System and method for automatically generating intelligent emergency disposal process of energy storage power station

Abstract

The invention relates to the technical field of energy storage power station emergency disposal, in particular to an energy storage power station intelligent emergency disposal process automatic generation system and method.The system adopts a scheme of multi-modal fault understanding, large language model reasoning and multi-target optimization scheduling, receives multi-modal input such as texts, alarm codes, temperature / voltage / alarm marks and the like, and sends the multi-modal input to a server; the method comprises the following steps: firstly, taking GPT-4o as a scene vector, carrying out double-tower Transform coding on the GPT-4o as a scene vector, generating a candidate plan by combining the GPT-4o subjected to LoRA fine tuning with a knowledge graph, screening an optimal scheme through a safety, timeliness and cost ternary optimization function psi and a particle swarm optimization algorithm, and carrying out digital twinborn simulation verification and weekly-level closed-loop learning iteration; according to the method, a standardized plan is generated in 1.7 seconds, the compiling efficiency is improved by 103 times, the execution consistency is improved by 23%, the annual operation and maintenance accident loss is reduced by 18%, the monthly newly-added scene coverage rate is 92%, the deployment operation and maintenance confirmation rate in a 100 MWh energy storage station is 96%, and the energy storage safety operation and maintenance intelligent level is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of emergency disposal of energy storage power stations, in particular to an intelligent emergency disposal process automatic generation system and method for energy storage power stations. BACKGROUND

[0002] Large-scale energy storage power stations have complex operating scenarios, and common faults include cell thermal runaway, PCS pressure loss, gas leakage, BMS communication abnormalities, and power grid mutations. Corresponding energy storage emergency plans are needed for these faults, and existing plans are mainly written by humans.

[0003] The existing energy storage emergency plans written by humans have the problems of general trigger conditions and lack of pertinence, different step granularities and difficulty in execution, low update frequency, and inability to cover new scenarios. At the same time, existing technologies mostly use templated static documents in PDF / Word format, which cannot be parsed by machines, the rule engine used is relatively rigid, new scenarios need to be hard-coded, and there is a lack of reasoning chain, making it difficult to explain the source and logic of emergency steps, and unable to meet the actual needs of complex fault emergency disposal of large-scale energy storage power stations. Therefore, in view of the above status quo, there is an urgent need to develop an intelligent emergency disposal process automatic generation system and method for energy storage power stations to overcome the deficiencies in current actual applications. SUMMARY

[0004] The purpose of the present application is to provide an intelligent emergency disposal process automatic generation system and method for energy storage power stations to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides the following technical solutions: An intelligent emergency disposal process automatic generation system for energy storage power stations, comprising: A scene input layer for receiving and encoding multi-modal fault information to obtain a fault scene semantic vector; A plan generation layer for generating and verifying a candidate emergency disposal plan based on the fault scene semantic vector and energy storage power station configuration information, in combination with a large language model and a knowledge graph; A scheme evaluation and optimization layer for quantitatively evaluating the candidate emergency disposal plan using a multi-objective optimization function and selecting an optimal emergency disposal plan through an optimization algorithm; An execution output layer for outputting the optimal emergency disposal plan and writing a plan execution log back to the knowledge graph to provide data support for knowledge base iteration.

[0006] As a further solution of the present application, the multi-modal fault information received by the scene input layer includes text description, alarm code, and telemetry data, and the telemetry data includes energy storage cluster temperature, bus voltage, and alarm flag.

[0007] As a further aspect of the present invention: the scene input layer adopts a multimodal encoder with a dual-tower Transformer structure, which encodes the multimodal fault information into text input and numerical input parts respectively, and then obtains the fault scene semantic vector by concatenating the vectors.

[0008] As a further aspect of the present invention: the large language model in the contingency plan generation layer is a LoRA-tuned GPT-4o model, and after the candidate emergency response plan is verified for consistency by the knowledge graph, steps that do not conform to the actual configuration and fault handling logic of the energy storage power station are corrected.

