Nuclear power station operation optimization method and system, electronic equipment and storage medium

Abstract

The invention provides a nuclear power station operation optimization method and system, electronic equipment and a storage medium. The nuclear power station operation optimization method comprises the steps of classifying acquired historical operation data and real-time operation data of a nuclear power station; inputting high-security-level and strong-real-time data obtained from a data classification result into an AI algorithm module, and inputting obtained non-real-time data or unstructured data into a large model module; the AI algorithm prediction result and the large model prediction result are subjected to tradeoff fusion processing, and a fault detection result of the nuclear power station is obtained; inputting a fault detection result into a nuclear power station analog machine for operation; and optimizing the operation of the nuclear power station based on the simulation operation result. According to the invention, the AI algorithm module and the large model module are combined, the simulation operation result is obtained by combining the nuclear power station simulator, and the operation of the nuclear power station is optimized based on the simulation operation result, so that the intelligent decision-making capability is improved, the verification cost is reduced, and a closed-loop iterative optimization system is constructed.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and nuclear engineering technology, and in particular to an optimization method, system, electronic device and storage medium for the operation of a nuclear power plant. Background Technology

[0002] Currently, nuclear power plant operation optimization mainly relies on two types of methods: (1) Manual experience-driven: Operators formulate operating strategies based on historical data and experience, which has the disadvantages of slow response speed, single strategy, and difficulty in dealing with complex operating conditions. (2) Offline model optimization: Strategies are generated based on mechanistic models or shallow machine learning algorithms, but it requires a large number of physical experiments for verification, which has a long cycle (usually several months), high cost, and is difficult to adapt to the needs of real-time dynamic adjustment.

[0003] Regarding nuclear power plant operation optimization, current methods employ neural network models to optimize mathematical models in simulators, enabling full-range nuclear power plant simulators to more accurately reflect the actual operating status of nuclear power plants. However, these methods lack verification methods for candidate strategies and processing and analysis of unstructured data. Another approach involves inputting target parameter data from the nuclear power plant into a condition diagnosis model to diagnose its operating conditions. Since this model outputs the probability distribution of the nuclear power plant under different operating conditions, it offers greater interpretability and more reliable diagnostic results compared to AI models that only output one operating condition. However, this method addresses the diagnosis and interpretability analysis of accident conditions, rather than verifying candidate strategies, and thus lacks a closed-loop optimization mechanism.

[0004] However, existing technologies have not yet achieved deep collaboration between large models and simulators, and existing nuclear power operation optimization methods suffer from problems such as insufficient intelligence, high verification costs, and lack of closed-loop iteration. Summary of the Invention

[0005] The technical problem to be solved by this disclosure is to overcome the shortcomings of existing nuclear power plant operation optimization methods, such as insufficient intelligence, high verification costs, and lack of closed-loop iteration, and to provide an optimization method, system, electronic equipment, and storage medium for nuclear power plant operation.

[0006] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0007] The first aspect of this disclosure provides an optimization method for the operation of a nuclear power plant, the optimization method comprising:

[0008] Acquire historical and real-time operational data of nuclear power plants;

[0009] The historical operational data and the real-time operational data are classified to obtain data classification results;

[0010] High-security and real-time data, as well as non-real-time or unstructured data, are obtained from the data classification results.

[0011] The high-security and real-time data are input into the AI ​​algorithm module to obtain the AI ​​algorithm prediction results, and the non-real-time or unstructured data are input into the large model module to obtain the large model prediction results.

[0012] The AI ​​algorithm prediction results are combined with the large model prediction results to obtain the fault detection results of the nuclear power plant.

[0013] The fault measurement results of the nuclear power plant are input into the nuclear power plant simulator for operation to obtain the simulation operation results;

[0014] The operation of the nuclear power plant is optimized based on the simulation results.

[0015] Preferably, the step of optimizing the operation of the nuclear power plant based on the simulation results includes:

[0016] In response to the simulation results achieving the expected results, the simulation results are taken as the actual operating results, and the nuclear power plant is optimized based on the actual operating results.

[0017] Preferably, the step of optimizing the operation of the nuclear power plant based on the simulation results further includes:

[0018] In response to the simulation results not achieving the expected results, the process returns to the step of classifying the historical and real-time running data to obtain the data classification results.

[0019] Preferably, the step of classifying the historical operating data and the real-time operating data to obtain the data classification result includes:

[0020] The historical operation data and the real-time operation data are classified according to preset data tags to obtain data classification results.

[0021] Preferably, the step of inputting the non-real-time data or unstructured data into the large model module to obtain the large model prediction result includes:

[0022] Based on the large model module, semantic understanding and deep feature extraction are performed on the non-real-time or unstructured data to obtain the large model prediction results of potential decision-related factors in operational behavior, text description, and image information.

[0023] And / or,

[0024] The step of balancing and fusing the AI ​​algorithm prediction results with the large model prediction results to obtain the fault detection results of the nuclear power plant includes:

[0025] The AI ​​algorithm prediction results and the large model prediction results are weighed and fused based on confidence factors and risk assessment indicators to obtain the fault prediction results of the nuclear power plant.

[0026] Preferably, the preset data tags include at least one of the following: security level dimension, real-time dimension, and data format dimension;

[0027] The security level dimension includes at least one of high level, medium level, and low level;

[0028] The real-time dimension includes at least one of strong real-time, weak real-time, and non-real-time.

[0029] The data format dimensions include structured and unstructured.

[0030] A second aspect of this disclosure provides an optimization system for the operation of a nuclear power plant, the optimization system comprising:

[0031] The first acquisition module is used to acquire historical and real-time operating data of the nuclear power plant.

[0032] The classification module is used to classify the historical operation data and the real-time operation data to obtain data classification results;

[0033] The second acquisition module is used to acquire high-security-level and highly real-time data, as well as non-real-time data or unstructured data from the data classification results.

[0034] The third acquisition module is used to input the high-security and high-real-time data into the AI ​​algorithm module to obtain the AI ​​algorithm prediction results, and to input the non-real-time data or unstructured data into the large model module to obtain the large model prediction results.

[0035] The processing module is used to weigh and fuse the prediction results of the AI ​​algorithm with the prediction results of the large model to obtain the fault detection results of the nuclear power plant.

[0036] The operation module is used to input the fault measurement results of the nuclear power plant into the nuclear power plant simulator for operation and to obtain the simulation operation results;

[0037] An optimization module is used to optimize the operation of the nuclear power plant based on the simulation results.

[0038] Preferably, the optimization module is configured to, in response to the simulation results achieving the expected effect, use the simulation results as the actual operating results, and optimize the nuclear power plant based on the actual operating results.

[0039] Preferably, the optimization module is further configured to invoke the classification module in response to the simulation results not achieving the expected results.

[0040] Preferably, the classification module is used to classify the historical operating data and the real-time operating data according to preset data tags to obtain data classification results.

[0041] Preferably, the third acquisition module is used to perform semantic understanding and deep feature extraction on the non-real-time data or unstructured data based on the large model module, and obtain the large model prediction results of potential decision-related factors in operation behavior, text description, and image information;

[0042] And / or,

[0043] The processing module is used to weigh and fuse the prediction results of the AI ​​algorithm and the prediction results of the large model based on the confidence factor and risk assessment index to obtain the fault prediction results of the nuclear power plant.

[0044] Preferably, the preset data tags include at least one of the following: security level dimension, real-time dimension, and data format dimension;

[0045] The security level dimension includes at least one of high level, medium level, and low level;

[0046] The real-time dimension includes at least one of strong real-time, weak real-time, and non-real-time.

[0047] The data format dimensions include structured and unstructured.

[0048] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and for running on the processor, wherein the processor executes the computer program to implement the optimization method for nuclear power plant operation described in the first aspect.

[0049] The fourth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the optimization method for nuclear power plant operation described in the first aspect.

[0050] The fifth aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the optimization method for nuclear power plant operation as described in the first aspect.

[0051] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of this disclosure.

[0052] The positive and progressive effects of this disclosure are as follows:

[0053] This disclosure inputs high-security and real-time data after data classification into an AI algorithm module, while inputting non-real-time or unstructured data into a large model module. By balancing and fusing the AI ​​algorithm prediction results with the large model prediction results, the fault prediction results of the nuclear power plant are obtained. Then, combined with the nuclear power plant simulator, the simulated operation results are obtained. Based on the simulated operation results, the operation of the nuclear power plant is optimized, which improves the intelligent decision-making capability, reduces the high verification cost, and constructs a closed-loop iterative optimization system. Attached Figure Description

[0054] Figure 1 A flowchart of an optimization method for nuclear power plant operation provided in Embodiment 1 of this disclosure.

[0055] Figure 2 This is a schematic diagram of the modules of the optimized system for nuclear power plant operation provided in Embodiment 2 of this disclosure.

[0056] Figure 3 This is a schematic diagram of the electronic device used in the optimized method for operating a nuclear power plant according to Embodiment 3 of this disclosure. Detailed Implementation

[0057] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0058] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the claims or the context of the embodiments, and should not constitute an unnecessary limitation due to the use of such prefixes. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0059] In this embodiment of the disclosure, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information comply with relevant laws and regulations and do not violate public order and good morals.

[0060] Example 1

[0061] Figure 1 A flowchart of an optimization method for nuclear power plant operation provided in Embodiment 1 of this disclosure is shown below. Figure 1 As shown, the optimization method includes:

[0062] S1. Obtain historical and real-time operating data of the nuclear power plant;

[0063] In this embodiment, the overall architecture for optimizing nuclear power plant operation includes a data acquisition module, a data strategy allocation module, a general large-scale model module, a self-developed AI algorithm module, a decision fusion and feedback module, an optimization control module, and an intelligent decision advisor module. Specifically, the data acquisition module is used to acquire historical operating data of the nuclear power plant and real-time operating data generated by the nuclear power simulation system; the data strategy allocation module is used to allocate data of different dimensions to corresponding processing modules according to a preset data labeling system; the general large-scale model module is used to process unreal-time unstructured data and generate auxiliary decision suggestions; the self-developed AI algorithm module is used to process high-priority or high-real-time data to ensure the controllability and response timeliness of the core control logic; the decision fusion and feedback module is used to comprehensively analyze the output results of the two types of algorithms and generate the final decision; the optimization control module applies the decision to the nuclear power plant simulator, simulation platform, or actual system and provides feedback on the execution effect to form a closed-loop optimization; the intelligent decision advisor module is used to backtrack the control effect and perform model tuning to achieve system self-learning and closed-loop optimization.

