the at least one fire area based on the global fire status data through the fire extinguishing decision-making intelligent agent, to obtain the fire extinguishing priority and resource allocation ratio of each fire area, and to obtain the optimal fire extinguishing action sequence for each fire area in the at least one fire area based on the fire extinguishing priority and resource allocation ratio of each fire area. The optimal fire extinguishing action sequence includes the fire extinguishing actions that the fire extinguishing system needs to perform at each time step in multiple consecutive time steps. The fire extinguishing execution module is used to map the optimal fire extinguishing action sequence of each fire zone in the at least one fire zone into a fire command, and control the fire extinguishing system to extinguish the fire based on the fire command.