[0009] As a further aspect of the present invention: the multi-objective optimization function is a ternary optimization function of safety, timeliness and cost, and the optimization algorithm is a particle swarm optimization algorithm.

[0010] A method for automatically generating intelligent emergency response procedures for energy storage power stations includes the following steps: S1: Encode the input multimodal fault information to obtain the fault scene semantic vector; S2: Based on the semantic vector of the fault scenario and the configuration information of the energy storage power station, candidate emergency response plans are generated through a large language model and then verified and corrected using a knowledge graph. S3: The candidate emergency response plans are quantitatively evaluated using a multi-objective optimization function, and the optimal emergency response plan is selected using an optimization algorithm. S4: Output the optimal emergency response plan, write the execution log back to the knowledge graph, and iterate the knowledge base through automatic verification and closed-loop learning.

[0011] As a further aspect of the present invention: In step S1, a multimodal encoder with a dual-tower Transformer structure is used to encode and concatenate the text input part and the numerical input part in the multimodal fault information to obtain the semantic vector of the fault scene.

[0012] As a further aspect of the present invention: in step S2, the fault information corresponding to the semantic vector of the fault scenario and the configuration information of the energy storage power station are input into the LoRA-tuned GPT-4o model in a preset format to generate the candidate emergency response plan.

[0013] As a further aspect of the present invention: In step S3, the multi-objective optimization function calculates the safety score, execution time, and economic cost of each step in each candidate emergency response plan, and performs normalization calculation by combining the weight coefficients. The optimization algorithm is a particle swarm optimization algorithm, which selects the optimal emergency response plan that minimizes the value of the multi-objective optimization function.

[0014] As a further aspect of the present invention: in step S4, the success rate of the execution of the optimal emergency response plan is evaluated by digital twin simulation, and the RLHF reward model is used to perform weekly incremental fine-tuning of the system to realize the absorption and iteration of new scenarios by the knowledge base.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The system employs a multimodal fault understanding, large language model (LoRA fine-tuned GPT-4o) inference, and multi-objective optimization scheduling technology. It supports multimodal inputs such as text, alarm codes, and telemetry data (temperature, voltage, alarm flags), which are encoded and processed. LLM combined with a knowledge graph rapidly generates candidate plans, which are then optimized and selected for output. This reduces plan generation time to only 1.7 seconds, improving compilation efficiency by 10%. 3 times; The system uses a ternary multi-objective optimization function (Ψ) to quantitatively evaluate each step of candidate contingency plans based on safety, timeliness, and cost, and selects the optimal contingency plan. The output emergency steps are interpretable, quantifiable, and executable, and support direct access via API / HMI, thereby improving execution consistency by 23% and reducing annual maintenance accident losses by 18%. The system possesses a self-evolving knowledge base and an automatic verification and closed-loop learning framework, enabling it to absorb new scenarios weekly and generate test cases automatically. Through digital twin simulation evaluation and the RLHF reward model, it performs weekly incremental fine-tuning, achieving a monthly coverage rate of 92% for newly added scenarios. The system integrates LLM reasoning and graph knowledge to establish a closed loop of "fault identification → contingency plan generation → verification and execution → knowledge accumulation". It outputs machine-parseable JSON, Markdown and API Hook, and writes execution logs back to the graph. During the deployment of the 100MWh energy storage station in East China (Q1-Q2 2025), it successfully handled 47 faults, with a maintenance confirmation rate of 96% and an average maintenance time saving of 42 minutes. Attached Figure Description

[0016] Figure 1 This is an overall structural diagram of the automatic generation system for emergency response procedures of energy storage power stations in this embodiment of the invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0019] Please see Figure 1 This invention provides an automatic generation system and method for intelligent emergency response processes in energy storage power stations, aiming to address the numerous defects and shortcomings of existing technologies in the manual writing of energy storage emergency plans. It adopts a core scheme of "multimodal fault understanding + large language model (LLM) reasoning + multi-objective optimization scheduling" to achieve the goals of inputting fault scenarios into standardized plans within seconds, making the steps interpretable, quantifiable, and executable, and enabling the knowledge base to evolve automatically. The following is a detailed description of the specific implementation process.