[0064] In this embodiment, the data acquisition modules are divided into two main categories: real-time data acquisition modules and other data acquisition modules. The real-time data acquisition modules primarily interface with the nuclear power plant's DCS system, simulator / simulation platform, PLC controller, etc., to collect real-time operational data such as pressure, water temperature, current, voltage, and control rod positions. The sampling frequency can reach over 100Hz, ensuring the system's high sensitivity to changes in key status parameters. The other data acquisition modules acquire historical operational data, including but not limited to historical operating condition logs, operator operation records, inspection videos, maintenance logs, environmental meteorological parameters, and video images, through API interfaces or batch import methods. This provides data support for subsequent trend recognition and semantic reasoning.

[0065] The real-time data acquisition module employs a multi-threaded distributed acquisition mechanism, combined with message queue tools such as Kafka or RabbitMQ, to buffer high-frequency data (e.g., real-time running data) into an in-memory database (such as Redis or InfluxDB), achieving millisecond-level data stream processing. Non-real-time data is stored in an archived, structured manner using a data warehouse (such as Hive or ClickHouse), facilitating unified scheduling and retrieval later.

[0066] S2. Classify historical and real-time operational data to obtain data classification results;

[0067] S3. Obtain high-security and real-time data, as well as non-real-time or unstructured data from the data classification results;

[0068] S4. Input high-security and real-time data into the AI ​​algorithm module to obtain AI algorithm prediction results, and input non-real-time or unstructured data into the large model module to obtain large model prediction results.

[0069] In this embodiment, the general large model module mainly undertakes the task of intelligent understanding and trend prediction of non-real-time or unstructured data. The system introduces a finely tuned general large model (such as GPT-4, DeepSeek, etc.) here, and performs Prompt Tuning and LoRA minor adjustments in combination with nuclear power professional corpus to adapt to the professional field.

[0070] Preprocessing layer: Entity extraction is performed on text content using NER tools, such as "pressure vessel temperature", "valve opening angle", and "control rod insertion depth". For image data, OCR and image recognition models are used to extract numbers / symbols and construct a standard input format. For time-series data, event vectorization is performed to construct an event flow graph for deep reasoning.

[0071] Model layer: Based on the Transformer structure, it performs multi-task training and has the capabilities of semantic reasoning, procedure verification, and trend prediction. It uses contextual information to generate diagnostic conclusions and probability assessments of the current working condition, such as "main pump abnormality probability 87%".

[0072] Output layer: The output includes highly readable diagnostic text, decision suggestion forms, risk level scores, etc., and is configured with reliability evaluation values ​​for adoption and verification by the decision fusion module.

[0073] The self-developed AI algorithm module is the core of the overall architecture, primarily handling data with high real-time and high security requirements. Its technical implementation consists of three parts:

[0074] 1. Physics Model-Driven Part:

[0075] Construct dynamic equations and thermal-hydraulic models for point reactors, and combine them with the Runge-Kutta numerical solution method to simulate core physical quantities such as core thermal power and coolant flow rate;

[0076] The system uses a state-space model to predict system response and supports online steady-state / transient simulation.

[0077] 2. Data-driven optimization section:

[0078] For nonlinear aging effects and complex boundary condition problems that are difficult to model, LSTM recurrent neural networks and graph neural networks are used for modeling.

[0079] Input features include historical fault data, operation records, environmental disturbance parameters, etc., to improve the robustness of abnormal operating condition prediction.

[0080] 3. Real-time performance guarantee:

[0081] Optimize model inference time by introducing model quantization techniques (such as FP32→FP16), GPU parallel computing, and ASIC hardware acceleration.

[0082] The measured delay is controlled within 5ms, which meets the real-time response requirements of the online control system.

[0083] S5. The AI ​​algorithm prediction results are weighed and fused with the large model prediction results to obtain the fault detection results of the nuclear power plant.

[0084] In this embodiment, the self-developed AI algorithm module employs technologies including but not limited to differential equation modeling, Kalman filtering, customized LSTM, and graph neural networks to improve prediction accuracy and control response in complex nonlinear scenarios.

[0085] The general large model module incorporates Prompt Tuning and LoRA fine-tuning technologies, enabling it to possess specialized capabilities such as anomaly identification and procedure verification in nuclear power scenarios.

[0086] In this embodiment, a label priority matrix is ​​used to automatically allocate data paths, ensuring that different types of data are diverted to the optimal algorithm channel. Before the fusion output, a confidence factor and risk assessment index are introduced for the results of each module, forming an interpretable composite intelligent judgment.

[0087] S6. Input the fault measurement results of the nuclear power plant into the nuclear power plant simulator and run it to obtain the simulation operation results;

[0088] S7. Optimize the operation of the nuclear power plant based on the simulation results.

[0089] In this embodiment, the optimized control module converts the fused control strategy into control outputs, which directly act on the nuclear power plant simulator, simulation platform, or actual control system. Its controlled objects include: thermal power regulation, control rod position correction, coolant flow rate control, and alarm level triggering.

[0090] The intelligent decision advisor module simultaneously records the deviation between the actual execution effect and the expected goal, generates optimization logs and performance evaluation reports, and feeds them back to the AI ​​model training set to achieve self-learning and version iteration.

[0091] For example, when an "abnormal temperature rise in the main pump" occurs during the operation of a nuclear power unit, the real-time data acquisition module collects the temperature change trend and labels it as high safety level and strong real-time performance. The data is then diverted to the self-developed AI algorithm module, which predicts potential damage risks based on historical trends. At the same time, the large model performs semantic analysis on the maintenance logs to uncover corresponding solutions for similar faults in the past. The fusion module combines the two outputs to generate "The main pump can continue to operate for 40 minutes. It is recommended to perform pre-maintenance preparations according to procedure C5." The control module adjusts the operating parameters based on the judgment, controls the power to decrease, and slowly starts the auxiliary pump. The feedback module collects the actual results and updates the database.

[0092] This embodiment realizes a complete closed-loop process from data acquisition to intelligent decision-making, and from model reasoning to control optimization, and has high security, real-time performance and intelligent adaptive capabilities.

[0093] This embodiment leverages the multimodal data processing and deep reasoning capabilities of a general-purpose large model module to integrate structured data (such as sensor parameters) and unstructured data (such as maintenance logs and accident reports) from nuclear power plant operation, overcoming the limitations of traditional rule-based systems. By generating dynamic optimization strategies through reinforcement learning and evolutionary algorithms, it achieves autonomous analysis and intelligent decision-making for complex operating conditions (such as equipment aging and extreme weather), reducing reliance on human experience and enhancing intelligent decision-making capabilities. Furthermore, by introducing the high-fidelity simulation capabilities of a full-range nuclear power simulator, the candidate strategies generated by the large model module are virtually verified. Through simulator simulations of normal operating conditions, accident scenarios, and emergency operations, the safety and feasibility of the strategies are evaluated, avoiding the high risks and costs of actual unit testing. For example, when adjusting reactor control parameters or formulating maintenance plans, the simulator can predict equipment lifespan loss and power fluctuations in advance, ensuring that the strategies comply with nuclear power safety regulations and reducing verification risks and costs. In addition, a closed-loop mechanism is established for the entire process: "data acquisition - model optimization - simulation verification - strategy implementation - feedback iteration." By leveraging real-time feedback from actual operational data, the training data and algorithm parameters of the large model module are dynamically adjusted, enabling the system to adapt to changes in equipment status and load fluctuations, and continuously improve optimization performance. For example, based on simulator verification results, the control rod adjustment strategy is optimized to achieve a dynamic balance between power generation efficiency and safety, thus constructing a closed-loop iterative optimization system.

[0094] In an optional embodiment, S7 includes:

[0095] In response to the simulation results achieving the expected results, the simulation results are used as the actual operating results, and the nuclear power plant is optimized based on the actual operating results.

[0096] In an optional embodiment, S7 further includes:

[0097] If the simulation results do not meet expectations, return to step S2.

[0098] In an optional embodiment, S2 includes:

[0099] Based on preset data labels, historical and real-time operational data are classified to obtain data classification results.

[0100] In an optional embodiment, the preset data tags include at least one of the following: security level dimension, real-time dimension, and data format dimension;

[0101] The security level dimension includes at least one of high level, medium level, and low level;

[0102] The real-time dimension includes at least one of strong real-time, weak real-time, and non-real-time.

[0103] Data format dimensions include structured and unstructured.

[0104] In this embodiment, to address the diverse challenges of different data dimensions, timeliness, and processing sensitivity, a data tagging system (e.g., preset data tags) is provided in the data preprocessing layer. This data tagging system specifically includes the following dimensions:

[0105] 1. Security Level: Used to distinguish between core control data and auxiliary analysis data; the security level classifies data into three levels: high, medium, and low, according to the degree of impact on system security.

[0106] High-level parameters, such as reactor power control parameters, core temperature, and shutdown trigger signals, must be processed by a self-developed AI algorithm module with strong controllability.

[0107] Medium level: such as equipment life assessment, material change cycle recommendations, energy consumption trends;

[0108] Low-level: such as operation and maintenance logs, personnel scheduling records, etc.

[0109] 2. Real-time performance level: Used to indicate the frequency of data updates and processing time requirements;

[0110] Strong real-time: Data update frequency ≥100Hz, such as sensor streams, control signals, etc., with delay control requirement <10ms;

[0111] Weak real-time: such as parameter refresh cycle on the order of 1 second, suitable for periodically sampled data;

[0112] Non-real-time: such as offline data like historical logs, documents, and images.

[0113] 3. Data format: Used to distinguish between structured and unstructured data and to determine the corresponding processing method.

[0114] Structured data: such as numerical sensor data, CSV format tables;

[0115] Unstructured data: such as images, videos, PDF files, and text operation and maintenance records.

[0116] Through the aforementioned data tagging system, the data strategy allocation module can automatically distribute different data to the self-developed AI algorithm module or the general large model module based on the priority decision matrix, thereby improving resource utilization efficiency and processing matching degree.

[0117] In an optional embodiment, S4 includes:

[0118] Based on the large model module, semantic understanding and deep feature extraction are performed on non-real-time or unstructured data to obtain large model prediction results of potential decision-related factors in operational behavior, text description, and image information;

[0119] In an optional embodiment, S5 includes:

[0120] The AI ​​algorithm prediction results and the large model prediction results are weighed and fused based on confidence factors and risk assessment indicators to obtain the fault prediction results of nuclear power plants.

[0121] In this embodiment, semantic understanding and deep feature extraction are performed on non-real-time or unstructured data based on the general large model module to obtain potential decision-related factors in operation behavior, text description, and image information.

[0122] Based on a self-developed AI algorithm module, physical modeling and state prediction are performed on high-safety-level and real-time data to ensure high-precision estimation of key parameters such as reactor power and coolant temperature.

