Intelligent diagnosis method and system for abnormal working condition of nuclear power unit

Abstract

The present application relates to a kind of nuclear power unit abnormal condition intelligent diagnosis method and system, it is related to nuclear power unit technical field, including the time series dataset of fusion unit operating state and power grid disturbance characteristics;Extract the deep feature vector reflecting the coupling relationship of equipment level thermal hydraulic anomaly and system level electromagnetic transient disturbance;Risk index set is constructed, including the instantaneous value of core damage frequency, the probability of large-scale power outage condition and the importance of comprehensive risk;Risk index set is input as state space into multi-agent deep deterministic policy gradient algorithm framework, generates hierarchical communication instruction sequence and power adjustment suggestion;Through time-sensitive network gateway, realize double-channel sub-microsecond clock synchronization and seamless redundant switching;Collect the reactor power response curve after instruction execution and power grid frequency recovery trajectory and feedback to step one.The present application guarantees the collaborative control reliability of nuclear power plant reactor power regulation system and power grid automatic generation control device.

Description

Technical Field

[0001] This invention relates to the field of nuclear power unit technology, and more specifically, to an intelligent diagnostic method and system for abnormal operating conditions of nuclear power units. Background Technology

[0002] As nuclear power plants play an increasingly prominent role as baseload power sources in new power systems, the need for coordinated assurance of reactor operation safety and grid stability is becoming increasingly urgent. Current diagnostic technologies for abnormal operating conditions in nuclear power plants primarily employ logic diagnostic systems based on fixed thresholds and offline probabilistic safety assessment methods. These methods trigger alarms by setting thresholds for single physical parameters such as reactor coolant system pressure and temperature, and combine this with periodic manual review to complete risk assessments. While these technologies can identify basic anomalies, they suffer from inherent drawbacks such as fragmented diagnostic dimensions, delayed risk quantification, and dispersed communication and dispatch. Fixed thresholds cannot adapt to the dynamic coupling characteristics of nuclear power plant power regulation with grid frequency fluctuations and voltage dips; offline assessments struggle to support second-level real-time decision-making requirements; and the lack of a real-time risk information exchange channel between the nuclear power plant's operation control system and the grid dispatch center leads to delays in power supply response under abnormal operating conditions. Furthermore, existing technologies mostly use a single wired communication method to transmit control commands, without considering communication redundancy guarantees under extreme natural disasters or network attacks. This makes it difficult to meet the high reliability and low latency collaborative control requirements of nuclear power plants under abnormal operating conditions, and cannot effectively cope with the cascading risks of faults on both sides of the nuclear power plant and the power grid.

[0003] To address the issues of independent anomaly diagnosis and grid risk dispatching in existing technologies, and the lack of a linkage mechanism between risk quantification and real-time communication, this application provides an intelligent diagnostic method for abnormal operating conditions of nuclear power units. This method constructs a closed-loop technical system integrating physical information neural networks, dynamic event tree Bayesian networks, multi-agent deep reinforcement learning, and heterogeneous redundant communication. It extracts cross-domain coupling features between nuclear power plant thermal-hydraulic parameters and grid electrical quantities by embedding a neural network with reactor neutron dynamics equations and thermal-hydraulic conservation equations. Real-time risk quantification of core damage frequency and grid stability margin is achieved through the fusion calculation of dynamic event trees and Bayesian networks. A multi-agent deep deterministic strategy gradient algorithm optimizes the hierarchical communication strategies and power adjustment schemes for the nuclear power plant shift leader, grid dispatch center, and emergency command center. A heterogeneous redundant transmission architecture composed of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication ensures reliable transmission of control commands to the nuclear power plant reactor power regulation system and the grid automatic generation control device. Finally, execution feedback is used to dynamically correct the anomaly detection threshold. This method realizes an integrated closed loop of intelligent diagnosis of abnormal operating conditions in nuclear power plants, quantitative risk assessment, and intelligent communication scheduling for power supply protection. It significantly improves the accuracy of anomaly identification and emergency response speed of nuclear power plants in complex power grid environments, and effectively ensures the operational safety of nuclear power plants and the reliability of power grid supply. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent diagnostic method and system for abnormal operating conditions of nuclear power units, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides an intelligent diagnostic method for abnormal operating conditions of nuclear power units, including: Step 1: Collect real-time monitoring data on the pressure, temperature, neutron flux density of the reactor coolant system of the nuclear power unit, as well as the frequency deviation and voltage fluctuation amplitude on the grid side. After spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data, construct a time-series dataset that integrates the unit's operating status and grid disturbance characteristics. Step 2: Feed the time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and thermal-hydraulic conservation equations. Solve the physical equation residuals through an automatic differentiation mechanism to constrain network training, forming a dynamic response feature of the unit that integrates physical laws and data-driven approaches. Map the dynamic response feature of the unit to a heterogeneous graph neural network topology composed of the primary coolant pressure boundary nodes and the voltage phasors of the key power grid bus. Learn the cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response through spatiotemporal graph convolution operations, constructing the spatiotemporal correlation feature of unit-grid coupling. Load the spatiotemporal correlation feature into a variational autoencoder structure to model the latent space distribution of normal and abnormal operating conditions. Infer the probability distribution of potential abnormal modes by maximizing the lower bound of evidence, and extract a deep feature vector reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances. Step 3: Using deep feature vectors as evidence variables, a dynamic event tree analysis framework is constructed with reactor protection system triggering, turbine digital electro-hydraulic control system fast shutdown action, and grid low-frequency low-pressure load shedding device activation as the top events. Accident evolution branch paths are generated through adaptive importance sampling to form a candidate accident sequence set. This candidate accident sequence set is then loaded into a variational inference-based Bayesian network inference engine. The posterior distribution of equipment state failure probabilities is updated using deep feature vectors, and the occurrence probability of each accident sequence is calculated. Based on the occurrence probability of each accident sequence, the Frosley-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. This is combined with real-time calculated core damage frequency and grid transient voltage stability margin assessment values ​​to construct a risk index set including instantaneous core damage frequency values, large-scale power outage condition probabilities, and comprehensive risk importance. Step 4: Input the risk index set as the state space into the multi-agent deep deterministic policy gradient algorithm framework. Configure the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. Design a reward function with the goal of minimizing the weighted entropy value of the risk index set. Optimize the communication protocol priority and power adjustment strategy under different risk levels through multi-scenario game learning in a digital twin environment. Output automatic generation control and regulation commands for the power grid dispatch center, reactor power regulation rod position suggestions for the nuclear power plant shift supervisor, and load transfer paths for the emergency command center. Generate a hierarchical communication command sequence and power adjustment suggestions. Step 5: Encode the hierarchical communication command sequence and power adjustment suggestions into general object-oriented substation event messages and manufacturing message specification messages that conform to substation communication standards. Transmit these messages through a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. Enable ultra-reliable low-latency communication mode and configure a dual-transmitter selective reception protection mechanism in the fifth-generation mobile communication network slice. Use orthogonal frequency division multiplexing multi-carrier modulation technology in the power line carrier channel. Achieve sub-microsecond clock synchronization and seamless redundancy switching for dual channels through a time-sensitive network gateway. Send the collaborative control commands to the nuclear power unit reactor power regulation system and the grid automatic generation control device respectively. Collect the reactor power response curve and grid frequency recovery trajectory after command execution and feed them back to Step 1 as input conditions for the next round of dynamic correction of abnormal monitoring data thresholds.

[0005] Preferably, step one includes: Piezoresistive pressure sensors and platinum resistance temperature sensors are installed in the hot and cold sections of the reactor coolant system of the nuclear power unit. A self-powered neutron detector is installed at the core outlet. A synchronous phasor measurement unit is installed on the high-voltage bus of the power grid. The original analog signals of pressure, temperature, neutron flux density, frequency deviation, and voltage fluctuation amplitude are collected at a sampling rate of 100 frames per second. After quantization and anti-aliasing filtering by a 16-bit analog-to-digital converter, the original digital sampling sequence is formed. The original digital sampling sequence is input into the clock synchronization engine based on the core period characteristics. The IEEE 1588 precise time protocol is used to align the clock between the nuclear power unit's distributed control system and the power grid's wide-area measurement system. The difference in sampling time between heterogeneous systems is compensated by cubic spline interpolation algorithm. The unified timestamp across systems is reconstructed with a period of 10 milliseconds to build a unified timestamp dataset. The unified timestamp dataset was loaded into a data cleaning engine based on sliding window Z-score anomaly detection and Min-Max normalization. The threshold for neutron flux density signal spike detection was set to 3 times the standard deviation, and the threshold for grid frequency deviation step detection was set to 0.5 Hz. After removing outlier samples, Kalman filtering was used to compensate for missing values ​​through state estimation. A 60-second sliding window was used for moving average smoothing to construct a time series dataset that integrates unit operating status and grid disturbance characteristics.

[0006] Preferably, step two includes: The time-series dataset is fed into a physical information neural network architecture, which embeds the reactor point neutron dynamics equation and the thermal-hydraulic conservation equation. The residual loss function of the physical equation is calculated through an automatic differentiation mechanism, and the total loss function is minimized through the Adam optimizer to form the dynamic response characteristics of the unit that integrate physical laws and data-driven approaches. The dynamic response characteristics of the unit are mapped to the set of primary loop coolant pressure boundary nodes and the set of voltage phasors of key power grid buses. A heterogeneous graph neural network topology is constructed, in which the node feature matrix contains the reactor power level and the grid frequency deviation, and the edge weight matrix represents the thermal-hydraulic impedance and the transmission line admittance. The cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response is learned through spatiotemporal graph convolution operation, thus constructing the spatiotemporal correlation characteristics of unit-grid coupling. Spatiotemporal correlation features are loaded into a variational autoencoder structure. The encoder network compresses the high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The decoder network reconstructs the distribution of normal operating conditions features. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence. A deep feature vector reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances is extracted.