[0020] The intelligent emergency response process automatic generation system for energy storage power stations in this embodiment adopts a four-layer structure design. Each layer works collaboratively to complete the automatic generation and optimization of the emergency response process, as detailed below: I. Scene Input Layer The scene input layer supports multimodal fault information input, including text descriptions, alarm codes, and telemetry data. The telemetry data includes the energy storage cluster temperature vector T, the bus voltage vector V, and the alarm flag set A.

[0021] For the input multimodal data, the system processes it using a multimodal encoder. This encoder employs a dual-tower Transformer structure, dividing the input sequence x into a text input portion. Sum of numbers input section Each through text encoding function right Encoding, numerical encoding functions right Encode the data, then concatenate the encoded results to obtain the fault scene semantic vector s, i.e.: ; in, ,‖ represents the vector concatenation operation.

[0022] This design can comprehensively receive different types of fault information and transform them into machine-processable scenario vectors s through a unified encoding method, ensuring the comprehensiveness and accuracy of fault scenario understanding and providing a precise data foundation for subsequent contingency plan generation.

[0023] II. Contingency Plan Generation Layer The core of the contingency plan generation layer is a multi-step contingency plan generator with LLM and graph alignment, where LLM adopts the LoRA-fine-tuned GPT-4o model.

[0024] The system organizes the fault information corresponding to the scene vector s obtained from the scene input layer into...<FAULT_JSON> The format, combined with the configuration information of the energy storage power station.<PLANT_CFG> As a prompt input to the fine-tuned GPT-4o model, for example: <FAULT_JSON>for: {"symptom":"PCSUndervoltage","battery_maxT":58,"gas_ppm":80};"The fault symptoms are PCS undervoltage, battery maximum temperature 58℃, and gas concentration in the cabinet 80ppm"; <PLANT_CFG> for: {"rated_MW":50,"cooling":"liquid"};"The energy storage power station has a rated power of 50 megawatts and is cooled by liquid cooling."

[0025] The GPT-4o model generates a list of candidate solutions in YAML format, containing multiple steps, based on the input Prompt. These candidate solutions are denoted as P1,…,P… n Each candidate plan P i Contains m steps .

[0026] After generating candidate plans, the system uses a knowledge graph to verify the consistency of each candidate plan and corrects steps that do not conform to the actual configuration of the power plant and the logic of fault handling.

[0027] This layer leverages the powerful reasoning capabilities of LLM combined with the verification functions of knowledge graphs to quickly generate candidate contingency plans that meet the needs of real-world scenarios. It also effectively ensures the rationality and consistency of emergency response steps, solving the problems of rigidity in existing rule engines and the need for hard coding of new scenarios. Furthermore, it makes emergency response steps interpretable, clarifying their origin and logic.

[0028] III. Scheme Evaluation and Optimization Layer The scheme evaluation and optimization layer adopts a ternary multi-objective optimization function Ψ, which considers safety, timeliness, and cost, to evaluate each candidate scheme P. i Each step To conduct a quantitative assessment, the first step is to calculate each step. Security score ( ∈[0,1]), execution time (Unit: s) and economic costs (Unit: CNY): .

[0029] set up This represents the maximum execution time of all steps in the same batch of candidate plans. To maximize the economic cost of all steps in the same batch of candidate plans, define a multi-objective optimization function: Where α, β, and γ are weighting coefficients, and α+β+γ=1.

[0030] The system uses the Particle Swarm Optimization (PSO) algorithm to search for candidate solutions that minimize the optimization function Ψ, i.e.: Should This is the optimal emergency response plan.