[0123] The outputs of the general large model module and the self-developed AI algorithm module are then transformed into decision probability vectors through a fusion module. A confidence evaluation mechanism is used for fusion judgment, thereby realizing collaborative decision output.

[0124] Specifically, to avoid the uncertainty of a single algorithm affecting the core control, this embodiment introduces an inter-model collaboration mechanism through a decision fusion module:

[0125] 1. Introduce a credibility assessment mechanism to perform semantic matching and confidence scoring on the output results of large model modules;

[0126] 2. Use the output of the self-developed AI algorithm module as the main reference and compare its logical consistency with the candidate suggestions of the large model module;

[0127] 3. If a candidate suggestion receives high confidence and passes the expert system rules verification, it will be included in the fusion output;

[0128] 4. All fusion decision results enter the feedback module, where they are compared and corrected based on the post-execution status, forming a data-driven self-learning optimization mechanism.

[0129] This embodiment aims to integrate actual nuclear power plant operation data with virtual simulation models to construct a comprehensive system encompassing data acquisition, strategy allocation, intelligent decision-making, and optimized control. The system acquires multi-source heterogeneous data in real time through a data acquisition module, including nuclear power simulator output signals, historical operation records, maintenance logs, and inspection videos. The data strategy allocation module then categorizes and processes this data according to safety level, real-time performance, and data format. High-safety and highly real-time data are precisely processed by a self-developed AI algorithm module, while non-real-time and complex data are analyzed using a general large-scale model to extract pattern features. The system integrates physical mechanism models, machine learning algorithms, and multimodal large-scale models to construct high-confidence intelligent decision-making schemes. Through an optimized control module, it achieves closed-loop regulation of the nuclear power plant's operating status, offering advantages such as fast response speed, high processing accuracy, and adaptability to complex operating conditions. It is suitable for various scenarios, including nuclear power plant operation optimization, fault prediction, and energy efficiency improvement, significantly enhancing the intelligence level and operational safety of the nuclear power system. Furthermore, leveraging the multimodal learning and knowledge graph technology of the large-scale model module, potential risks (such as pipe cracks and sensor drift) are identified, and emergency response plans are simulated in the simulator. The standardized verification process using simulators ensures the reliability of the strategy while reducing delays in responding to new types of accidents. Furthermore, the closed-loop mechanism enables the system to autonomously evolve as operational data accumulates, adapting to performance changes in nuclear power equipment over long-term operation. Through deep collaboration between domestically developed large-scale models (such as DeepSeek) and independently controllable simulators, a safe and efficient intelligent decision-making platform for nuclear power is constructed. This platform reduces reliance on foreign technologies, enhances the digitalization level of the nuclear power industry, provides technical support for intelligent management of nuclear power companies, and helps achieve the goal of "zero accidents" and improved economic efficiency.

[0130] In the specific implementation process, by combining the actual operation of the nuclear power plant with a virtual simulation model, and utilizing advanced simulation technology, artificial intelligence, big data analysis, and other means, intelligent management and optimized control of the nuclear power plant are achieved. Specifically, the overall architecture for optimizing nuclear power plant operation includes a data acquisition module, a data allocation strategy module, a general large model module, a self-developed AI algorithm module, a decision fusion and feedback module, an optimized control module, and an intelligent decision advisor module.

[0131] The data acquisition module includes other data acquisition modules and a real-time data acquisition module. Both modules are connected to the data strategy allocation module. As a core foundational component of the entire system, the data acquisition module is responsible for efficiently aggregating multi-dimensional, multi-source, and multi-granular data from the nuclear power plant's operation, providing reliable data support for subsequent model training, intelligent decision-making, simulation verification, and optimized control. From a data source perspective, this module primarily consists of two sub-modules: a real-time data acquisition module and other data acquisition modules. These modules connect to the data strategy allocation module via a unified data interface protocol, forming a flexible, scalable, and highly reliable data flow system.

[0132] The real-time data acquisition module is primarily designed for high-speed, continuous, and structured data streams generated during the operation of nuclear power plants, including but not limited to the following key operating parameters:

[0133] 1. Physical quantity measurement data: such as reactor core outlet temperature, main steam temperature, coolant flow rate, pressure vessel pressure, auxiliary pump flow rate, etc. This data is typically collected by various high-precision industrial sensors deployed in key locations and uploaded via DCS (Distributed Control System) or SCADA system.

[0134] 2. Electrical quantity monitoring data: including main generator output voltage, current, active power, reactive power, power factor, etc. This type of data is used to determine load changes, energy balance status, and the operating efficiency of frequency regulation and voltage regulation systems.

[0135] 3. Control system feedback signals: such as control rod position signals, regulating valve opening, pump unit operating status (start / stop), interlock signals, etc. These signals not only reflect the current control command execution status of the system, but also provide key information for judging operational stability and predicting faults.

[0136] 4. Simulation platform output data: In a high-fidelity simulation system, the real-time operating parameters simulated by virtual devices (such as virtual instrument readings, simulated alarm signals, etc.) are also included in the real-time data category to ensure the synchronization and reliability of simulation with real operation.

[0137] These real-time data are characterized by high update frequency (≥100Hz), high degree of structure, and strong data integrity, making them suitable for direct input into self-developed AI algorithm modules for highly timely intelligent processing, such as rapid dynamic response, model prediction and correction, and online edge detection. During data access, to ensure low latency and high availability of data transmission, the system adopts high-performance industrial communication protocols (such as OPC UA, Modbus TCP, etc.) and supports functions such as breakpoint resumption, dual-channel redundant acquisition, and timestamp synchronization to guarantee data real-time performance and security.

[0138] Other data acquisition modules primarily focus on non-real-time, low-frequency, semi-structured, or unstructured data sources. While these data are not essential for real-time control of nuclear power plant operations, they play a crucial role in fault diagnosis, experience reuse, trend analysis, and knowledge graph construction. Specifically, they include the following categories:

[0139] 1. Historical operation records: such as historical operation data stored in the form of SQL logs, CSV data tables or PDF documents, covering various past start-up and shutdown conditions, load adjustment processes, abnormal event records, operation reports, etc., which can be used to extract feature patterns, train prediction models, and reproduce abnormal conditions.

[0140] 2. Inspection video and image data: Video streams and image snapshots from factory camera systems, wearable terminals or drone inspections reflect detailed information such as equipment surface condition, instrument pointer position, and signs of condensate leakage. They are suitable for input into the multimodal processing module of the large model module for semantic extraction and image recognition.

[0141] 3. Maintenance Log: Records the maintenance cycle, inspection process, problem description, handling results, and replacement parts for various equipment. This type of data is mostly in text form, with information scattered and expressed in various ways, but it contains a large amount of fault causal chains and maintenance experience, which is extremely valuable for intelligent pre-maintenance and risk assessment.

[0142] 4. Environmental climate and external conditions data: External factors such as temperature, humidity, wind speed, external power grid frequency fluctuations, and earthquake monitoring data, although not directly involved in reactor control, have a significant impact on the performance of auxiliary systems (such as cooling systems and ventilation systems) and are one of the important inputs for building a global operational status analysis model.

[0143] To ensure unified access to different types of data sources, the system is equipped with various format conversion and protocol adaptation tools (such as document parsers, OCR engines, and video-to-image modules) in other data acquisition modules, achieving unified vectorized encoding between text, images, videos, and structured tables. Simultaneously, it supports data acquisition through multiple methods, including REST API, batch import via FTP, and direct database connection, ensuring the comprehensiveness and flexibility of historical data access.

[0144] Whether it's real-time data or other types of data, all data is aggregated through standardized data interfaces to the data strategy allocation module for further tagging, prioritization, and downstream task scheduling. This module constructs a unified data standard system and a multi-dimensional tagging system (such as security level, data format, and real-time tags) to perform semantic understanding and usage guidance on the data, thereby achieving precise scheduling and collaborative utilization of data between the large model and the self-developed AI algorithm module.

[0145] For example:

[0146] The time-series data of a certain pressure sensor will be automatically marked as "structured / real-time / high priority / safety related" and will be preferentially input into the self-developed AI algorithm module for short-cycle dynamic prediction.

[0147] A maintenance record PDF document will be marked as "unstructured / non-real-time / secondary priority / maintenance assistance" for use in generating maintenance suggestions and knowledge graph supplements for large model modules;

[0148] A video inspection segment is labeled as "unstructured / image / medium priority / fault identification," triggering the multimodal recognition module to screen for apparent anomalies.

[0149] Through this organic integration and intelligent allocation, the data acquisition module not only completes the basic function of "collecting data", but also becomes the central hub of "intelligent data flow" in the entire closed-loop system, providing strong data support for model inference, control command generation, and operation strategy optimization.

[0150] The data strategy allocation module establishes a multi-dimensional data tagging system, which classifies data sequentially according to preset priority strategies based on security level, real-time performance, and data format.

[0151] In this embodiment, the data strategy allocation module undertakes the crucial data central scheduling function, serving as a bridge connecting the data acquisition module and downstream intelligent processing modules (including the self-developed AI algorithm module, large model module, decision feedback module, etc.). This data strategy allocation module not only completes data classification, preprocessing, and tag management, but also formulates differentiated data processing paths based on multiple dimensions such as data security level, real-time requirements, and degree of structure, ensuring that various types of data flow and are used within the system in the most reasonable, efficient, and secure manner.

[0152] To achieve precise management and intelligent distribution of massive amounts of heterogeneous data, the data strategy allocation module has constructed a systematic multi-dimensional data tagging system. This data tagging system revolves around three core dimensions:

[0153] 1. Data Security Level: Divided into high security level (such as nuclear safety related signals), medium security level (such as equipment status parameters), and low security level (such as environmental monitoring data).

[0154] 2. Real-time dimension (Data Timeliness): Divided into strong real-time data (update frequency ≥ 100Hz), weak real-time data (1Hz~100Hz), and non-real-time data (updated on an hourly or daily basis).

[0155] 3. Data Format Dimension: Divided into structured data (such as time series tables, SQL databases), semi-structured data (such as XML / JSON logs), and unstructured data (such as natural language text, images, videos, and audio).