[0007] Preferably, step three includes: Using deep feature vectors as evidence variables, a dynamic event tree analysis framework is constructed. The reactor protection system triggering, the turbine digital electro-hydraulic control system fast shutdown action, and the power grid low-frequency low-pressure load shedding device start-up are set as top events. An adaptive importance sampling algorithm is adopted to adjust the branch sampling weights according to the probability of abnormal modes in the deep feature vectors. The accident evolution branch path is generated through sequence Monte Carlo simulation to form a candidate accident sequence set. The candidate accident sequence set is loaded into a Bayesian network inference engine based on variational inference, and a directed acyclic graph with nodes representing equipment status, human actions, and environmental parameters is constructed. The conditional probability table is updated using deep feature vectors as evidence input, and the occurrence probability of each accident sequence is calculated. Based on the occurrence probability of each accident sequence, the Frosley-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The risk index set is constructed by integrating the real-time calculated core damage frequency and the power grid transient voltage stability margin assessment value.

[0008] Preferably, step four includes: The multi-agent deep deterministic policy gradient algorithm framework uses the risk index set as the input to the state space. The power grid dispatch center, the nuclear power plant shift leader, and the emergency command center are configured as agents with independent local observations. A reward function is designed with the goal of minimizing the weighted entropy value of the risk index set. The communication protocol priority and power adjustment strategy are optimized through multi-scenario game learning in a digital twin environment. The digital twin environment includes a reactor physics and thermal model, a turbine speed regulation model, and a power grid transient stability model. Three types of training scenarios are set: normal operating condition, design baseline accident condition, and above-design baseline accident condition. A centralized training and decentralized execution mechanism is adopted. Each agent generates actions independently through the policy network. After evaluation by a global value function, the network parameters are updated through policy gradient ascent to optimize the communication protocol priority and power adjustment strategy under different risk levels. It outputs automatic generation control and regulation commands to the power grid dispatch center, reactor power regulation rod position suggestions to the nuclear power plant operating shift leader, and load transfer paths to the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions.

[0009] Secondly, this application also provides an intelligent diagnostic system for abnormal operating conditions of nuclear power units, including: Step 1, Acquisition Module: This module is used to collect real-time monitoring data on the pressure, temperature, neutron flux density of the reactor coolant system of the nuclear power unit, as well as the frequency deviation and voltage fluctuation amplitude on the grid side. After spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data, a time-series dataset integrating the unit's operating status and grid disturbance characteristics is constructed. Step 2, Association Module: This module feeds the time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and thermal-hydraulic conservation equations. The residuals of the physical equations are solved using an automatic differentiation mechanism to constrain network training, forming dynamic response characteristics of the unit that integrate physical laws and data-driven approaches. These dynamic response characteristics are mapped to a heterogeneous graph neural network topology composed of primary coolant pressure boundary nodes and grid key bus voltage phasors. Spatiotemporal graph convolution operations are used to learn the cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response, constructing spatiotemporal correlation characteristics of the unit-grid coupling. These spatiotemporal correlation characteristics are then loaded into a variational autoencoder structure to model the latent space distribution of normal and abnormal operating conditions. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence, extracting deep feature vectors reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances. Step 3: Construction Module: This module uses deep feature vectors as evidence variables to construct a dynamic event tree analysis framework with reactor protection system triggering, turbine digital electro-hydraulic control system fast shutdown action, and grid low-frequency low-pressure load shedding device activation as the top events. Adaptive importance sampling generates accident evolution branch paths, forming a candidate accident sequence set. The candidate accident sequence set is loaded into a variational inference-based Bayesian network inference engine, updating the posterior distribution of equipment state failure probabilities with deep feature vectors, and calculating the occurrence probability of each accident sequence. Based on the occurrence probability of each accident sequence, the Frosley-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. This is combined with real-time calculated core damage frequency and grid transient voltage stability margin assessment values ​​to construct a risk index set including instantaneous core damage frequency values, large-scale power outage condition probabilities, and comprehensive risk importance. Step 4, Output Generation Module: This module uses the risk index set as input to the state space of a multi-agent deep deterministic policy gradient algorithm framework. It configures the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. It designs a reward function with the goal of minimizing the weighted entropy value of the risk index set. Through multi-scenario game learning in a digital twin environment, it optimizes the communication protocol priority and power adjustment strategy under different risk levels. It outputs automatic generation control and regulation commands for the power grid dispatch center, reactor power regulation rod position suggestions for the nuclear power plant shift supervisor, and load transfer paths for the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions. Step 5, Diagnostic Module: This module encodes the hierarchical communication command sequence and power adjustment suggestions into general object-oriented substation event messages and manufacturing message specifications that conform to substation communication standards. These messages are transmitted via a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. In the fifth-generation mobile communication network slice, an ultra-reliable low-latency communication mode is enabled, and a dual-transmitter selective reception protection mechanism is configured. Orthogonal frequency division multiplexing multi-carrier modulation technology is used in the power line carrier channel. A time-sensitive network gateway achieves sub-microsecond clock synchronization and seamless redundancy switching between the two channels. Cooperative control commands are then sent to the nuclear power unit reactor power regulation system and the grid automatic generation control device, respectively. The reactor power response curve and grid frequency recovery trajectory after command execution are collected and fed back to Step 1 as input conditions for the next round of dynamic correction of abnormal monitoring data thresholds.

[0010] Thirdly, this application also provides an intelligent diagnostic device for abnormal operating conditions of nuclear power units, comprising: Memory, used to store computer programs; A processor is used to implement the steps of the intelligent diagnostic method for abnormal operating conditions of nuclear power units when executing the computer program.

[0011] Fourthly, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent diagnostic method for abnormal operating conditions of nuclear power units.

[0012] The beneficial effects of this invention are as follows: This invention achieves deep integration of nuclear power plant operating status and grid disturbance characteristics by collecting real-time monitoring data of nuclear power unit reactor coolant system pressure, temperature, neutron flux density, grid-side frequency deviation, and voltage fluctuation amplitude, and by performing spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data. This solves the problem of data separation between the unit side and the grid side in the prior art and improves the comprehensiveness and accuracy of abnormal operating condition identification.

[0013] This invention feeds a time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and the thermal-hydraulic conservation equations, and uses an automatic differentiation mechanism to solve the physical equation residuals to constrain network training. This achieves an organic integration of physical laws and data-driven models, overcomes the shortcomings of insufficient interpretability of pure data-driven models and poor adaptability of pure physical models, and enhances the physical interpretability and generalization ability of anomaly feature extraction.

[0014] This invention constructs the spatiotemporal correlation features of unit-grid coupling by mapping the dynamic response characteristics of the unit to a heterogeneous graph neural network topology composed of the primary loop coolant pressure boundary node and the voltage phasor of the key bus of the power grid, and learning cross-scale dynamic correlation through spatiotemporal graph convolution operation. This overcomes the limitation of separating the analysis of equipment-level anomalies and system-level disturbances in the existing technology, and realizes the coupled perception of thermal-hydraulic anomalies of nuclear power plants and electromagnetic transient disturbances of the power grid.

[0015] This invention loads spatiotemporal correlation features into a variational autoencoder structure and infers the probability distribution of potential abnormal modes by maximizing the lower bound of evidence. It extracts a deep feature vector that reflects the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances, thereby realizing quantitative modeling of the latent space distribution of normal and abnormal operating conditions and improving the early identification capability of abnormal modes.

[0016] This invention constructs a dynamic event tree analysis framework using deep feature vectors as evidence variables and generates accident evolution branch paths through adaptive importance sampling. It also loads a set of candidate accident sequences into a Bayesian network inference engine based on variational inference to update the posterior distribution of equipment state failure probabilities. This enables real-time calculation of the probability of accident sequences occurring under abnormal operating conditions in nuclear power plants, solves the problems of offline probabilistic safety assessment and poor timeliness in existing technologies, and supports second-level risk quantification decision-making.

[0017] This invention calculates the Frosley-Vesely importance of each basic event based on the probability of occurrence of each accident sequence using a risk importance comprehensive ranking algorithm, and integrates the real-time calculated core damage frequency and the power grid transient voltage stability margin assessment value to construct a risk index set. This achieves a unified quantitative representation of nuclear safety risk and power grid stability risk, overcoming the inadequacy of a single risk index in reflecting the overall safety status of the system.

[0018] This invention utilizes a multi-agent deep deterministic policy gradient algorithm framework, which inputs a risk index set into the state space and configures the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. It also designs a reward function that minimizes the weighted entropy value of the risk index set and optimizes communication protocol priority and power adjustment strategies through multi-scenario game learning in a digital twin environment. This achieves intelligent collaborative decision-making among nuclear power plants, the power grid, and emergency response, solving the problem of fragmented multi-party dispatching and lack of collaborative optimization in existing technologies, and reducing communication latency and load loss under abnormal operating conditions.

[0019] This invention achieves highly reliable, low-latency transmission of control commands under extreme operating conditions by generating hierarchical communication command sequences and power adjustment suggestions, transmitting them through a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication, enabling ultra-reliable low-latency communication mode and configuring a dual-transmitter selective-receiver protection mechanism, employing orthogonal frequency division multiplexing multi-carrier modulation technology and achieving dual-channel sub-microsecond clock synchronization and seamless redundancy switching through a time-sensitive network gateway. It overcomes the vulnerability of single communication methods to natural disasters or network attacks, and ensures the reliability of coordinated control between the nuclear power plant reactor power regulation system and the grid automatic generation control device.

[0020] This invention collects the reactor power response curve and the grid frequency recovery trajectory after the command is executed and feeds them back to step one as input conditions for the dynamic correction of the abnormal threshold of the next round of monitoring data. This realizes the adaptive closed-loop adjustment of the abnormal detection threshold, solves the problem that fixed thresholds cannot adapt to the dynamics of unit aging, seasonal changes and grid structure evolution, and ensures the long-term effectiveness and accuracy of the diagnostic system.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the intelligent diagnostic method for abnormal operating conditions of nuclear power units as described in this embodiment of the invention; Figure 2 This is a schematic diagram of the intelligent diagnostic system for abnormal operating conditions of nuclear power units as described in this embodiment of the invention; Figure 3 This is a schematic diagram of the intelligent diagnostic device for abnormal operating conditions of nuclear power units as described in an embodiment of the present invention.