[0031] This multi-objective optimization design can comprehensively balance the safety, timeliness, and economy of emergency response, select the optimal solution, effectively improve the scientificity and rationality of emergency response, improve execution consistency by 23%, and reduce annual operation and maintenance accident losses by 18%.

[0032] IV. Execution Output Layer The execution output layer combines the optimal solution obtained from the scheme evaluation and optimization layers. Output is provided in machine-parseable JSON, Markdown, and API Hook formats, supporting direct execution via API / HMI for quick access and implementation by on-site operations and maintenance personnel. Simultaneously, this layer writes the execution logs of the contingency plan back to the knowledge graph, providing data support for knowledge base updates and iterations.

[0033] This design solves the problem that existing templated static documents cannot be parsed by machines, improves the execution efficiency of the contingency plan, and lays the foundation for subsequent knowledge accumulation.

[0034] The automatic generation method for intelligent emergency response process of energy storage power station of the present invention specifically includes the following steps: Step 1: Scene Semantic Encoding. A multimodal encoder with a dual-tower Transformer structure is used to encode the input text, alarm code, temperature T, voltage V, and alarm flag A, etc., to obtain the fault scene semantic vector s, thereby realizing a unified representation of multimodal fault information and ensuring the comprehensiveness and accuracy of fault scene understanding.

[0035] Step 2: LLM contingency plan generation. Fault scenario information and power plant configuration information are input into the LoRA-tuned GPT-4o model in a specific Prompt format to generate a list of candidate contingency plans in YAML format. This list is then validated and corrected using a knowledge graph to ensure that the candidate contingency plans meet the actual scenario requirements, and that the steps are interpretable and reasonable.

[0036] Step 3: Multi-objective optimization and selection. First, estimate the safety score of each candidate solution step. Execution time and economic costs The execution time and economic cost are normalized, and then the optimal plan is selected by using a multi-objective optimization function Ψ and a particle swarm optimization algorithm. To achieve the best overall balance between safety, timeliness, and cost.

[0037] Step 4: Automatic Verification and Closed-Loop Update. The optimal contingency plan is verified through digital twin simulation. The execution success rate is evaluated, and the execution logs are used in the RLHF reward model to make weekly incremental fine-tuning of the system, enabling the knowledge base to absorb new scenarios and generate test cases on a weekly basis, realizing the self-evolution of the knowledge base, with a monthly new scenario coverage rate of 92%.

[0038] The following example further illustrates this: When the input fault scenario is a cell temperature of 62℃, PCS undervoltage, and H2 concentration of 120ppm in the cabinet, the system obtains the corresponding scenario vector s through scenario semantic encoding. This vector is then input into the fine-tuned GPT-4o model and, combined with the power plant configuration information, generates three candidate contingency plans. After evaluation by the multi-objective optimization function Ψ and selection by the particle swarm optimization algorithm, the optimal contingency plan is output with the following steps in sequence: current reduction → PCS disconnection → liquid cooling flow increase → ventilation. This contingency plan generation takes only 1.7 seconds, improving efficiency by 10% compared to manual drafting. 3 times.

[0039] This system has been deployed at a 100MWh energy storage station in East China (deployment period: Q1-Q2 2025). During this period, a total of 47 faults were handled. After confirmation by operation and maintenance personnel, the confirmation rate of the contingency plan meeting the actual handling needs reached 96%. Each fault handling saved an average of 42 minutes, significantly improving the efficiency and safety of emergency handling at the energy storage power station. This fully demonstrates that the system and method provided by this invention have good practicality and reliability.

[0040] In summary, this invention, by integrating LLM reasoning and graph knowledge, establishes a closed loop of "fault identification → contingency plan generation → verification and execution → knowledge accumulation," comprehensively solving the shortcomings of existing technologies, significantly improving the intelligent level of energy storage safety operation and maintenance, and has broad application prospects.