[0156] Based on the data labels across these three dimensions, the system has established a pre-defined priority strategy matrix to automatically classify, label, and allocate processing paths for each type of data. The following sections provide a detailed explanation of each strategy type:

[0157] The safety-first strategy refers to the fact that in the complex operating environment of nuclear power plants, some data is related to equipment safety, personnel safety, and even the nuclear safety boundary, requiring extremely high accuracy and controllability in its processing. Therefore, the system adopts a mandatory classification and control strategy for the safety level dimension:

[0158] High-safety-level data includes critical safety parameters such as reactor power limit thresholds, coolant pressure alarm signals, emergency shutdown trigger signals, and spent fuel pool water level alarms. This type of data must be processed by a self-developed AI algorithm module that has undergone safety certification. Its modeling logic is based on nuclear power plant operation mechanisms and control logic, and it can provide interpretable and traceable judgments. It strictly avoids the use of probabilistic language generation methods based on large models, thus mitigating the potential uncertainty risks associated with it.

[0159] For low-to-medium safety level data, such as equipment health index (HI), cooling water energy consumption, temperature difference efficiency coefficient, and maintenance cycle, candidate suggested solutions can first be generated by a general large model (such as a finely tuned DeepSeek). Then, the rationality and feasibility of these solutions are verified by a self-developed AI algorithm module. Finally, the data is handed over to the "decision fusion and feedback module" for unified scheduling and feedback closed-loop control. This data processing strategy improves reasoning efficiency and knowledge generalization ability while ensuring safety.

[0160] Real-time response strategy refers to the fact that nuclear power plant control systems are extremely sensitive to time response, especially in scenarios involving dynamic response, closed-loop control, and accident prediction, where data processing latency must be strictly controlled at the millisecond level. Therefore, the system has developed the following classification and processing strategies based on the real-time characteristics of the data:

[0161] High-latency real-time data (≥100Hz): Such as time-series signals like temperature, pressure, flow rate, control rod position signals, and main pump speed of core components. These signals have a high update frequency and strong dynamics, requiring low-latency processing capabilities of ≤10ms. This type of data is processed preferentially by a self-developed AI algorithm module. The kernel of this algorithm often includes differential equation solvers, dynamic models of control systems, Kalman filters, and multivariable state estimators, possessing high robustness and stability, and effectively supporting dynamic simulation and real-time control decision-making.

[0162] Non-real-time data, such as historical operating logs, equipment maintenance records, and operation summary reports, typically has an update cycle of minutes to hours. This type of data is suitable for inputting into general-purpose models for in-depth pattern mining and knowledge extraction. This type of data analysis focuses more on long-term trend prediction, potential fault location, and operational experience accumulation, and can support the construction of visualization analysis tools such as "operational error risk maps" and "critical equipment failure evolution paths."

[0163] Data format processing strategy refers to the different processing methods determined by different data formats. Based on the degree of data structure, the system defines the following path:

[0164] Structured data processing path: This includes time-series data collected by industrial instruments, operational reports exported from DCS systems, database tables, etc. This type of data is processed first by the self-developed AI algorithm module. Before being input into the model, data cleaning, normalization, and physical parameter mapping are performed. The algorithm combines nuclear reaction kinetic equations, energy conservation equations, and mass conservation equations for modeling and analysis, which is suitable for quantitative simulation, parameter optimization, and boundary calculation.

[0165] Unstructured data processing paths include: operation logs (natural language), maintenance case documents, accident investigation reports, graphical instrument readings, and monitoring videos. These data vary in format and have complex semantics, making them unsuitable for traditional algorithm processing. The system utilizes the NLP and multimodal parsing capabilities of a general-purpose large model to perform vectorized semantic encoding, keyword extraction, and image / text OCR fusion processing to extract implicit information such as "error operation mode," "fault correlation causal chain," and "abnormal equipment status in images."

[0166] Semi-structured data processing path: such as system operation configuration in XML format, alarm information records in JSON format, etc., although they contain a certain structure, there is free text nesting. The system first converts them into a structured form and then assigns them to the self-developed AI algorithm module or large model module channel according to their security level and real-time label.

[0167] To ensure that various types of data can quickly match the optimal processing path under label cross-referencing, the system introduces a label fusion and path optimization mechanism. After receiving a set of data, this mechanism automatically invokes the label inference engine, comprehensively considers its three-dimensional label weights, and determines the optimal processing method through decision tree or Bayesian path selection algorithms. For example:

[0168] If a set of data with an update frequency of 200Hz, a data type of pressure signal, and a high safety level has a tag fusion result of: high priority / strong real-time / structured, then the system will automatically schedule it to enter the "self-developed AI algorithm module fast processing channel".

[0169] If an operator error log text is tagged as medium priority / non-real-time / unstructured, it will be automatically routed to the "general large model semantic parsing module" and labeled as "supports decision suggestion input".

[0170] This fusion mechanism effectively avoids the fragmentation problem of data processing strategies and improves the intelligence and automation level of data scheduling. Through the implementation of the above data strategy allocation mechanism, this system maximizes data utilization efficiency while ensuring nuclear security and real-time response. It allows the self-developed AI algorithm module and the general large model module to each play to their strengths, building a collaborative data intelligence system of "precise modeling + generalized reasoning + intelligent fusion," which becomes an important engine supporting the continuous evolution of the closed-loop optimization system.

[0171] The general-purpose large model module comprises a preprocessing layer, a model layer, and an output layer. The preprocessing layer performs entity recognition on unstructured data (e.g., extracting key entities such as "pressure vessel temperature" and "control rod insertion depth") and vectorizes time-series data (e.g., converting sensor sequences into event encoding matrices). The model layer employs domain adaptation techniques (e.g., Prompt Tuning, LoRA) to fine-tune the general-purpose large model module for nuclear power scenarios, focusing on optimizing tasks such as abnormal operating condition diagnosis and compliance verification of operating procedures. The output layer generates probability distribution-based results (e.g., "the probability of steam generator leakage under current operating conditions is 87%)" and includes a confidence score for subsequent coupled decision-making.

[0172] The self-developed AI algorithm module comprises three parts: physics model-driven, data-driven optimization, and real-time performance assurance. The physics model-driven approach is based on nuclear engineering mechanistic models (such as point reactor dynamics equations and thermal-hydraulic models), employing numerical integration methods (such as Runge-Kutta) for deterministic calculations to output precise physical quantities (such as core power distribution and coolant temperature gradient). Data-driven optimization addresses complex nonlinear aspects difficult to model using mechanistic models (such as material aging effects during long-term operation), supplementing modeling with self-developed machine learning models (such as customized LSTM and graph neural networks) to improve the accuracy of edge case predictions. Real-time performance assurance ensures that the self-developed AI algorithm module meets the ≤5ms cycle time requirement in embedded real-time systems through model quantization (FP32→FP16) and hardware acceleration (GPU / ASIC deployment).

[0173] The decision fusion and feedback module primarily implements the fusion of multi-source model outputs, confidence-weighted decision-making, and a closed-loop feedback update mechanism to build a robust and reliable intelligent decision engine. This module consists of a candidate solution aggregation submodule, a confidence mechanism evaluation submodule, a fusion strategy execution submodule, and an operational feedback learning submodule.

[0174] The candidate solution aggregation submodule receives various decision suggestions from the general large model module and the self-developed AI algorithm module, including predictive indicators (such as equipment failure probability and operational deviation risk), optimized control variables (such as control rod adjustment schemes and coolant flow change schemes), and alarm suggestions (such as suggestions for early inspection or equipment status warnings). All candidate solutions are tagged and managed according to attributes such as data source, model type, and computation latency to facilitate subsequent integration and traceability.

[0175] Confidence Mechanism Evaluation Submodule: Evaluates the confidence level of each candidate solution based on the following dimensions:

[0176] 1. Model training coverage: Determines whether the current working condition falls within the training data distribution; if it exceeds the range, the confidence level is automatically lowered.

[0177] 2. Model applicability boundaries: For example, large models are good at handling pattern recognition tasks, but their confidence in high-precision numerical prediction is discounted;

[0178] 3. Real-time requirement satisfaction: If the solution generation delay exceeds the preset threshold (e.g., 10ms), the result will be used for long-term optimization and excluded from emergency response tasks;

[0179] 4. Multi-model consistency test: If the general large model module and the self-developed AI algorithm module make the same judgment on a certain event, the overall confidence level is increased; otherwise, it enters the manual review queue.

[0180] The fusion strategy execution submodule employs a multi-model weighted voting mechanism, Bayesian inference fusion method, or confidence-based dynamic strategy switching mechanism to ultimately generate an executable control strategy and early warning suggestions. In special circumstances, manual intervention or interruption of model decision-making by the upper-level control system is permitted to ensure operational safety.

[0181] The runtime feedback learning submodule: This module transmits the operational results (such as response latency, actual failure rate, and load fluctuations) back in real time for dynamic updates.

[0182] The optimized control module and the intelligent decision advisor module constitute the execution closed loop of intelligent control, which directly affects the adjustment of key operating parameters of the nuclear power plant, the action commands of control equipment, and the operation guidance of the operation and maintenance team.

[0183] The optimized control module can output control commands to the nuclear power simulation system or the actual control interface, and obtain execution status feedback in real time. By rapidly analyzing the fused optimal control strategy and generating commands, the optimized control module drives the full-range nuclear power simulator system to make precise adjustments. Its core components include:

[0184] 1. Control Strategy Parsing Unit: Transforms abstract layer decision results (such as "suggest reducing main pump power by 10%) into specific control commands (such as "adjust the main pump speed setpoint from 1450 rpm to 1305 rpm"), supporting multi-time scale adjustment (ms-level fast response / min-level rolling optimization).

[0185] 2. Control command issuance and synchronization unit: It adopts a distributed industrial protocol based on OPC-UA or DDS (Data Distribution Service) to send control commands to the simulation platform or DCS system, and at the same time realizes multi-channel redundant broadcasting and execution feedback acquisition.

[0186] 3. Safety threshold verification mechanism: Before all instructions are issued, they are logically checked by the constraint verifier of the self-developed AI algorithm module to ensure that no safety limits are triggered (such as "coolant flow rate is not lower than the minimum critical value" and "control rod position change rate must not exceed physical limit"), thus ensuring the stable operation of the system.

[0187] The intelligent decision advisor module, serving as the system's "human-machine interface layer," primarily provides model-based decision support and operational recommendations for operations engineers and scheduling managers. Its design features include:

[0188] 1. Multimodal interactive terminal: Supports multiple interaction methods including text, image, and voice. Dispatchers can input current operating conditions via natural language (e.g., "Why is the main steam temperature too high?"), and the system automatically retrieves cause analysis reports from the simulation model and historical database and generates visual charts.

[0189] 2. Operation Recommendation Generator: Based on the current operating status and model inference results, it automatically generates operation suggestions (e.g., "It is recommended to switch ventilation fan A to standby mode to reduce energy consumption", and attaches comparative data of similar historical operating conditions and expected energy-saving effects).