[0024] In the diagram: 701, Acquisition Module; 702, Association Module; 703, Construction Module; 704, Output Generation Module; 705, Diagnostic Module; 800, Intelligent Diagnostic Equipment for Abnormal Operating Conditions of Nuclear Power Units; 801, Processor; 802, Memory; 803, Multimedia Component; 804, I / O Interface; 805, Communication Component. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Example 1: This embodiment provides an intelligent diagnostic method for abnormal operating conditions of nuclear power units.

[0028] See Figure 1The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0029] S100, Step 1: Collect real-time monitoring data of the reactor coolant system pressure, temperature, neutron flux density, grid-side frequency deviation, and voltage fluctuation amplitude of the nuclear power unit. After spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data, construct a time-series dataset that integrates the unit's operating status and grid disturbance characteristics.

[0030] It is understood that step S100 includes S101, S102, and S103, wherein: S101. Piezoresistive pressure sensors and platinum resistance temperature sensors are installed in the hot and cold sections of the reactor coolant system of the nuclear power unit. A self-powered neutron detector is installed at the core outlet. A synchronous phasor measurement unit is installed on the high-voltage bus of the power grid. The original analog signals of pressure, temperature, neutron flux density, frequency deviation, and voltage fluctuation amplitude are collected at a sampling rate of 100 frames per second. After quantization and anti-aliasing filtering by a 16-bit analog-to-digital converter, the original digital sampling sequence is formed. It should be noted that, to address the sampling timing discrepancy caused by the heterogeneity of clock sources between the nuclear power unit's distributed control system and the power grid's wide-area measurement system, this step constructs a clock synchronization engine based on the reactor core cycle characteristics. This engine does not simply adopt a network time protocol, but deeply integrates the physical characteristics of the nuclear reactor: by real-time monitoring of the reactor cycle (i.e., the time required for a neutron density change of e times), it dynamically adjusts the synchronization priority. When the reactor cycle enters the range of 10 to 100 seconds, it is identified as a potential transient condition, and the synchronization priority of the power grid data is automatically increased to ensure the timeliness of the coupled analysis. The introduction of a cubic spline interpolation algorithm solves the problem of sampling rate differences in heterogeneous systems. This algorithm ensures the continuity of the first and second derivatives of the interpolation function, making it suitable for physical processes with limited rate of change, such as the thermal parameters of nuclear power units. Compared to linear interpolation, it better preserves the changing trends of coolant temperature and pressure. The peak detection of the neutron flux density signal adopts the sliding window Z-score method. The window length is taken as the least common multiple of the reactor core thermal response time constant and the power grid electromechanical oscillation period, ensuring that it can both capture transients introduced by reactivity and filter out harmonic interference from the power grid. The step detection threshold for grid frequency deviation needs to consider the dead-zone characteristics of the nuclear power unit speed control system. A 0.5 Hz threshold corresponds to the dead-zone boundary of the turbine governor's primary frequency regulation, preventing normal frequency regulation from triggering abnormal alarms. Kalman filtering state estimation compensation is designed for nuclear power unit sensor failure scenarios. By establishing a state-space model of coolant pressure, temperature, and neutron flux density, the gain matrix is ​​dynamically adjusted using observation residuals, maintaining data continuity even when a single sensor fails. The moving average smoothing window length is taken as the nuclear power unit load tracking cycle, preserving the normal power regulation trend while smoothing short-term grid disturbances. The constructed time-series dataset achieves time-series alignment between the slow-changing thermal processes of the nuclear power unit and the fast-changing electromagnetic processes of the grid, laying the foundation for subsequent cross-domain feature extraction.

[0031] S102. Input the original digital sampling sequence into the clock synchronization engine based on the core period characteristics, use the IEEE1588 precise time protocol to align the clock between the nuclear power unit's distributed control system and the power grid's wide-area measurement system, use the cubic spline interpolation algorithm to compensate for the sampling time differences of the heterogeneous systems, reconstruct the cross-system unified timestamp with a period of 10 milliseconds, and build a unified timestamp dataset. It should be noted that traditional graph neural networks are mostly used for homogeneous systems, while this step addresses the heterogeneous characteristics of the primary coolant pressure boundary nodes of nuclear power units and the voltage phasors of key power grid buses by designing a cross-domain mapping mechanism: the primary loop node features include coolant pressure, temperature, flow rate, and their first-order time derivatives, representing the thermal-hydraulic state; the power grid node features include voltage amplitude, phase angle, frequency, and their rate of change, representing the electromagnetic transient state; the thermal-hydraulic sub-block of the edge weight matrix is ​​calculated based on the coolant flow resistance coefficient and the network topology, the power grid sub-block is calculated based on the transmission line impedance and admittance matrix, and the cross-domain sub-block introduces the power-frequency static characteristic coefficient and the voltage-reactive power sensitivity coefficient to achieve explicit coupling between reactor power regulation and grid frequency response, and between turbine valve opening and voltage fluctuations. The spatiotemporal graph convolution operation uses one-dimensional causal convolution in the time dimension to ensure real-time performance, and uses spectral graph convolution in the spatial dimension to aggregate neighborhood node information. By stacking multiple layers of network, the dynamic transfer function of reactor power-frequency coupling characteristics and the cross-scale response law of turbine valve-voltage fluctuations are learned.

[0032] Furthermore, the encoder network employs a bidirectional long short-term memory (LSTM) network structure to capture long-range dependencies in time-series data, compressing high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The dimension of the latent variables is the sum of the number of independent safety parameters of nuclear power units and the number of key nodes in the power grid, ensuring a match between information bottlenecks and physical degrees of freedom. The decoder network uses a deconvolution structure to reconstruct the distribution of features under normal operating conditions, with the reconstruction error serving as a baseline for anomaly detection during the training phase. The process of maximizing the lower bound of evidence introduces a reparameterization technique, achieving randomness of latent variables through standard normal distribution sampling and learnable affine transformations, avoiding gradient vanishing caused by discrete sampling. The probability distribution inference of potential anomaly patterns uses Monte Carlo dropout to approximate the posterior, obtaining the predicted mean and variance through multiple samplings via forward propagation. When the variance exceeds a threshold, it is judged as an out-of-distribution anomaly, i.e., a potential accident mode that did not appear in the training data.

[0033] This step achieves the extraction of coupled features between thermal-hydraulic anomalies of nuclear power units and electromagnetic transient disturbances of the power grid by hard-constrained embedding of physical information neural networks and cross-domain topology construction of heterogeneous graph neural networks, overcoming the limitation of the separation between equipment-level and system-level analysis in traditional methods.

[0034] S103. Load the unified timestamp dataset into a data cleaning engine based on sliding window Z-score anomaly detection and Min-Max normalization. Set the neutron flux density signal peak detection threshold to 3 times the standard deviation and the grid frequency deviation step detection threshold to 0.5 Hz. After removing abnormal samples, use Kalman filtering to compensate for missing values ​​by state estimation. Use a 60-second sliding window for moving average smoothing to construct a time series dataset that integrates unit operating status and grid disturbance characteristics.

[0035] It should be noted that the Bayesian network inference engine integrates three types of nodes: equipment status, human factors, and environmental parameters. Equipment status nodes include the availability of reactor coolant pumps, the integrity of steam generator heat transfer tubes, and the response time of the turbine speed control system; their prior probabilities come from the equipment reliability database. Human factors nodes include operator diagnosis time, probability of operational errors, and deviations in procedure execution; their probability distributions come from full-range simulator experimental data of the nuclear power plant. Environmental parameter nodes include seismic intensity, extreme temperatures, and grid short-circuit capacity. When deep feature vectors are used as evidence input, they do not directly replace node probabilities but rather approximate the posterior distribution through variational inference: a variational distribution parameterized network is introduced, using deep feature vectors as input and output variational parameters. By minimizing the relative entropy between the variational distribution and the true posterior, the network parameters are optimized, achieving a smooth fusion of evidence and prior knowledge. Compared to traditional connection tree inference algorithms, this variational inference method can handle high-dimensional continuous evidence such as deep feature vectors, avoiding information loss caused by discretization.

[0036] The risk importance ranking algorithm uses the Frostley-Vesely importance index as its core indicator, which characterizes the necessity of a basic event for a top event, i.e., the probability increment of the top event occurring when the basic event occurs. The calculation process considers common-cause failures of the nuclear power unit and the power grid: when power grid failures and nuclear power unit equipment failures share a common cause, such as extreme weather, a common-cause failure coefficient is introduced to correct the conditional probability. Real-time calculation of core damage frequency employs a simplified thermal-hydraulic model coupled with a severe accident analysis program. Based on current coolant pressure, temperature, and neutron flux density, it rapidly estimates the fuel cladding temperature and core meltdown fraction, avoiding the computational burden of calling the full severe accident analysis program during operation. The power grid transient voltage stability margin assessment uses the extended equal-area criterion, calculating the difference between the critical cut-off time and the actual cut-off time based on the current power grid topology and generator power angle characteristics, quantifying the voltage collapse risk.

[0037] The technical effect of this step is that by fusing adaptive sampling of dynamic event trees with variational inference of Bayesian networks, real-time probability updates and risk importance quantification of the evolution path of abnormal operating conditions of nuclear power units are realized, supporting second-level risk assessment from diagnosis to decision-making.

[0038] S200, Step 2: Feed the time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and thermal-hydraulic conservation equations. Solve the physical equation residuals through an automatic differentiation mechanism to constrain network training, forming a dynamic response feature of the unit that integrates physical laws and data-driven approaches. Map the dynamic response feature of the unit to a heterogeneous graph neural network topology composed of the primary coolant pressure boundary nodes and the voltage phasors of the key power grid bus. Learn the cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response through spatiotemporal graph convolution operations, and construct the spatiotemporal correlation feature of unit-grid coupling. Load the spatiotemporal correlation feature into a variational autoencoder structure to model the latent space distribution of normal and abnormal operating conditions. Infer the probability distribution of potential abnormal modes by maximizing the lower bound of evidence, and extract a deep feature vector reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances.