[0041] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An automated system for generating intelligent emergency response procedures for energy storage power stations, characterized in that, include: The scene input layer is used to receive multimodal fault information and encode it to obtain a fault scene semantic vector; The contingency plan generation layer generates and verifies candidate emergency response plans based on the semantic vectors of the fault scenarios and the configuration information of the energy storage power station, combined with a large language model and knowledge graph. The scheme evaluation and optimization layer uses a multi-objective optimization function to quantitatively evaluate the candidate emergency response plans and selects the optimal emergency response plan through an optimization algorithm. The execution output layer outputs the optimal emergency response plan and writes the plan execution log back to the knowledge graph to provide data support for knowledge base iteration.

2. The automatic generation system for intelligent emergency response procedures of energy storage power stations according to claim 1, characterized in that, The multimodal fault information received by the scene input layer includes text descriptions, alarm codes, and telemetry data, including energy storage cluster temperature, bus voltage, and alarm flags.

3. The automatic generation system for intelligent emergency response procedures of energy storage power stations according to claim 1, characterized in that, The scene input layer adopts a multimodal encoder with a dual-tower Transformer structure. The multimodal fault information is divided into text input part and numerical input part, which are encoded separately and then concatenated to obtain the semantic vector of the fault scene.

4. The automatic generation system for intelligent emergency response procedures of energy storage power stations according to claim 1, characterized in that, The large language model in the contingency plan generation layer is the LoRA-tuned GPT-4o model. After the candidate emergency response plans are verified for consistency by the knowledge graph, steps that do not conform to the actual configuration and fault handling logic of the energy storage power station are corrected.

5. The automatic generation system for intelligent emergency response procedures of energy storage power stations according to claim 1, characterized in that, The multi-objective optimization function is a ternary optimization function of safety, timeliness, and cost, and the optimization algorithm is the particle swarm optimization algorithm.

6. A method for automatically generating intelligent emergency response procedures for energy storage power stations, characterized in that, Includes the following steps: S1: Encode the input multimodal fault information to obtain the fault scene semantic vector; S2: Based on the semantic vector of the fault scenario and the configuration information of the energy storage power station, candidate emergency response plans are generated through a large language model and then verified and corrected using a knowledge graph. S3: The candidate emergency response plans are quantitatively evaluated using a multi-objective optimization function, and the optimal emergency response plan is selected using an optimization algorithm. S4: Output the optimal emergency response plan, write the execution log back to the knowledge graph, and iterate the knowledge base through automatic verification and closed-loop learning.

7. The method for automatically generating intelligent emergency response procedures for energy storage power stations according to claim 6, characterized in that, In step S1, a multimodal encoder with a dual-tower Transformer structure is used to encode the text input part and the numerical input part in the multimodal fault information respectively and then concatenate them to obtain the semantic vector of the fault scene.

8. The method for automatically generating intelligent emergency response procedures for energy storage power stations according to claim 6, characterized in that, In step S2, the fault information corresponding to the semantic vector of the fault scenario and the configuration information of the energy storage power station are input into the LoRA-tuned GPT-4o model in a preset format to generate the candidate emergency response plan.

9. The method for automatically generating intelligent emergency response procedures for energy storage power stations according to claim 6, characterized in that, In step S3, the multi-objective optimization function calculates the safety score, execution time, and economic cost of each step in each candidate emergency response plan, and performs normalization calculations by combining weight coefficients. The optimization algorithm is a particle swarm optimization algorithm, which selects the optimal emergency response plan that minimizes the value of the multi-objective optimization function.

10. The method for automatically generating intelligent emergency response procedures for energy storage power stations according to claim 6, characterized in that, In step S4, the success rate of the optimal emergency response plan is evaluated by digital twin simulation, and the RLHF reward model is used to make weekly incremental fine-tuning of the system to realize the absorption and iteration of new scenarios by the knowledge base.

Images