[0190] 3. Knowledge Graph and Case Library Integration: The system integrates a nuclear power plant operation knowledge graph and a classic operating condition case library to perform semantic matching on the current abnormal state, push similar cases and handling experience, and help young operators respond quickly to complex situations.

[0191] 4. Feedback and incentive mechanism: Operators are encouraged to evaluate the model's recommended operations (e.g., mark them as "effective" or "ineffective"). The evaluation results will be used to reinforce the labels of the training data, providing real-world feedback for the model's continuous learning.

[0192] This embodiment inputs high-safety-level and highly real-time data, after data classification, into the AI ​​algorithm module, while inputting non-real-time or unstructured data into the large model module. By balancing and fusing the AI ​​algorithm predictions with the large model predictions, fault prediction results for the nuclear power plant are obtained. These results are then combined with a nuclear power plant simulator to obtain simulated operation results. Based on these simulated operation results, the operation of the nuclear power plant is optimized, improving intelligent decision-making capabilities, reducing high verification costs, and constructing a closed-loop iterative optimization system. Specifically, by combining core dimensions such as safety level, real-time performance, and data format, refined classification and intelligent strategy allocation of various types of data in the nuclear power system are achieved. High-safety-level data is processed by a controllable, self-developed AI algorithm module, ensuring operational safety and interpretable results; highly real-time data is processed by differential models and filtering algorithms, achieving millisecond-level response to meet the rapid dynamic requirements of the control system; non-real-time and unstructured data undergo semantic parsing and pattern mining through the large model, improving knowledge extraction and decision support capabilities. Through the tag fusion mechanism, the optimal processing path can be automatically matched for different data, realizing the high efficiency and intelligence of data flow. This effectively improves the data utilization efficiency, safety assurance level and intelligent analysis capability of the nuclear power full-range simulator, and provides a solid data foundation and processing support for subsequent intelligent operation and maintenance, auxiliary decision-making and fault prediction.

[0193] Example 2

[0194] Corresponding to the aforementioned embodiment of an optimization method for nuclear power plant operation, this disclosure also provides an embodiment of an optimization system for nuclear power plant operation.

[0195] Figure 2 This is a schematic diagram of a module of an optimization system for nuclear power plant operation provided in Embodiment 2 of this disclosure, as shown below. Figure 2 As shown, the optimization system includes:

[0196] The first acquisition module 21 is used to acquire historical and real-time operating data of the nuclear power plant.

[0197] In this embodiment, the overall architecture for optimizing nuclear power plant operation includes a data acquisition module, a data strategy allocation module, a general large-scale model module, a self-developed AI algorithm module, a decision fusion and feedback module, an optimization control module, and an intelligent decision advisor module. Specifically, the data acquisition module is used to acquire historical operating data of the nuclear power plant and real-time operating data generated by the nuclear power simulation system; the data strategy allocation module is used to allocate data of different dimensions to corresponding processing modules according to a preset data labeling system; the general large-scale model module is used to process unreal-time unstructured data and generate auxiliary decision suggestions; the self-developed AI algorithm module is used to process high-priority or high-real-time data to ensure the controllability and response timeliness of the core control logic; the decision fusion and feedback module is used to comprehensively analyze the output results of the two types of algorithms and generate the final decision; the optimization control module applies the decision to the nuclear power plant simulator, simulation platform, or actual system and provides feedback on the execution effect to form a closed-loop optimization; the intelligent decision advisor module is used to backtrack the control effect and perform model tuning to achieve system self-learning and closed-loop optimization.

[0198] In this embodiment, the data acquisition modules are divided into two main categories: real-time data acquisition modules and other data acquisition modules. The real-time data acquisition modules primarily interface with the nuclear power plant's DCS system, simulator / simulation platform, PLC controller, etc., to collect real-time operational data such as pressure, water temperature, current, voltage, and control rod positions. The sampling frequency can reach over 100Hz, ensuring the system's high sensitivity to changes in key status parameters. The other data acquisition modules acquire historical operational data, including but not limited to historical operating condition logs, operator operation records, inspection videos, maintenance logs, environmental meteorological parameters, and video images, through API interfaces or batch import methods. This provides data support for subsequent trend recognition and semantic reasoning.

[0199] The real-time data acquisition module employs a multi-threaded distributed acquisition mechanism, combined with message queue tools such as Kafka or RabbitMQ, to buffer high-frequency data (e.g., real-time running data) into an in-memory database (such as Redis or InfluxDB), achieving millisecond-level data stream processing. Non-real-time data is stored in an archived, structured manner using a data warehouse (such as Hive or ClickHouse), facilitating unified scheduling and retrieval later.

[0200] Classification module 22 is used to classify historical and real-time operating data to obtain data classification results;

[0201] The second acquisition module 23 is used to acquire high-security-level and highly real-time data, as well as non-real-time data or unstructured data from the data classification results.

[0202] The third acquisition module 24 is used to input high-security and real-time data into the AI ​​algorithm module to obtain AI algorithm prediction results, and to input non-real-time or unstructured data into the large model module to obtain large model prediction results.

[0203] In this embodiment, the general large model module mainly undertakes the task of intelligent understanding and trend prediction of non-real-time or unstructured data. The system introduces a finely tuned general large model (such as GPT-4, DeepSeek, etc.) here, and performs Prompt Tuning and LoRA minor adjustments in combination with nuclear power professional corpus to adapt to the professional field.

[0204] Preprocessing layer: Entity extraction is performed on text content using NER tools, such as "pressure vessel temperature", "valve opening angle", and "control rod insertion depth". For image data, OCR and image recognition models are used to extract numbers / symbols and construct a standard input format. For time-series data, event vectorization is performed to construct an event flow graph for deep reasoning.

[0205] Model layer: Based on the Transformer structure, it performs multi-task training and has the capabilities of semantic reasoning, procedure verification, and trend prediction. It uses contextual information to generate diagnostic conclusions and probability assessments of the current working condition, such as "main pump abnormality probability 87%".

[0206] Output layer: The output includes highly readable diagnostic text, decision suggestion forms, risk level scores, etc., and is configured with reliability evaluation values ​​for adoption and verification by the decision fusion module.

[0207] The self-developed AI algorithm module is the core of the overall architecture, primarily handling data with high real-time and high security requirements. Its technical implementation consists of three parts:

[0208] 1. Physics Model-Driven Part:

[0209] Construct dynamic equations and thermal-hydraulic models for point reactors, and combine them with the Runge-Kutta numerical solution method to simulate core physical quantities such as core thermal power and coolant flow rate;

[0210] The system uses a state-space model to predict system response and supports online steady-state / transient simulation.

[0211] 2. Data-driven optimization section:

[0212] For nonlinear aging effects and complex boundary condition problems that are difficult to model, LSTM recurrent neural networks and graph neural networks are used for modeling.

[0213] Input features include historical fault data, operation records, environmental disturbance parameters, etc., to improve the robustness of abnormal operating condition prediction.

[0214] 3. Real-time performance guarantee:

[0215] Optimize model inference time by introducing model quantization techniques (such as FP32→FP16), GPU parallel computing, and ASIC hardware acceleration.

[0216] The measured delay is controlled within 5ms, which meets the real-time response requirements of the online control system.

[0217] Processing module 25 is used to weigh and fuse the AI ​​algorithm prediction results with the large model prediction results to obtain the fault detection results of the nuclear power plant.

[0218] In this embodiment, the self-developed AI algorithm module employs technologies including but not limited to differential equation modeling, Kalman filtering, customized LSTM, and graph neural networks to improve prediction accuracy and control response in complex nonlinear scenarios.

[0219] The general large model module incorporates Prompt Tuning and LoRA fine-tuning technologies, enabling it to possess specialized capabilities such as anomaly identification and procedure verification in nuclear power scenarios.

[0220] In this embodiment, a label priority matrix is ​​used to automatically allocate data paths, ensuring that different types of data are diverted to the optimal algorithm channel. Before the fusion output, a confidence factor and risk assessment index are introduced for the results of each module, forming an interpretable composite intelligent judgment.

[0221] The operation module 26 is used to input the fault measurement results of the nuclear power plant into the nuclear power plant simulator for operation and to obtain the simulation operation results;

[0222] Optimization module 27 is used to optimize the operation of the nuclear power plant based on the simulation results.

[0223] In this embodiment, the optimized control module converts the fused control strategy into control outputs, which directly act on the nuclear power plant simulator, simulation platform, or actual control system. Its controlled objects include: thermal power regulation, control rod position correction, coolant flow rate control, and alarm level triggering.

[0224] The intelligent decision advisor module simultaneously records the deviation between the actual execution effect and the expected goal, generates optimization logs and performance evaluation reports, and feeds them back to the AI ​​model training set to achieve self-learning and version iteration.

[0225] For example, when an "abnormal temperature rise in the main pump" occurs during the operation of a nuclear power unit, the real-time data acquisition module collects the temperature change trend and labels it as high safety level and strong real-time performance. The data is then diverted to the self-developed AI algorithm module, which predicts potential damage risks based on historical trends. At the same time, the large model performs semantic analysis on the maintenance logs to uncover corresponding solutions for similar faults in the past. The fusion module combines the two outputs to generate "The main pump can continue to operate for 40 minutes. It is recommended to perform pre-maintenance preparations according to procedure C5." The control module adjusts the operating parameters based on the judgment, controls the power to decrease, and slowly starts the auxiliary pump. The feedback module collects the actual results and updates the database.

[0226] This embodiment realizes a complete closed-loop process from data acquisition to intelligent decision-making, and from model reasoning to control optimization, and has high security, real-time performance and intelligent adaptive capabilities.

[0227] This embodiment leverages the multimodal data processing and deep reasoning capabilities of a general-purpose large model module to integrate structured data (such as sensor parameters) and unstructured data (such as maintenance logs and accident reports) from nuclear power plant operation, overcoming the limitations of traditional rule-based systems. By generating dynamic optimization strategies through reinforcement learning and evolutionary algorithms, it achieves autonomous analysis and intelligent decision-making for complex operating conditions (such as equipment aging and extreme weather), reducing reliance on human experience and enhancing intelligent decision-making capabilities. Furthermore, by introducing the high-fidelity simulation capabilities of a full-range nuclear power simulator, the candidate strategies generated by the large model module are virtually verified. Through simulator simulations of normal operating conditions, accident scenarios, and emergency operations, the safety and feasibility of the strategies are evaluated, avoiding the high risks and costs of actual unit testing. For example, when adjusting reactor control parameters or formulating maintenance plans, the simulator can predict equipment lifespan loss and power fluctuations in advance, ensuring that the strategies comply with nuclear power safety regulations and reducing verification risks and costs. In addition, a closed-loop mechanism is established for the entire process: "data acquisition - model optimization - simulation verification - strategy implementation - feedback iteration." By leveraging real-time feedback from actual operational data, the training data and algorithm parameters of the large model module are dynamically adjusted, enabling the system to adapt to changes in equipment status and load fluctuations, and continuously improve optimization performance. For example, based on simulator verification results, the control rod adjustment strategy is optimized to achieve a dynamic balance between power generation efficiency and safety, thus constructing a closed-loop iterative optimization system.