[0039] It is understood that step S200 includes S201, S202, and S203, wherein: S201. The time-series dataset is fed into a physical information neural network architecture. This network embeds the reactor point reactor neutron dynamics equations and the thermal-hydraulic conservation equations. The residual loss function of the physical equations is calculated through an automatic differentiation mechanism, and the total loss function is minimized iteratively by the Adam optimizer, forming a dynamic response characteristic of the unit that integrates physical laws and data-driven approaches. The formula for calculating the residual loss function of the physical equations is as follows: In the formula, This represents the physical constraint loss value. Let be the neutron density, and t be the time. For reactivity, To delay the issuance of neutron share 0.0065, The neutron generation time is 0.0001 seconds, and λ is the physical constraint weight of 0.5.

[0040] It should be noted that the construction of the physical information neural network is not simply a matter of adding a regularization term to the loss function. Instead, it uses the point reactor neutron dynamics equation and the thermal-hydraulic conservation equation as hard constraints on the network output, forcing the neural network to learn a mapping relationship that conforms to physical laws. The point reactor neutron dynamics equation describes the evolution of neutron density over time, considering the coupling effect between instantaneous neutrons and six sets of delayed neutron precursor nuclei. Its rigidity stems from the huge timescale difference between the neutron generation time and the half-life of the delayed neutron precursor nuclei. The thermal-hydraulic conservation equation describes the coolant energy balance, considering the dynamic processes of core heat release, primary loop heat transfer, and secondary loop steam generation.

[0041] Understandably, traditional numerical differentiation uses finite difference approximations, which suffer from both truncation and rounding errors, especially for rigid problems like neutron density where the rate of change spans six orders of magnitude, making step size selection difficult. Automatic differentiation, by tracking the derivative rules of basic operations through a computational graph, transforms the gradient calculation of the physical equation residuals into chain multiplication of the Jacobian matrix, avoiding discretization errors. At the implementation level, the computational graph construction needs to consider the coupling nodes between the neutron dynamics equations and the thermal-hydraulic equations: changes in neutron density affect core power release, thus altering coolant temperature; changes in coolant temperature affect neutron density through feedback from the reactivity temperature coefficient, forming a closed-loop coupling. The application of the Adam optimizer requires a specially designed learning rate scheduling strategy, with physical constraint weights dynamically adjusted. Smaller weights are assigned in the early stages of training to ensure data fitting ability, while increased weights in later stages enhance physical consistency.

[0042] By embedding hard constraints into physical equations and using an automatic differentiation mechanism, a lossless fusion of data-driven models and physical laws is achieved, enhancing the physical interpretability and extrapolation capability of anomaly feature extraction.

[0043] S202. The dynamic response characteristics of the unit are mapped to the set of primary loop coolant pressure boundary nodes and the set of grid key bus voltage phasors. A heterogeneous graph neural network topology is constructed, in which the node feature matrix contains the reactor power level and grid frequency deviation, and the edge weight matrix represents the thermal-hydraulic impedance and transmission line admittance. The cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response is learned through spatiotemporal graph convolution operation, and the spatiotemporal correlation characteristics of unit-grid coupling are constructed. It should be noted that traditional graph neural networks assume isomorphic nodes and simple edge relationships, making them unable to handle coupled systems like nuclear power units and power grids, which have vastly different physical mechanisms and time scales. The node set design of the heterogeneous graph neural network topology reflects the essential physical differences: the primary coolant pressure boundary nodes correspond to key locations such as the reactor pressure vessel, the primary side of the steam generator, and the inlet and outlet of the main pump, and their feature vectors include pressure, temperature, flow rate, and their rate of change, reflecting the thermal-hydraulic state; the voltage phasor nodes of the key power grid bus correspond to electrical nodes such as the high-voltage side of the nuclear power plant's step-up transformer, switchyard, and transmission lines, and their feature vectors include voltage amplitude, phase angle, frequency, and their rate of change, reflecting the electromagnetic transient state.

[0044] The construction of the edge weight matrix is ​​the core innovation of this step. The thermal-hydraulic sub-block is calculated based on a one-dimensional flow network model, considering the coolant flow resistance coefficient, local resistance coefficient, and pipe network topology. The edge weight is directly proportional to the square of the flow rate and inversely proportional to the fifth power of the pipe diameter, reflecting the similarity law of fluid mechanics. The power grid sub-block is calculated based on the AC power flow equation, considering the transmission line resistance, reactance, and ground admittance. The edge weight is inversely proportional to the line impedance, reflecting Ohm's law and Kirchhoff's law. The design of the cross-domain sub-block reflects the coupling mechanism between the nuclear power unit and the power grid. The power-frequency static characteristic coefficient characterizes the primary frequency regulation capability of the turbine speed regulation system, reflecting the steady-state relationship between reactor power changes and grid frequency deviation. The voltage-reactive power sensitivity coefficient characterizes the voltage regulation capability of the generator excitation system, reflecting the dynamic relationship between turbine valve opening changes and grid voltage fluctuations. The spatiotemporal graph convolution operation employs one-dimensional causal convolution in the time dimension to ensure real-time performance and avoid future information leakage. The kernel length is taken as the least common multiple of the thermal response time constant of the nuclear power unit and the electromechanical oscillation period of the power grid, covering dynamics across multiple time scales from seconds to minutes. In the spatial dimension, spectral graph convolution is used, and frequency domain filtering is achieved through eigenvalue decomposition of the Laplacian matrix. Low-frequency components correspond to slow, spatially smooth trends, while high-frequency components correspond to rapid fluctuations of local anomalies. Multi-layer network stacking learns higher-order neighborhood relationships. The first layer captures the thermal-hydraulic transfer or electrical connection relationships of directly connected nodes, the second layer captures the coupling effect of indirect paths, and the deep network abstracts the dynamic transfer function of reactor power-frequency coupling characteristics and the cross-scale response law of turbine valve-voltage fluctuations. Through heterogeneous node mapping and cross-domain edge weight design, a unified graph topology representation of the nuclear power unit's thermal-hydraulic system and the power grid's electrical system is realized, learning the cross-scale dynamic correlation law between equipment-level anomalies and system-level disturbances.

[0045] S203. The spatiotemporal correlation features are loaded into the variational autoencoder structure. The encoder network compresses the high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The decoder network reconstructs the distribution of normal operating conditions features. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence. The deep feature vector reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances is extracted.

[0046] It should be noted that the introduction of variational autoencoders achieves a unification of dimensionality reduction and generative modeling: the encoder network compresses high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space, while outputting the mean and variance parameters of the latent variables to characterize the uncertainty in the dimensionality reduction process; the decoder network samples and reconstructs the original features from the latent variables, and the reconstruction error measures the model's ability to characterize normal operating conditions. The encoder network adopts a bidirectional long short-term memory network structure, with the forward layer capturing the causal dependencies of time-series data and the backward layer capturing future contextual information. The latent states are concatenated and input into the fully connected layer to output the latent variable parameters. The selection of the latent variable dimension requires a trade-off between information compression and physical degrees of freedom: too low a dimension leads to the loss of reactor-specific safety parameters and information on key grid nodes, while too high a dimension results in a loss of sensitivity to anomaly detection. Typically, the dimension is the sum of the number of independent safety parameters of nuclear power units and the number of key grid nodes. The decoder network adopts a deconvolution structure, with the latent variables gradually expanded to the original feature dimension through upsampling and convolution operations. The activation function is chosen as hyperbolic tangent to match the normalized feature range.

[0047] The process of maximizing the lower bound of evidence introduces a reparameterization technique to solve the gradient backpropagation problem in random sampling: the randomness of latent variables is achieved through sampling from a standard normal distribution and learnable affine transformations. The sampling results are differentiable with respect to network parameters, supporting end-to-end training. The probability distribution inference of anomalous patterns uses Monte Carlo dropout to approximate a Bayesian posterior: during the testing phase, dropout is kept on, and multiple forward propagations are performed to obtain the mean and variance of the predicted distribution. When the variance exceeds a threshold, it is judged as an out-of-distribution anomaly, i.e., a potential accident mode not covered by the training data. This uncertainty quantification is crucial for nuclear power units. Traditional deterministic models output a single predicted value and cannot distinguish between cognitive uncertainty (the model's ignorance of unknown operating conditions) and accidental uncertainty (the inherent randomness of the system). The variational autoencoder, however, explicitly quantifies cognitive uncertainty through the posterior distribution of latent variables, issuing early warnings for new anomalous patterns.

[0048] S300, Step 3: Using deep feature vectors as evidence variables, construct a dynamic event tree analysis framework with reactor protection system triggering, turbine digital electro-hydraulic control system fast shutdown action, and grid low-frequency low-pressure load shedding device activation as the top events. Generate accident evolution branch paths through adaptive importance sampling to form a candidate accident sequence set. Load the candidate accident sequence set into a Bayesian network inference engine based on variational inference, update the posterior distribution of equipment state failure probability with deep feature vectors, and calculate the occurrence probability of each accident sequence. Based on the occurrence probability of each accident sequence, calculate the Frosley-Vesely importance of each basic event using a risk importance comprehensive ranking algorithm. Integrate the real-time calculated core damage frequency and grid transient voltage stability margin assessment value to construct a risk index set including instantaneous core damage frequency, probability of large-scale power outage conditions, and comprehensive risk importance.