[0228] In an optional embodiment, the optimization module is configured to, in response to the simulation results achieving the expected results, use the simulation results as the actual operating results and optimize the nuclear power plant based on the actual operating results.

[0229] In an optional embodiment, the optimization module is further configured to invoke the classification module in response to the simulation results not meeting expectations.

[0230] In an optional embodiment, the classification module is used to classify historical running data and real-time running data according to preset data labels to obtain data classification results.

[0231] In an optional embodiment, the preset data tags include at least one of the following: security level dimension, real-time dimension, and data format dimension;

[0232] The security level dimension includes at least one of high level, medium level, and low level;

[0233] The real-time dimension includes at least one of strong real-time, weak real-time, and non-real-time.

[0234] Data format dimensions include structured and unstructured.

[0235] In this embodiment, to address the diverse challenges of different data dimensions, timeliness, and processing sensitivity, a data tagging system (e.g., preset data tags) is provided in the data preprocessing layer. This data tagging system specifically includes the following dimensions:

[0236] 1. Security Level: Used to distinguish between core control data and auxiliary analysis data; the security level classifies data into three levels: high, medium, and low, according to the degree of impact on system security.

[0237] High-level parameters, such as reactor power control parameters, core temperature, and shutdown trigger signals, must be processed by a self-developed AI algorithm module with strong controllability.

[0238] Medium level: such as equipment life assessment, material change cycle recommendations, energy consumption trends;

[0239] Low-level: such as operation and maintenance logs, personnel scheduling records, etc.

[0240] 2. Real-time performance level: Used to indicate the frequency of data updates and processing time requirements;

[0241] Strong real-time: Data update frequency ≥100Hz, such as sensor streams, control signals, etc., with delay control requirement <10ms;

[0242] Weak real-time: such as parameter refresh cycle on the order of 1 second, suitable for periodically sampled data;

[0243] Non-real-time: such as offline data like historical logs, documents, and images.

[0244] 3. Data format: Used to distinguish between structured and unstructured data and to determine the corresponding processing method.

[0245] Structured data: such as numerical sensor data, CSV format tables;

[0246] Unstructured data: such as images, videos, PDF files, and text operation and maintenance records.

[0247] Through the aforementioned data tagging system, the data strategy allocation module can automatically distribute different data to the self-developed AI algorithm module or the general large model module based on the priority decision matrix, thereby improving resource utilization efficiency and processing matching degree.

[0248] In an optional embodiment, the third acquisition module is used to perform semantic understanding and deep feature extraction on non-real-time data or unstructured data based on the large model module, and obtain the large model prediction results of potential decision-related factors in operation behavior, text description, and image information.

[0249] In an optional embodiment, the processing module is used to perform a weighted fusion process on the AI ​​algorithm prediction results and the large model prediction results based on confidence factors and risk assessment indicators to obtain the fault measurement results of the nuclear power plant.

[0250] In this embodiment, semantic understanding and deep feature extraction are performed on non-real-time or unstructured data based on the general large model module to obtain potential decision-related factors in operation behavior, text description, and image information.

[0251] Based on a self-developed AI algorithm module, physical modeling and state prediction are performed on high-safety-level and real-time data to ensure high-precision estimation of key parameters such as reactor power and coolant temperature.

[0252] The outputs of the general large model module and the self-developed AI algorithm module are then transformed into decision probability vectors through a fusion module. A confidence evaluation mechanism is used for fusion judgment, thereby realizing collaborative decision output.

[0253] Specifically, to avoid the uncertainty of a single algorithm affecting the core control, this embodiment introduces an inter-model collaboration mechanism through a decision fusion module:

[0254] 1. Introduce a credibility assessment mechanism to perform semantic matching and confidence scoring on the output results of large model modules;

[0255] 2. Use the output of the self-developed AI algorithm module as the main reference and compare its logical consistency with the candidate suggestions of the large model module;

[0256] 3. If a candidate suggestion receives high confidence and passes the expert system rules verification, it will be included in the fusion output;

[0257] 4. All fusion decision results enter the feedback module, where they are compared and corrected based on the post-execution status, forming a data-driven self-learning optimization mechanism.

[0258] This embodiment aims to integrate actual nuclear power plant operation data with virtual simulation models to construct a comprehensive system encompassing data acquisition, strategy allocation, intelligent decision-making, and optimized control. The system acquires multi-source heterogeneous data in real time through a data acquisition module, including nuclear power simulator output signals, historical operation records, maintenance logs, and inspection videos. The data strategy allocation module then categorizes and processes this data according to safety level, real-time performance, and data format. High-safety and highly real-time data are precisely processed by a self-developed AI algorithm module, while non-real-time and complex data are analyzed using a general large-scale model to extract pattern features. The system integrates physical mechanism models, machine learning algorithms, and multimodal large-scale models to construct high-confidence intelligent decision-making schemes. Through an optimized control module, it achieves closed-loop regulation of the nuclear power plant's operating status, offering advantages such as fast response speed, high processing accuracy, and adaptability to complex operating conditions. It is suitable for various scenarios, including nuclear power plant operation optimization, fault prediction, and energy efficiency improvement, significantly enhancing the intelligence level and operational safety of the nuclear power system. Furthermore, leveraging the multimodal learning and knowledge graph technology of the large-scale model module, potential risks (such as pipe cracks and sensor drift) are identified, and emergency response plans are simulated in the simulator. The standardized verification process using simulators ensures the reliability of the strategy while reducing delays in responding to new types of accidents. Furthermore, the closed-loop mechanism enables the system to autonomously evolve as operational data accumulates, adapting to performance changes in nuclear power equipment over long-term operation. Through deep collaboration between domestically developed large-scale models (such as DeepSeek) and independently controllable simulators, a safe and efficient intelligent decision-making platform for nuclear power is constructed. This platform reduces reliance on foreign technologies, enhances the digitalization level of the nuclear power industry, provides technical support for intelligent management of nuclear power companies, and helps achieve the goal of "zero accidents" and improved economic efficiency.

[0259] In the specific implementation process, by combining the actual operation of the nuclear power plant with a virtual simulation model, and utilizing advanced simulation technology, artificial intelligence, big data analysis, and other means, intelligent management and optimized control of the nuclear power plant are achieved. Specifically, the overall architecture for optimizing nuclear power plant operation includes a data acquisition module, a data allocation strategy module, a general large model module, a self-developed AI algorithm module, a decision fusion and feedback module, an optimized control module, and an intelligent decision advisor module.

[0260] The data acquisition module includes other data acquisition modules and a real-time data acquisition module. Both modules are connected to the data strategy allocation module. As a core foundational component of the entire system, the data acquisition module is responsible for efficiently aggregating multi-dimensional, multi-source, and multi-granular data from the nuclear power plant's operation, providing reliable data support for subsequent model training, intelligent decision-making, simulation verification, and optimized control. From a data source perspective, this module primarily consists of two sub-modules: a real-time data acquisition module and other data acquisition modules. These modules connect to the data strategy allocation module via a unified data interface protocol, forming a flexible, scalable, and highly reliable data flow system.

[0261] The real-time data acquisition module is primarily designed for high-speed, continuous, and structured data streams generated during the operation of nuclear power plants, including but not limited to the following key operating parameters:

[0262] 1. Physical quantity measurement data: such as reactor core outlet temperature, main steam temperature, coolant flow rate, pressure vessel pressure, auxiliary pump flow rate, etc. This data is typically collected by various high-precision industrial sensors deployed in key locations and uploaded via DCS (Distributed Control System) or SCADA system.

[0263] 2. Electrical quantity monitoring data: including main generator output voltage, current, active power, reactive power, power factor, etc. This type of data is used to determine load changes, energy balance status, and the operating efficiency of frequency regulation and voltage regulation systems.

[0264] 3. Control system feedback signals: such as control rod position signals, regulating valve opening, pump unit operating status (start / stop), interlock signals, etc. These signals not only reflect the current control command execution status of the system, but also provide key information for judging operational stability and predicting faults.

[0265] 4. Simulation platform output data: In a high-fidelity simulation system, the real-time operating parameters simulated by virtual devices (such as virtual instrument readings, simulated alarm signals, etc.) are also included in the real-time data category to ensure the synchronization and reliability of simulation with real operation.

[0266] These real-time data are characterized by high update frequency (≥100Hz), high degree of structure, and strong data integrity, making them suitable for direct input into self-developed AI algorithm modules for highly timely intelligent processing, such as rapid dynamic response, model prediction and correction, and online edge detection. During data access, to ensure low latency and high availability of data transmission, the system adopts high-performance industrial communication protocols (such as OPC UA, Modbus TCP, etc.) and supports functions such as breakpoint resumption, dual-channel redundant acquisition, and timestamp synchronization to guarantee data real-time performance and security.

[0267] Other data acquisition modules primarily focus on non-real-time, low-frequency, semi-structured, or unstructured data sources. While these data are not essential for real-time control of nuclear power plant operations, they play a crucial role in fault diagnosis, experience reuse, trend analysis, and knowledge graph construction. Specifically, they include the following categories:

[0268] 1. Historical operation records: such as historical operation data stored in the form of SQL logs, CSV data tables or PDF documents, covering various past start-up and shutdown conditions, load adjustment processes, abnormal event records, operation reports, etc., which can be used to extract feature patterns, train prediction models, and reproduce abnormal conditions.

[0269] 2. Inspection video and image data: Video streams and image snapshots from factory camera systems, wearable terminals or drone inspections reflect detailed information such as equipment surface condition, instrument pointer position, and signs of condensate leakage. They are suitable for input into the multimodal processing module of the large model module for semantic extraction and image recognition.

[0270] 3. Maintenance Log: Records the maintenance cycle, inspection process, problem description, handling results, and replacement parts for various equipment. This type of data is mostly in text form, with information scattered and expressed in various ways, but it contains a large amount of fault causal chains and maintenance experience, which is extremely valuable for intelligent pre-maintenance and risk assessment.