[0049] It is understood that step S300 includes S301, S302, and S303, wherein: S301. Using deep feature vectors as evidence variables, a dynamic event tree analysis framework is constructed. The reactor protection system triggering, the turbine digital electro-hydraulic control system fast shutdown action, and the grid low-frequency low-pressure load shedding device start-up are set as top events. An adaptive importance sampling algorithm is adopted to adjust the branch sampling weights according to the probability of abnormal modes in the deep feature vectors. The accident evolution branch path is generated through sequence Monte Carlo simulation to form a candidate accident sequence set. It should be noted that the introduction of the adaptive importance sampling algorithm solves the problem of low computational efficiency in the traditional Monte Carlo method. The sampling weight adjustment mechanism deeply integrates the deep feature vector output from step two: when the deep feature vector indicates an abnormal increase in the reactor power-frequency coupling coefficient, i.e., the neutron density change rate shows a strong correlation with the grid frequency deviation, the branch that causes a sudden drop in grid frequency leading to an emergency reactor shutdown is sampled first; when it indicates an abnormal turbine valve-voltage sensitivity, i.e., the valve opening change shows a superlinear relationship with grid voltage fluctuations, the branch that causes a grid voltage collapse leading to turbine protection action is sampled first. This adaptive mechanism concentrates limited computational resources on high-risk scenarios, avoiding the ineffective consumption of uniform sampling on low-probability branches.

[0050] The implementation of sequential Monte Carlo simulation requires the construction of a state transition model: each branch node corresponds to a change in system state, including the discretization level of physical quantities such as reactor power level, coolant temperature and pressure, and grid frequency and voltage, as well as the triggering conditions for operator actions and automatic protection actions. The state transition probability is jointly determined by equipment reliability data, human factor reliability data, and the dynamic model of the physical process. In practical engineering applications, the innovation of this framework lies in its real-time adaptability to operating conditions. Traditional probabilistic safety assessments are based on offline calculations under design baseline conditions and cannot reflect parameter drift caused by unit aging, fuel cycle, and seasonal changes. The dynamic event tree, through real-time input of deep feature vectors, embeds the probability of abnormal patterns of the current operating condition into branch sampling, realizing a paradigm shift from offline static assessment to online dynamic assessment.

[0051] The technical benefits of this small step are as follows: through adaptive importance sampling and sequential Monte Carlo simulation, the evolution path of abnormal operating conditions of nuclear power units is generated in real time and the candidate accident sequence set is dynamically constructed, which significantly improves the timeliness and pertinence of risk assessment.

[0052] S302. Load the candidate accident sequence set into a Bayesian network inference engine based on variational inference, construct a directed acyclic graph with nodes representing device status, human actions, and environmental parameters, update the conditional probability table using deep feature vectors as evidence input, and calculate the occurrence probability of each accident sequence. The formula for calculating the probability of an accident sequence is as follows: In the formula, The probability of the updated accident sequence occurring. For prior probability, is the score of the first principal component after dimensionality reduction by principal component analysis of the deep feature vector, and k is the evidence strength coefficient; It should be noted that the construction of the Bayesian network topology reflects the causal relationships of abnormal operating conditions in nuclear power units: equipment status nodes include the availability status of reactor coolant pumps, the integrity of steam generator heat transfer tubes, and the response time of the turbine speed control system, etc., whose prior probabilities come from the equipment reliability database and online condition monitoring; human factor action nodes include operator diagnosis time, probability of operational errors, and deviations in procedure execution, etc., whose probability distributions come from experimental data from the full-range simulator of the nuclear power plant and feedback from operational experience; environmental parameter nodes include seismic intensity, extreme temperatures, and grid short-circuit capacity, etc., serving as external condition inputs. The directed edges between nodes represent the direction of causal influence, such as equipment failure leading to system function degradation, system function degradation triggering operator intervention, and operator errors exacerbating the evolution of the accident.

[0053] When deep feature vectors are used as evidence input, they are not simply replaced with node probability tables, but rather smoothly fused through a variational network. The input to the variational network is the low-dimensional representation of the deep feature vectors after dimensionality reduction via principal component analysis (PCA), and the output is the variational parameters of the probability distribution of each node. The purpose of PCA dimensionality reduction is to eliminate redundant information in the deep feature vectors while retaining the principal component scores that reflect the essence of the anomalous patterns. The setting of the evidence strength coefficient reflects the adjustment of the confidence level of the deep feature vectors: when the coefficient is large, the deep feature vectors have a significant corrective effect on the prior probabilities, which is suitable for situations where the anomalous patterns are clear; when the coefficient is small, the dominant role of the prior probabilities is retained, which is suitable for situations where the anomalous patterns are ambiguous. The Sigmoid function maps the linear combination to the probability interval, ensuring that the updated probability values ​​are physically reasonable.

[0054] During nuclear power plant operation, evidence sources include sensor measurements, operator observations, diagnostic system outputs, and external notifications. Traditional methods struggle to quantify the conflicts and uncertainties among different pieces of evidence. The variational inference framework explicitly models the reliability differences of evidence through evidence weight coefficients and variational distribution variance, achieving adaptive fusion of multi-source evidence. Furthermore, this framework supports online learning: new accident sequence data are continuously input, variational network parameters are updated via stochastic gradient descent, and the probability table of the Bayesian network optimizes with accumulated operational experience, forming a continuous evolutionary mechanism for risk perception. By fusing variational inference approximate reasoning with deep feature vector evidence, real-time updates and calculations of the probability of accident sequences are achieved, overcoming the discretization loss and slow convergence problems of traditional methods.

[0055] S303. Based on the occurrence probability of each accident sequence, the Fosaili-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The real-time calculated core damage frequency and the power grid transient voltage stability margin assessment value are then integrated to construct a risk index set.

[0056] It should be noted that the Frosley-Vesely importance score characterizes the necessity of a basic event for a top event, i.e., the conditional probability increment of the top event when the basic event occurs. This index differs from risk reduction and risk achievement scores, directly reflecting the criticality of the basic event in the accident evolution, and is suitable for identifying priority targets in resource-constrained scenarios. The calculation process needs to consider the common causes of failures in nuclear power units and the power grid: when power grid failures and nuclear power plant equipment failures share common causes such as extreme weather, a common cause failure coefficient is introduced to correct the conditional probability, avoiding risk underestimation due to the independence assumption.

[0057] The real-time calculation of core damage frequency employs a coupling strategy between a simplified thermal-hydraulic model and a severe accident analysis program. Since calling the full severe accident analysis program during operation is computationally burdensome, this step uses a neural network-based alternative model: taking the current coolant pressure, temperature, and neutron flux density as inputs, it quickly estimates the peak fuel cladding temperature and core meltdown fraction, outputting an approximate value for the core damage frequency. This alternative model is trained on numerous offline severe accident analysis program examples, ensuring approximate accuracy while achieving millisecond-level response. The grid transient voltage stability margin assessment uses the extended equal-area criterion. Based on the current grid topology and generator power angle characteristics, it calculates the difference between the critical cut-off time and the actual cut-off time, quantifying the voltage collapse risk. A positive margin value indicates system stability; a smaller value indicates a thinner stability margin; and a negative value indicates that voltage instability is inevitable.

[0058] Furthermore, the fusion of risk indicator sets employs a weighted Euclidean norm, with weight coefficients reflecting the priority of nuclear safety versus grid safety. As a baseload power source, the safe shutdown of nuclear power units has a significant impact on the grid; therefore, core damage frequency is given the highest weight. Grid stability margin is the next highest, characterizing the lateral propagation risk of abnormal operating conditions. Risk importance has the lowest weight, used to identify critical weaknesses. Normalization of the indicator set ensures the comparability of risks with different dimensions. The final output risk indicator set consists of scalar values ​​between 0 and 1; higher values ​​indicate higher overall risk, facilitating the state space input of subsequent multi-agent decision-making systems.

[0059] By ranking risks by importance and integrating risks from multiple dimensions, a comprehensive risk index set with a unified dimension was constructed, realizing the quantitative integration of nuclear safety risks and power grid stability risks, and providing a decision-making basis for subsequent intelligent communication scheduling.

[0060] S400, Step 4: Input the risk index set as the state space into the multi-agent deep deterministic policy gradient algorithm framework. Configure the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. Design a reward function with the goal of minimizing the weighted entropy value of the risk index set. Optimize the communication protocol priority and power adjustment strategy under different risk levels through multi-scenario game learning in a digital twin environment. Output automatic generation control and regulation commands for the power grid dispatch center, reactor power regulation rod position suggestions for the nuclear power plant shift supervisor, and load transfer paths for the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions.

[0061] It is understood that in this step, S400 includes S401, S402, and S403, wherein: S401. A multi-agent deep deterministic policy gradient algorithm framework is established, which uses the risk index set as the input to the state space. The power grid dispatch center, the nuclear power plant shift leader, and the emergency command center are configured as agents with independent local observations. A reward function is designed with the goal of minimizing the weighted entropy value of the risk index set. It should be noted that the centralized training and decentralized execution mechanism adopted in this step combines the advantages of both: during the training phase, the global value function gathers the observation and action information of all agents to learn the optimal collaborative strategy; during the execution phase, each agent makes independent decisions based only on local observations, reducing communication dependence and computational latency. The power grid dispatch center agent has global information on the power grid side, such as the entire network topology, load distribution, and renewable energy output. Its local observations are the power grid stability margin component of the risk indicators, and its action space is the automatic generation control adjustment quantity, i.e., the adjustment command of the nuclear power unit power setpoint. The nuclear power plant shift supervisor agent has plant-side information such as the reactor physical state, equipment availability, and operator configuration. Its local observations are the core damage frequency component of the risk indicators, and its action space is the control rod reactivity adjustment quantity and the operation sequence of auxiliary equipment. The emergency command center agent has external information such as the distribution of emergency resources, load priority, and social impact assessment. Its local observations are the risk importance component of the risk indicators, and its action space is the load transfer path and the emergency power supply startup sequence. The action spaces of the three intelligent agents are coupled: the power regulation commands of the power grid dispatch center change the operating conditions of the nuclear power plant; the equipment operation of the nuclear power plant's long-term operating hours affects the equivalent impedance of the power grid; and the load shedding of the emergency command center changes the power balance of the power grid.

[0062] Weighted processing reflects the priority differences of different risk sources, with the entropy value of core damage frequency having the highest weight, followed by grid stability margin, and then risk importance having the lowest weight. The action amplitude penalty term prevents system oscillations caused by over-regulation; automatic generation control adjustments, control rod reactive adjustments, and load shedding are multiplied by their respective penalty coefficients. The load loss penalty term quantifies the socio-economic cost of power supply assurance and is proportional to the capacity and importance level of the shedding load. The communication latency penalty term strengthens real-time response capabilities; the longer the latency, the heavier the penalty, prompting the agent to choose low-latency communication protocols and decision paths.

[0063] S402. Communication protocol priority and power adjustment strategies are optimized through multi-scenario game learning in a digital twin environment. The digital twin environment includes a reactor physics and thermal model, a turbine speed regulation model, and a power grid transient stability model. Three training scenarios are set: normal operating condition, design baseline accident condition, and above-design baseline accident condition. A centralized training and decentralized execution mechanism is adopted. Each agent generates actions independently through a policy network. After evaluation by a global value function, the network parameters are updated through policy gradient ascent to optimize the communication protocol priority and power adjustment strategies under different risk levels. It should be noted that traditional graph neural networks are mostly used for homogeneous systems, while this step addresses the heterogeneous characteristics of the primary coolant pressure boundary nodes of nuclear power units and the voltage phasors of key power grid buses by designing a cross-domain mapping mechanism: the primary loop node features include coolant pressure, temperature, flow rate, and their first-order time derivatives, representing the thermal-hydraulic state; the power grid node features include voltage amplitude, phase angle, frequency, and their rate of change, representing the electromagnetic transient state; the thermal-hydraulic sub-block of the edge weight matrix is ​​calculated based on the coolant flow resistance coefficient and the network topology, the power grid sub-block is calculated based on the transmission line impedance and admittance matrix, and the cross-domain sub-block introduces the power-frequency static characteristic coefficient and the voltage-reactive power sensitivity coefficient to achieve explicit coupling between reactor power regulation and grid frequency response, and between turbine valve opening and voltage fluctuations. The spatiotemporal graph convolution operation uses one-dimensional causal convolution in the time dimension to ensure real-time performance, and uses spectral graph convolution in the spatial dimension to aggregate neighborhood node information. By stacking multiple layers of network, the dynamic transfer function of reactor power-frequency coupling characteristics and the cross-scale response law of turbine valve-voltage fluctuations are learned.

[0064] The loading of the variational autoencoder structure enables generative modeling of normal and abnormal operating conditions. The encoder network employs a bidirectional long short-term memory network structure to capture long-range dependencies in time-series data, compressing high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The dimension of the latent variables is the sum of the number of independent safety parameters of the nuclear power unit and the number of key nodes in the power grid, ensuring a match between information bottlenecks and physical degrees of freedom. The decoder network uses a deconvolution structure to reconstruct the distribution of features under normal operating conditions, with the reconstruction error serving as the baseline for anomaly detection during the training phase. The process of maximizing the lower bound of evidence introduces a reparameterization technique, achieving randomness of latent variables through standard normal distribution sampling and learnable affine transformations, avoiding gradient vanishing caused by discrete sampling. The probability distribution inference of potential abnormal modes uses Monte Carlo dropout to approximate the posterior, obtaining the predicted mean and variance through multiple samplings via forward propagation. When the variance exceeds a threshold, it is judged as an out-of-distribution anomaly, i.e., a potential accident mode that did not appear in the training data.

[0065] S403 outputs automatic generation control and regulation commands to the power grid dispatch center, reactor power regulation rod position suggestions to the nuclear power plant shift leader, and load transfer paths to the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions.

[0066] It should be noted that the selection of the top event in the dynamic event tree analysis framework reflects the coupling risk sources between the nuclear power unit and the power grid: the reactor protection system trigger corresponds to the last line of defense in depth for nuclear safety; the fast shutdown action of the turbine digital electro-hydraulic control system corresponds to the mechanical-electrical interface between the nuclear power plant and the power grid; and the activation of the power grid low-frequency and low-pressure load shedding device corresponds to the reverse propagation of a power grid fault to the nuclear power unit. Unlike traditional static event trees, the branch generation of the dynamic event tree uses adaptive importance sampling, and the sampling weights are adjusted in real time according to the depth feature vector output in step two: when the depth feature vector indicates an abnormal increase in the reactor power-frequency coupling coefficient, the branch that causes a sudden drop in power grid frequency leading to an emergency reactor shutdown is sampled first; when it indicates an abnormal turbine control valve-voltage sensitivity, the branch that causes a power grid voltage collapse leading to turbine protection action is sampled first. This adaptive mechanism concentrates computational resources on high-risk scenarios, avoiding the low efficiency problem of uniform sampling in traditional Monte Carlo methods.

[0067] The equipment status nodes include the availability of reactor coolant pumps, the integrity of steam generator heat transfer tubes, and the response time of the turbine speed control system, with prior probabilities derived from the equipment reliability database. Human factor action nodes include operator diagnosis time, probability of operational errors, and deviations in procedure execution, with probability distributions derived from full-range simulator experimental data of the nuclear power plant. Environmental parameter nodes include seismic intensity, extreme temperatures, and grid short-circuit capacity. When deep feature vectors are used as evidence input, they do not directly replace node probabilities but rather approximate the posterior distribution through variational inference: a variational distribution parameterized network is introduced, using deep feature vectors as input and output variational parameters. By minimizing the relative entropy between the variational distribution and the true posterior, the network parameters are optimized, achieving a smooth fusion of evidence and prior knowledge. Compared to traditional connection tree inference algorithms, this variational inference method can handle high-dimensional continuous evidence such as deep feature vectors, avoiding information loss caused by discretization.

[0068] The risk importance ranking algorithm uses the Frostley-Vesely importance index as its core indicator, which characterizes the necessity of a basic event for a top event, i.e., the probability increment of the top event occurring when the basic event occurs. The calculation process considers common-cause failures of the nuclear power unit and the power grid: when power grid failures and nuclear power unit equipment failures share a common cause, such as extreme weather, a common-cause failure coefficient is introduced to correct the conditional probability. Real-time calculation of core damage frequency employs a simplified thermal-hydraulic model coupled with a severe accident analysis program. Based on current coolant pressure, temperature, and neutron flux density, it rapidly estimates the fuel cladding temperature and core meltdown fraction, avoiding the computational burden of calling the full severe accident analysis program during operation. The power grid transient voltage stability margin assessment uses the extended equal-area criterion, calculating the difference between the critical cut-off time and the actual cut-off time based on the current power grid topology and generator power angle characteristics, quantifying the voltage collapse risk.

[0069] S500, Step 5: Encode the hierarchical communication command sequence and power adjustment suggestions into general object-oriented substation event messages and manufacturing message specification messages that conform to substation communication standards. Transmit these messages through a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. Enable ultra-reliable low-latency communication mode and configure a dual-transmitter selective reception protection mechanism in the fifth-generation mobile communication network slice. Use orthogonal frequency division multiplexing multi-carrier modulation technology in the power line carrier channel. Achieve dual-channel sub-microsecond clock synchronization and seamless redundancy switching through a time-sensitive network gateway. Send the collaborative control commands to the nuclear power unit reactor power regulation system and the grid automatic generation control device respectively. Collect the reactor power response curve and grid frequency recovery trajectory after command execution and feed them back to Step 1 as input conditions for the next round of dynamic correction of abnormal monitoring data thresholds.

[0070] It is understood that in this step, S500 includes S501, S502, and S503, wherein: S501. The hierarchical communication command sequence and power adjustment suggestions are encoded into general object-oriented substation event messages and manufacturing message specification messages conforming to the IEC 61850 standard. The general object-oriented substation event message dataset contains power adjustment commands, rod position suggestions, and load transfer path identifiers. The manufacturing message specification message contains numerical encoding of risk index sets. Logical node data is mapped into binary bit streams using the ASN.1 basic encoding rules to form the encoded communication messages. S502 uses a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. In the fifth-generation mobile communication network slice, an ultra-reliable low-latency communication mode is enabled with a 1-millisecond air interface latency and 99.999% reliability. A dual-transmit and selective-receive protection mechanism is adopted for redundant deployment on the core network side. In the power line carrier channel, orthogonal frequency division multiplexing multi-carrier modulation technology is used with a subcarrier spacing of 15 kHz. The modulation mode is adaptively selected from orthogonal phase shift keying to 256 orthogonal amplitude modulation. Bit loading is dynamically adjusted according to channel state information to form a heterogeneous redundant transmission channel. S503. Through a time-sensitive network gateway, dual-channel sub-microsecond clock synchronization and seamless redundancy switching are achieved. Cooperative control commands are sent to the nuclear power unit reactor power regulation system and the grid automatic generation control device, respectively. The reactor power response curve and grid frequency recovery trajectory after the command execution are collected, the dynamic correction coefficient is calculated, and the corrected abnormal threshold is fed back to step one as the input condition for the next round of dynamic correction of abnormal threshold of monitoring data.

[0071] It should be noted that the dynamic feedback correction formula for the threshold is as follows: In the formula, The anomaly detection threshold after the (k+1)th round of correction is fed back to step one as the input for the next round of monitoring. Let γ be the original threshold for the k-th round, and γ be the correction rate of 0.1. This is the actual power response value. The desired power response value, This is the rated power.

[0072] It should be noted that the use of a time-sensitive network gateway to achieve dual-channel sub-microsecond clock synchronization ensures precise timing coordination and data consistency between the nuclear power unit and the grid control device; the seamless redundancy switching mechanism eliminates the risk of single-point failure and ensures the continuous and reliable issuance of control commands; the synchronous distribution of coordinated control commands to the reactor power regulation system and the grid AGC device realizes coordinated and optimized control of power generation, grid, load and storage; the real-time acquisition of reactor power response curves and grid frequency recovery trajectories, the calculation of dynamic correction coefficients and the feedback to the anomaly threshold setting stage form a closed-loop adaptive control mechanism, which enables the monitoring threshold to be dynamically adjusted according to the system operating status, significantly improving the accuracy of anomaly detection and the adaptability of the control strategy.