[0271] 4. Environmental climate and external conditions data: External factors such as temperature, humidity, wind speed, external power grid frequency fluctuations, and earthquake monitoring data, although not directly involved in reactor control, have a significant impact on the performance of auxiliary systems (such as cooling systems and ventilation systems) and are one of the important inputs for building a global operational status analysis model.

[0272] To ensure unified access to different types of data sources, the system is equipped with various format conversion and protocol adaptation tools (such as document parsers, OCR engines, and video-to-image modules) in other data acquisition modules, achieving unified vectorized encoding between text, images, videos, and structured tables. Simultaneously, it supports data acquisition through multiple methods, including REST API, batch import via FTP, and direct database connection, ensuring the comprehensiveness and flexibility of historical data access.

[0273] Whether it's real-time data or other types of data, all data is aggregated through standardized data interfaces to the data strategy allocation module for further tagging, prioritization, and downstream task scheduling. This module constructs a unified data standard system and a multi-dimensional tagging system (such as security level, data format, and real-time tags) to perform semantic understanding and usage guidance on the data, thereby achieving precise scheduling and collaborative utilization of data between the large model and the self-developed AI algorithm module.

[0274] For example:

[0275] The time-series data of a certain pressure sensor will be automatically marked as "structured / real-time / high priority / safety related" and will be preferentially input into the self-developed AI algorithm module for short-cycle dynamic prediction.

[0276] A maintenance record PDF document will be marked as "unstructured / non-real-time / secondary priority / maintenance assistance" for use in generating maintenance suggestions and knowledge graph supplements for large model modules;

[0277] A video inspection segment is labeled as "unstructured / image / medium priority / fault identification," triggering the multimodal recognition module to screen for apparent anomalies.

[0278] Through this organic integration and intelligent allocation, the data acquisition module not only completes the basic function of "collecting data", but also becomes the central hub of "intelligent data flow" in the entire closed-loop system, providing strong data support for model inference, control command generation, and operation strategy optimization.

[0279] The data strategy allocation module establishes a multi-dimensional data tagging system, which classifies data sequentially according to preset priority strategies based on security level, real-time performance, and data format.

[0280] In this embodiment, the data strategy allocation module undertakes the crucial data central scheduling function, serving as a bridge connecting the data acquisition module and downstream intelligent processing modules (including the self-developed AI algorithm module, large model module, decision feedback module, etc.). This data strategy allocation module not only completes data classification, preprocessing, and tag management, but also formulates differentiated data processing paths based on multiple dimensions such as data security level, real-time requirements, and degree of structure, ensuring that various types of data flow and are used within the system in the most reasonable, efficient, and secure manner.

[0281] To achieve precise management and intelligent distribution of massive amounts of heterogeneous data, the data strategy allocation module has constructed a systematic multi-dimensional data tagging system. This data tagging system revolves around three core dimensions:

[0282] 1. Data Security Level: Divided into high security level (such as nuclear safety related signals), medium security level (such as equipment status parameters), and low security level (such as environmental monitoring data).

[0283] 2. Real-time dimension (Data Timeliness): Divided into strong real-time data (update frequency ≥ 100Hz), weak real-time data (1Hz~100Hz), and non-real-time data (updated on an hourly or daily basis).

[0284] 3. Data Format Dimension: Divided into structured data (such as time series tables, SQL databases), semi-structured data (such as XML / JSON logs), and unstructured data (such as natural language text, images, videos, and audio).

[0285] Based on the data labels across these three dimensions, the system has established a pre-defined priority strategy matrix to automatically classify, label, and allocate processing paths for each type of data. The following sections provide a detailed explanation of each strategy type:

[0286] The safety-first strategy refers to the fact that in the complex operating environment of nuclear power plants, some data is related to equipment safety, personnel safety, and even the nuclear safety boundary, requiring extremely high accuracy and controllability in its processing. Therefore, the system adopts a mandatory classification and control strategy for the safety level dimension:

[0287] High-safety-level data includes critical safety parameters such as reactor power limit thresholds, coolant pressure alarm signals, emergency shutdown trigger signals, and spent fuel pool water level alarms. This type of data must be processed by a self-developed AI algorithm module that has undergone safety certification. Its modeling logic is based on nuclear power plant operation mechanisms and control logic, and it can provide interpretable and traceable judgments. It strictly avoids the use of probabilistic language generation methods based on large models, thus mitigating the potential uncertainty risks associated with it.

[0288] For low-to-medium safety level data, such as equipment health index (HI), cooling water energy consumption, temperature difference efficiency coefficient, and maintenance cycle, candidate suggested solutions can first be generated by a general large model (such as a finely tuned DeepSeek). Then, the rationality and feasibility of these solutions are verified by a self-developed AI algorithm module. Finally, the data is handed over to the "decision fusion and feedback module" for unified scheduling and feedback closed-loop control. This data processing strategy improves reasoning efficiency and knowledge generalization ability while ensuring safety.

[0289] Real-time response strategy refers to the fact that nuclear power plant control systems are extremely sensitive to time response, especially in scenarios involving dynamic response, closed-loop control, and accident prediction, where data processing latency must be strictly controlled at the millisecond level. Therefore, the system has developed the following classification and processing strategies based on the real-time characteristics of the data:

[0290] High-latency real-time data (≥100Hz): Such as time-series signals like temperature, pressure, flow rate, control rod position signals, and main pump speed of core components. These signals have a high update frequency and strong dynamics, requiring low-latency processing capabilities of ≤10ms. This type of data is processed preferentially by a self-developed AI algorithm module. The kernel of this algorithm often includes differential equation solvers, dynamic models of control systems, Kalman filters, and multivariable state estimators, possessing high robustness and stability, and effectively supporting dynamic simulation and real-time control decision-making.

[0291] Non-real-time data, such as historical operating logs, equipment maintenance records, and operation summary reports, typically has an update cycle of minutes to hours. This type of data is suitable for inputting into general-purpose models for in-depth pattern mining and knowledge extraction. This type of data analysis focuses more on long-term trend prediction, potential fault location, and operational experience accumulation, and can support the construction of visualization analysis tools such as "operational error risk maps" and "critical equipment failure evolution paths."

[0292] Data format processing strategy refers to the different processing methods determined by different data formats. Based on the degree of data structure, the system defines the following path:

[0293] Structured data processing path: This includes time-series data collected by industrial instruments, operational reports exported from DCS systems, database tables, etc. This type of data is processed first by the self-developed AI algorithm module. Before being input into the model, data cleaning, normalization, and physical parameter mapping are performed. The algorithm combines nuclear reaction kinetic equations, energy conservation equations, and mass conservation equations for modeling and analysis, which is suitable for quantitative simulation, parameter optimization, and boundary calculation.

[0294] Unstructured data processing paths include: operation logs (natural language), maintenance case documents, accident investigation reports, graphical instrument readings, and monitoring videos. These data vary in format and have complex semantics, making them unsuitable for traditional algorithm processing. The system utilizes the NLP and multimodal parsing capabilities of a general-purpose large model to perform vectorized semantic encoding, keyword extraction, and image / text OCR fusion processing to extract implicit information such as "error operation mode," "fault correlation causal chain," and "abnormal equipment status in images."

[0295] Semi-structured data processing path: such as system operation configuration in XML format, alarm information records in JSON format, etc., although they contain a certain structure, there is free text nesting. The system first converts them into a structured form and then assigns them to the self-developed AI algorithm module or large model module channel according to their security level and real-time label.

[0296] To ensure that various types of data can quickly match the optimal processing path under label cross-referencing, the system introduces a label fusion and path optimization mechanism. After receiving a set of data, this mechanism automatically invokes the label inference engine, comprehensively considers its three-dimensional label weights, and determines the optimal processing method through decision tree or Bayesian path selection algorithms. For example:

[0297] If a set of data with an update frequency of 200Hz, a data type of pressure signal, and a high safety level has a tag fusion result of: high priority / strong real-time / structured, then the system will automatically schedule it to enter the "self-developed AI algorithm module fast processing channel".

[0298] If an operator error log text is tagged as medium priority / non-real-time / unstructured, it will be automatically routed to the "general large model semantic parsing module" and labeled as "supports decision suggestion input".

[0299] This fusion mechanism effectively avoids the fragmentation problem of data processing strategies and improves the intelligence and automation level of data scheduling. Through the implementation of the above data strategy allocation mechanism, this system maximizes data utilization efficiency while ensuring nuclear security and real-time response. It allows the self-developed AI algorithm module and the general large model module to each play to their strengths, building a collaborative data intelligence system of "precise modeling + generalized reasoning + intelligent fusion," which becomes an important engine supporting the continuous evolution of the closed-loop optimization system.

[0300] The general-purpose large model module comprises a preprocessing layer, a model layer, and an output layer. The preprocessing layer performs entity recognition on unstructured data (e.g., extracting key entities such as "pressure vessel temperature" and "control rod insertion depth") and vectorizes time-series data (e.g., converting sensor sequences into event encoding matrices). The model layer employs domain adaptation techniques (e.g., Prompt Tuning, LoRA) to fine-tune the general-purpose large model module for nuclear power scenarios, focusing on optimizing tasks such as abnormal operating condition diagnosis and compliance verification of operating procedures. The output layer generates probability distribution-based results (e.g., "the probability of steam generator leakage under current operating conditions is 87%)" and includes a confidence score for subsequent coupled decision-making.

[0301] The self-developed AI algorithm module comprises three parts: physics model-driven, data-driven optimization, and real-time performance assurance. The physics model-driven approach is based on nuclear engineering mechanistic models (such as point reactor dynamics equations and thermal-hydraulic models), employing numerical integration methods (such as Runge-Kutta) for deterministic calculations to output precise physical quantities (such as core power distribution and coolant temperature gradient). Data-driven optimization addresses complex nonlinear aspects difficult to model using mechanistic models (such as material aging effects during long-term operation), supplementing modeling with self-developed machine learning models (such as customized LSTM and graph neural networks) to improve the accuracy of edge case predictions. Real-time performance assurance ensures that the self-developed AI algorithm module meets the ≤5ms cycle time requirement in embedded real-time systems through model quantization (FP32→FP16) and hardware acceleration (GPU / ASIC deployment).

[0302] The decision fusion and feedback module primarily implements the fusion of multi-source model outputs, confidence-weighted decision-making, and a closed-loop feedback update mechanism to build a robust and reliable intelligent decision engine. This module consists of a candidate solution aggregation submodule, a confidence mechanism evaluation submodule, a fusion strategy execution submodule, and an operational feedback learning submodule.