[0073] Example 2: like Figure 2 As shown, this embodiment provides an intelligent diagnostic system for abnormal operating conditions of nuclear power units. See [link to documentation]. Figure 2 The system includes: Step 1, Acquisition Module 701: Used to collect real-time monitoring data of the reactor coolant system pressure, temperature, neutron flux density, and grid-side frequency deviation and voltage fluctuation amplitude of the nuclear power unit. After spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data, a time-series dataset integrating the unit's operating status and grid disturbance characteristics is constructed. Step 2, Module 702: This module feeds the time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and thermal-hydraulic conservation equations. The residuals of the physical equations are solved using an automatic differentiation mechanism to constrain network training, forming dynamic response characteristics of the unit that integrate physical laws and data-driven approaches. These dynamic response characteristics are mapped to a heterogeneous graph neural network topology composed of primary coolant pressure boundary nodes and grid key bus voltage phasors. Spatiotemporal graph convolution operations are used to learn the cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response, constructing spatiotemporal correlation characteristics of the unit-grid coupling. These spatiotemporal correlation characteristics are then loaded into a variational autoencoder structure to model the latent space distribution of normal and abnormal operating conditions. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence, and a deep feature vector reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances is extracted. Step 3, Module 703: This module is used to construct a dynamic event tree analysis framework with deep feature vectors as evidence variables, taking reactor protection system triggering, turbine digital electro-hydraulic control system fast shutdown action, and grid low-frequency low-pressure load shedding device activation as top events. It generates accident evolution branch paths through adaptive importance sampling, forming a candidate accident sequence set. The candidate accident sequence set is loaded into a Bayesian network inference engine based on variational inference, and the posterior distribution of equipment state failure probability is updated using deep feature vectors to calculate the probability of occurrence of each accident sequence. Based on the probability of occurrence of each accident sequence, the Frosley-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. This is combined with the real-time calculated core damage frequency and the grid transient voltage stability margin assessment value to construct a risk index set including instantaneous core damage frequency values, large-scale power outage condition probabilities, and comprehensive risk importance. Step 4, Output Generation Module 704: This module uses the risk index set as input to the state space of a multi-agent deep deterministic policy gradient algorithm framework. It configures the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. It designs a reward function with the goal of minimizing the weighted entropy value of the risk index set. Through multi-scenario game learning in a digital twin environment, it optimizes the communication protocol priority and power adjustment strategy under different risk levels. It outputs automatic generation control and regulation commands for the power grid dispatch center, reactor power regulation rod position suggestions for the nuclear power plant shift supervisor, and load transfer paths for the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions. Step 5, Diagnostic Module 705: This module encodes the hierarchical communication command sequence and power adjustment suggestions into general object-oriented substation event messages and manufacturing message specification messages that conform to substation communication standards. These messages are transmitted via a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. An ultra-reliable low-latency communication mode is enabled in the fifth-generation mobile communication network slice, and a dual-transmitter selective reception protection mechanism is configured. Orthogonal frequency division multiplexing multi-carrier modulation technology is used in the power line carrier channel. Sub-microsecond clock synchronization and seamless redundancy switching between the two channels are achieved through a time-sensitive network gateway. Cooperative control commands are then sent to the nuclear power unit reactor power regulation system and the grid automatic generation control device, respectively. The reactor power response curve and grid frequency recovery trajectory after command execution are collected and fed back to Step 1 as input conditions for the next round of dynamic correction of abnormal monitoring data thresholds.

[0074] Specifically, the acquisition module 701 includes: The first processing unit is used to install piezoresistive pressure sensors and platinum resistance temperature sensors in the hot and cold sections of the reactor coolant system of the nuclear power unit, install a self-powered neutron detector at the core outlet, and install a synchronous phasor measurement unit on the high-voltage bus of the power grid side. It collects the original analog signals of pressure, temperature, neutron flux density, frequency deviation, and voltage fluctuation amplitude at a sampling rate of 100 frames per second. After quantization and anti-aliasing filtering by a 16-bit analog-to-digital converter, the original digital sampling sequence is formed. The first building unit is used to input the original digital sampling sequence into the clock synchronization engine based on the core period characteristics. It uses the IEEE 1588 precise time protocol to align the clock between the nuclear power unit's distributed control system and the power grid's wide-area measurement system. It uses a cubic spline interpolation algorithm to compensate for the sampling time differences of heterogeneous systems and reconstructs a unified timestamp across systems with a period of 10 milliseconds to build a unified timestamp dataset. The second building block is used to load the unified timestamp dataset into a data cleaning engine based on sliding window Z-score anomaly detection and Min-Max normalization. The threshold for neutron flux density signal spike detection is set to 3 times the standard deviation, and the threshold for grid frequency deviation step detection is set to 0.5 Hz. After removing outlier samples, Kalman filtering is used to compensate for missing values ​​through state estimation. A 60-second sliding window is used for moving average smoothing to build a time series dataset that integrates unit operating status and grid disturbance characteristics.

[0075] Specifically, the association module 702 includes: Iterative Unit: Used to feed time-series datasets into a physical information neural network architecture. This network embeds the reactor point reactor neutron dynamics equations and the thermo-hydraulic conservation equations. It calculates the residual loss function of the physical equations through an automatic differentiation mechanism, and then iteratively minimizes the total loss function using the Adam optimizer, forming a dynamic response characteristic of the unit that integrates physical laws and data-driven approaches. The formula for calculating the residual loss function of the physical equations is as follows: In the formula, This represents the physical constraint loss value. Let be the neutron density, and t be the time. For reactivity, To delay the issuance of neutron share 0.0065, The neutron generation time is 0.0001 seconds, and λ is the physical constraint weight of 0.5.

[0076] The third building unit is used to map the dynamic response characteristics of the unit to the set of primary coolant pressure boundary nodes and the set of voltage phasors of the key bus of the power grid, and to build a heterogeneous graph neural network topology. The node feature matrix contains the reactor power level and the frequency deviation of the power grid, and the edge weight matrix represents the thermal-hydraulic impedance and the transmission line admittance. Through spatiotemporal graph convolution operation, the cross-scale dynamic correlation between the reactor power-frequency coupling characteristics and the turbine valve opening-voltage fluctuation response is learned, and the spatiotemporal correlation characteristics of the unit-grid coupling are constructed. Extraction Unit: Used to load spatiotemporal correlation features into a variational autoencoder structure. The encoder network compresses high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The decoder network reconstructs the distribution of normal operating condition features, infers the probability distribution of potential abnormal modes by maximizing the lower bound of evidence, and extracts deep feature vectors that reflect the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances.

[0077] Specifically, the construction module 703 includes: The fourth building unit is used to construct a dynamic event tree analysis framework using deep feature vectors as evidence variables. It sets the reactor protection system trigger, the turbine digital electro-hydraulic control system fast shutdown action, and the grid low-frequency low-pressure load shedding device as the top events. It adopts an adaptive importance sampling algorithm to adjust the branch sampling weights according to the probability of abnormal modes in the deep feature vectors. The accident evolution branch path is generated through sequence Monte Carlo simulation to form a set of candidate accident sequences. The fifth building unit is used to load the candidate accident sequence set into the variational inference-based Bayesian network inference engine, construct a directed acyclic graph with nodes representing device status, human actions, and environmental parameters, update the conditional probability table using deep feature vectors as evidence input, and calculate the occurrence probability of each accident sequence. The formula for calculating the probability of an accident sequence is as follows: In the formula, The probability of the updated accident sequence occurring. For prior probability, is the score of the first principal component after dimensionality reduction by principal component analysis of the deep feature vector, and k is the evidence strength coefficient; Fusion Unit: Based on the occurrence probability of each accident sequence, the Fuseli-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The unit then merges the real-time calculated core damage frequency with the power grid transient voltage stability margin assessment value to construct a risk index set.

[0078] Specifically, the output generation module 704 includes: Input Unit: Used to input the risk index set as the state space into the multi-agent deep deterministic policy gradient algorithm framework, configuring the power grid dispatch center, nuclear power plant shift leader and emergency command center as agents with independent local observations, and designing a reward function with the goal of minimizing the weighted entropy value of the risk index set; Evaluation Strategy Unit: Used to optimize communication protocol priority and power adjustment strategies through multi-scenario game learning in a digital twin environment. The digital twin environment includes a reactor physics and thermal model, a turbine speed regulation model, and a power grid transient stability model. Three training scenarios are set: normal operating condition, design baseline accident condition, and above-design baseline accident condition. A centralized training and decentralized execution mechanism is adopted. Each agent generates actions independently through the policy network. The network parameters are updated through policy gradient ascent after evaluation by a global value function to optimize the communication protocol priority and power adjustment strategies under different risk levels. Generation Unit: Used to output automatic generation control and regulation commands to the power grid dispatch center, reactor power regulation rod position suggestions to the nuclear power plant operating shift leader, and load transfer paths to the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions.

[0079] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0080] Example 3: Corresponding to the above method embodiments, this embodiment also provides an intelligent diagnostic device for abnormal operating conditions of nuclear power units. The intelligent diagnostic device for abnormal operating conditions of nuclear power units described below and the intelligent diagnostic method for abnormal operating conditions of nuclear power units described above can be referred to in correspondence.

[0081] Figure 3 This is a block diagram illustrating an intelligent diagnostic device 800 for abnormal operating conditions of a nuclear power unit, according to an exemplary embodiment. Figure 3 As shown, the intelligent diagnostic device 800 for abnormal operating conditions of nuclear power units includes a processor 801 and a memory 802. The intelligent diagnostic device 800 for abnormal operating conditions of nuclear power units also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0082] The processor 801 controls the overall operation of the intelligent diagnostic device 800 for abnormal operating conditions of the nuclear power unit, to complete all or part of the steps in the aforementioned intelligent diagnostic method for abnormal operating conditions of the nuclear power unit. The memory 802 stores various types of data to support the operation of the intelligent diagnostic device 800 for abnormal operating conditions of the nuclear power unit. This data may include, for example, instructions for any application or method operating on the intelligent diagnostic device 800 for abnormal operating conditions of the nuclear power unit, as well as application-related data, such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, or buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the nuclear power unit's intelligent diagnostic device 800 for abnormal operating conditions and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0083] In an exemplary embodiment, the intelligent diagnostic device 800 for abnormal operating conditions of nuclear power units may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the aforementioned intelligent diagnostic method for abnormal operating conditions of nuclear power units.