[0303] The candidate solution aggregation submodule receives various decision suggestions from the general large model module and the self-developed AI algorithm module, including predictive indicators (such as equipment failure probability and operational deviation risk), optimized control variables (such as control rod adjustment schemes and coolant flow change schemes), and alarm suggestions (such as suggestions for early inspection or equipment status warnings). All candidate solutions are tagged and managed according to attributes such as data source, model type, and computation latency to facilitate subsequent integration and traceability.

[0304] Confidence Mechanism Evaluation Submodule: Evaluates the confidence level of each candidate solution based on the following dimensions:

[0305] 1. Model training coverage: Determines whether the current working condition falls within the training data distribution; if it exceeds the range, the confidence level is automatically lowered.

[0306] 2. Model applicability boundaries: For example, large models are good at handling pattern recognition tasks, but their confidence in high-precision numerical prediction is discounted;

[0307] 3. Real-time requirement satisfaction: If the solution generation delay exceeds the preset threshold (e.g., 10ms), the result will be used for long-term optimization and excluded from emergency response tasks;

[0308] 4. Multi-model consistency test: If the general large model module and the self-developed AI algorithm module make the same judgment on a certain event, the overall confidence level is increased; otherwise, it enters the manual review queue.

[0309] The fusion strategy execution submodule employs a multi-model weighted voting mechanism, Bayesian inference fusion method, or confidence-based dynamic strategy switching mechanism to ultimately generate an executable control strategy and early warning suggestions. In special circumstances, manual intervention or interruption of model decision-making by the upper-level control system is permitted to ensure operational safety.

[0310] The runtime feedback learning submodule: This module transmits the operational results (such as response latency, actual failure rate, and load fluctuations) back in real time for dynamic updates.

[0311] The optimized control module and the intelligent decision advisor module constitute the execution closed loop of intelligent control, which directly affects the adjustment of key operating parameters of the nuclear power plant, the action commands of control equipment, and the operation guidance of the operation and maintenance team.

[0312] The optimized control module can output control commands to the nuclear power simulation system or the actual control interface, and obtain execution status feedback in real time. By rapidly analyzing the fused optimal control strategy and generating commands, the optimized control module drives the full-range nuclear power simulator system to make precise adjustments. Its core components include:

[0313] 1. Control Strategy Parsing Unit: Transforms abstract layer decision results (such as "suggest reducing main pump power by 10%) into specific control commands (such as "adjust the main pump speed setpoint from 1450 rpm to 1305 rpm"), supporting multi-time scale adjustment (ms-level fast response / min-level rolling optimization).

[0314] 2. Control command issuance and synchronization unit: It adopts a distributed industrial protocol based on OPC-UA or DDS (Data Distribution Service) to send control commands to the simulation platform or DCS system, and at the same time realizes multi-channel redundant broadcasting and execution feedback acquisition.

[0315] 3. Safety threshold verification mechanism: Before all instructions are issued, they are logically checked by the constraint verifier of the self-developed AI algorithm module to ensure that no safety limits are triggered (such as "coolant flow rate is not lower than the minimum critical value" and "control rod position change rate must not exceed physical limit"), thus ensuring the stable operation of the system.

[0316] The intelligent decision advisor module, serving as the system's "human-machine interface layer," primarily provides model-based decision support and operational recommendations for operations engineers and scheduling managers. Its design features include:

[0317] 1. Multimodal interactive terminal: Supports multiple interaction methods including text, image, and voice. Dispatchers can input current operating conditions via natural language (e.g., "Why is the main steam temperature too high?"), and the system automatically retrieves cause analysis reports from the simulation model and historical database and generates visual charts.

[0318] 2. Operation Recommendation Generator: Based on the current operating status and model inference results, it automatically generates operation suggestions (e.g., "It is recommended to switch ventilation fan A to standby mode to reduce energy consumption", and attaches comparative data of similar historical operating conditions and expected energy-saving effects).

[0319] 3. Knowledge Graph and Case Library Integration: The system integrates a nuclear power plant operation knowledge graph and a classic operating condition case library to perform semantic matching on the current abnormal state, push similar cases and handling experience, and help young operators respond quickly to complex situations.

[0320] 4. Feedback and incentive mechanism: Operators are encouraged to evaluate the model's recommended operations (e.g., mark them as "effective" or "ineffective"). The evaluation results will be used to reinforce the labels of the training data, providing real-world feedback for the model's continuous learning.

[0321] This embodiment inputs high-safety-level and highly real-time data, after data classification, into the AI ​​algorithm module, while inputting non-real-time or unstructured data into the large model module. By balancing and fusing the AI ​​algorithm predictions with the large model predictions, fault prediction results for the nuclear power plant are obtained. These results are then combined with a nuclear power plant simulator to obtain simulated operation results. Based on these simulated operation results, the operation of the nuclear power plant is optimized, improving intelligent decision-making capabilities, reducing high verification costs, and constructing a closed-loop iterative optimization system. Specifically, by combining core dimensions such as safety level, real-time performance, and data format, refined classification and intelligent strategy allocation of various types of data in the nuclear power system are achieved. High-safety-level data is processed by a controllable, self-developed AI algorithm module, ensuring operational safety and interpretable results; highly real-time data is processed by differential models and filtering algorithms, achieving millisecond-level response to meet the rapid dynamic requirements of the control system; non-real-time and unstructured data undergo semantic parsing and pattern mining through the large model, improving knowledge extraction and decision support capabilities. Through the tag fusion mechanism, the optimal processing path can be automatically matched for different data, realizing the high efficiency and intelligence of data flow. This effectively improves the data utilization efficiency, safety assurance level and intelligent analysis capability of the nuclear power full-range simulator, and provides a solid data foundation and processing support for subsequent intelligent operation and maintenance, auxiliary decision-making and fault prediction.

[0322] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0323] Example 3

[0324] Figure 3 This is a schematic diagram of the structure of an electronic device according to Embodiment 3 of this disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the optimization method for nuclear power plant operation described in any of the above embodiments. Figure 3 The electronic device 90 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0325] like Figure 3As shown, the electronic device 90 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 90 may include, but are not limited to: at least one processor 91, at least one memory 92, and a bus 93 connecting different system components (including memory 92 and processor 91).

[0326] Bus 93 includes a data bus, an address bus, and a control bus.

[0327] The memory 92 may include volatile memory, such as random access memory (RAM) 921 and / or cache memory 922, and may further include read-only memory (ROM) 923.

[0328] The memory 92 may also include a program tool 925 (or utility) having a set (at least one) program module 924, such program module 924 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0329] The processor 91 executes various functional applications and data processing by running computer programs stored in the memory 92, such as the nuclear power plant operation optimization method provided in any of the above embodiments.

[0330] Electronic device 90 can also communicate with one or more external devices 94 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 95. Furthermore, electronic device 90 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 96. Figure 3 As shown, network adapter 96 communicates with other modules of electronic device 90 via bus 93. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with electronic device 90, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0331] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0332] Example 4

[0333] Embodiment 4 of this disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the optimization method for nuclear power plant operation provided in any of the above embodiments.

[0334] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0335] Example 5

[0336] Embodiment 5 of this disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the optimization method for nuclear power plant operation described in any of the above embodiments.

[0337] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0338] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. An optimization method for the operation of a nuclear power plant, characterized in that, The optimization method includes: Acquire historical and real-time operational data of nuclear power plants; The historical operational data and the real-time operational data are classified to obtain data classification results; High-security and real-time data, as well as non-real-time or unstructured data, are obtained from the data classification results. The high-security and real-time data are input into the AI ​​algorithm module to obtain the AI ​​algorithm prediction results, and the non-real-time or unstructured data are input into the large model module to obtain the large model prediction results. The AI ​​algorithm prediction results are combined with the large model prediction results to obtain the fault detection results of the nuclear power plant. The fault measurement results of the nuclear power plant are input into the nuclear power plant simulator for operation to obtain the simulation operation results; The operation of the nuclear power plant was optimized based on the simulation results.

2. The method for optimizing nuclear power plant operation as described in claim 1, characterized in that, The steps for optimizing the operation of the nuclear power plant based on the simulation results include: In response to the simulation results achieving the expected results, the simulation results are taken as the actual operating results, and the nuclear power plant is optimized based on the actual operating results.

3. The method for optimizing nuclear power plant operation as described in claim 1, characterized in that, The step of optimizing the nuclear power plant operation based on the simulation results also includes: If the simulation results do not meet expectations, the process returns to the step of classifying the historical and real-time running data to obtain the data classification results.

4. The method for optimizing nuclear power plant operation as described in claim 1, characterized in that, The step of classifying the historical operational data and the real-time operational data to obtain data classification results includes: The historical operation data and the real-time operation data are classified according to preset data tags to obtain data classification results.

5. The method for optimizing the operation of a nuclear power plant as described in claim 1, characterized in that, The step of inputting the non-real-time or unstructured data into the large model module to obtain the large model prediction results includes: Based on the large model module, semantic understanding and deep feature extraction are performed on the non-real-time or unstructured data to obtain the large model prediction results of potential decision-related factors in operational behavior, text description, and image information. And / or, The step of balancing and fusing the AI ​​algorithm prediction results with the large model prediction results to obtain the fault detection results of the nuclear power plant includes: The AI ​​algorithm prediction results and the large model prediction results are weighed and fused based on confidence factors and risk assessment indicators to obtain the fault prediction results of the nuclear power plant.

6. The method for optimizing nuclear power plant operation as described in claim 4, characterized in that, The preset data tags include at least one of the following dimensions: security level, real-time performance, and data format. The security level dimension includes at least one of high level, medium level, and low level; The real-time dimension includes at least one of strong real-time, weak real-time, and non-real-time. The data format dimensions include structured and unstructured.

7. An optimization system for nuclear power plant operation, characterized in that, The optimization system includes: The first acquisition module is used to acquire historical and real-time operating data of the nuclear power plant. The classification module is used to classify the historical operation data and the real-time operation data to obtain data classification results; The second acquisition module is used to acquire high-security-level and highly real-time data, as well as non-real-time data or unstructured data from the data classification results. The third acquisition module is used to input the high-security and high-real-time data into the AI ​​algorithm module to obtain the AI ​​algorithm prediction results, and to input the non-real-time data or unstructured data into the large model module to obtain the large model prediction results. The processing module is used to weigh and fuse the prediction results of the AI ​​algorithm with the prediction results of the large model to obtain the fault detection results of the nuclear power plant. The operation module is used to input the fault measurement results of the nuclear power plant into the nuclear power plant simulator for operation and to obtain the simulation operation results; An optimization module is used to optimize the operation of the nuclear power plant based on the simulation results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the optimized method for operating a nuclear power plant as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the optimization method for operating a nuclear power plant as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the optimization method for nuclear power plant operation as described in any one of claims 1 to 6.

Images