[0084] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the intelligent diagnostic method for abnormal operating conditions of nuclear power units described above. For example, the computer-readable storage medium may be the memory 802 including program instructions described above. These program instructions may be executed by the processor 801 of the intelligent diagnostic device 800 for abnormal operating conditions of nuclear power units to complete the intelligent diagnostic method for abnormal operating conditions of nuclear power units described above.

[0085] Example 4: Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the intelligent diagnostic method for abnormal operating conditions of nuclear power units described above.

[0086] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the intelligent diagnosis method for abnormal operating conditions of nuclear power units in the above method embodiments.

[0087] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent diagnosis of abnormal operating conditions in nuclear power units, characterized in that, include: Step 1: Collect real-time monitoring data on the pressure, temperature, neutron flux density of the reactor coolant system of the nuclear power unit, as well as the frequency deviation and voltage fluctuation amplitude on the grid side. After spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data, construct a time-series dataset that integrates the unit's operating status and grid disturbance characteristics. Step 2: Feed the time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and thermal-hydraulic conservation equations. Solve the physical equation residuals through an automatic differentiation mechanism to constrain network training, forming a dynamic response feature of the unit that integrates physical laws and data-driven approaches. Map the dynamic response feature of the unit to a heterogeneous graph neural network topology composed of the primary coolant pressure boundary nodes and the voltage phasors of the key power grid bus. Learn the cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response through spatiotemporal graph convolution operations, and construct the spatiotemporal correlation feature of unit-grid coupling. Spatiotemporal correlation features are loaded into the variational autoencoder structure to model the latent space distribution of normal and abnormal operating conditions. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence, and deep feature vectors reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances are extracted. Step 3: Using deep feature vectors as evidence variables, construct a dynamic event tree analysis framework with reactor protection system triggering, turbine digital electro-hydraulic control system fast shutdown action, and grid low-frequency low-pressure load shedding device activation as the top events. Generate accident evolution branch paths through adaptive importance sampling to form a candidate accident sequence set. Load the candidate accident sequence set into a Bayesian network inference engine based on variational inference, update the posterior distribution of equipment state failure probability with deep feature vectors, and calculate the occurrence probability of each accident sequence. Based on the occurrence probability of each accident sequence, the Fousseli-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The real-time calculated core damage frequency and the power grid transient voltage stability margin assessment value are integrated to construct a risk index set that includes the instantaneous value of core damage frequency, the probability of large-scale power outage conditions, and the comprehensive risk importance. Step 4: Input the risk index set as the state space into the multi-agent deep deterministic policy gradient algorithm framework. Configure the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. Design a reward function with the goal of minimizing the weighted entropy value of the risk index set. Optimize the communication protocol priority and power adjustment strategy under different risk levels through multi-scenario game learning in a digital twin environment. Output automatic generation control and regulation commands for the power grid dispatch center, reactor power regulation rod position suggestions for the nuclear power plant shift supervisor, and load transfer paths for the emergency command center. Generate a hierarchical communication command sequence and power adjustment suggestions. Step 5: Encode the hierarchical communication command sequence and power adjustment suggestions into general object-oriented substation event messages and manufacturing message specification messages that conform to substation communication standards. Transmit these messages through a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. Enable ultra-reliable low-latency communication mode and configure a dual-transmitter selective reception protection mechanism in the fifth-generation mobile communication network slice. Use orthogonal frequency division multiplexing multi-carrier modulation technology in the power line carrier channel. Achieve dual-channel sub-microsecond clock synchronization and seamless redundancy switching through a time-sensitive network gateway. Send the collaborative control commands to the nuclear power unit reactor power regulation system and the grid automatic generation control device respectively. Collect the reactor power response curve and grid frequency recovery trajectory after command execution and feed them back to Step 1 as input conditions for the next round of dynamic correction of abnormal monitoring data thresholds. Step two includes: The time-series dataset is fed into a physical information neural network architecture. This network embeds the reactor point reactor neutron dynamics equations and the thermo-hydraulic conservation equations. An automatic differentiation mechanism calculates the residual loss function of the physical equations, and the Adam optimizer iteratively minimizes the total loss function, forming a dynamic response characteristic of the unit that integrates physical laws and data-driven approaches. The formula for calculating the residual loss function of the physical equations is as follows: In the formula, This represents the physical constraint loss value. Let be the neutron density, and t be the time. For reactivity, To delay the issuance of neutron share 0.0065, The neutron generation time is 0.0001 seconds, and λ is the physical constraint weight of 0.

5. The dynamic response characteristics of the unit are mapped to the set of primary loop coolant pressure boundary nodes and the set of voltage phasors of key power grid buses. A heterogeneous graph neural network topology is constructed, in which the node feature matrix contains the reactor power level and the grid frequency deviation, and the edge weight matrix represents the thermal-hydraulic impedance and the transmission line admittance. The cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response is learned through spatiotemporal graph convolution operation, thus constructing the spatiotemporal correlation characteristics of unit-grid coupling. Spatiotemporal correlation features are loaded into a variational autoencoder structure. The encoder network compresses the high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The decoder network reconstructs the distribution of normal operating conditions features. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence. A deep feature vector reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances is extracted. Step three includes: using deep feature vectors as evidence variables, constructing a dynamic event tree analysis framework, setting the reactor protection system triggering, the turbine digital electro-hydraulic control system fast shutdown action, and the grid low-frequency low-pressure load shedding device start-up as top events, using an adaptive importance sampling algorithm, adjusting the branch sampling weights according to the probability of abnormal modes in the deep feature vectors, generating accident evolution branch paths through sequence Monte Carlo simulation, and forming a candidate accident sequence set; Among them, the candidate accident sequence set contains multiple accident evolution branch paths. Each path is composed of multiple branch nodes connected in series according to the state transition relationship. Each branch node includes the discrete state level of physical quantities such as reactor power, coolant temperature and pressure, grid frequency, and voltage, as well as the trigger state of automatic protection actions and human factors operation. The transition probability between nodes is jointly determined by equipment reliability data, human factors reliability data and physical process dynamic model. The candidate incident sequence set is loaded into a Bayesian network inference engine based on variational inference. A directed acyclic graph is constructed with nodes representing device status, human actions, and environmental parameters. The conditional probability table is updated using deep feature vectors as evidence input, and the occurrence probability of each incident sequence is calculated. The formula for calculating the probability of an incident sequence is as follows: In the formula, The probability of the updated accident sequence occurring. For prior probability, , where k is the first principal component score after dimensionality reduction by principal component analysis of the deep feature vector, and e is the natural constant. Based on the occurrence probability of each accident sequence, the Frosley-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The risk index set is constructed by integrating the real-time calculated core damage frequency and the power grid transient voltage stability margin assessment value.

2. The intelligent diagnostic method for abnormal operating conditions of nuclear power units according to claim 1, characterized in that, Step one includes: Piezoresistive pressure sensors and platinum resistance temperature sensors are installed in the hot and cold sections of the reactor coolant system of the nuclear power unit. A self-powered neutron detector is installed at the core outlet. A synchronous phasor measurement unit is installed on the high-voltage bus of the power grid. The original analog signals of pressure, temperature, neutron flux density, frequency deviation, and voltage fluctuation amplitude are collected at a sampling rate of 100 frames per second. After quantization and anti-aliasing filtering by a 16-bit analog-to-digital converter, the original digital sampling sequence is formed. The original digital sampling sequence is input into the clock synchronization engine based on the core period characteristics. The IEEE 1588 precise time protocol is used to align the clock between the nuclear power unit's distributed control system and the power grid's wide-area measurement system. The difference in sampling time between heterogeneous systems is compensated by cubic spline interpolation algorithm. The unified timestamp across systems is reconstructed with a period of 10 milliseconds to build a unified timestamp dataset. The unified timestamp dataset was loaded into a data cleaning engine based on sliding window Z-score anomaly detection and Min-Max normalization. The threshold for neutron flux density signal spike detection was set to 3 times the standard deviation, and the threshold for grid frequency deviation step detection was set to 0.5 Hz. After removing outlier samples, Kalman filtering was used to compensate for missing values ​​through state estimation. A 60-second sliding window was used for moving average smoothing to construct a time series dataset that integrates unit operating status and grid disturbance characteristics.

3. The intelligent diagnostic method for abnormal operating conditions of nuclear power units according to claim 1, characterized in that, Step four includes: The multi-agent deep deterministic policy gradient algorithm framework uses the risk index set as the input to the state space. The power grid dispatch center, the nuclear power plant shift leader, and the emergency command center are configured as agents with independent local observations. A reward function is designed with the goal of minimizing the weighted entropy value of the risk index set. The communication protocol priority and power adjustment strategy are optimized through multi-scenario game learning in a digital twin environment. The digital twin environment includes a reactor physics and thermal model, a turbine speed regulation model, and a power grid transient stability model. Three types of training scenarios are set: normal operating condition, design baseline accident condition, and above-design baseline accident condition. A centralized training and decentralized execution mechanism is adopted. Each agent generates actions independently through the policy network. After evaluation by a global value function, the network parameters are updated through policy gradient ascent to optimize the communication protocol priority and power adjustment strategy under different risk levels. It outputs automatic generation control and regulation commands to the power grid dispatch center, reactor power regulation rod position suggestions to the nuclear power plant operating shift leader, and load transfer paths to the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions.

4. An intelligent diagnostic system for abnormal operating conditions of nuclear power units, based on the intelligent diagnostic method for abnormal operating conditions of nuclear power units as described in claim 1, characterized in that, include: Step 1, Acquisition Module: This module is used to collect real-time monitoring data on the pressure, temperature, neutron flux density of the reactor coolant system of the nuclear power unit, as well as the frequency deviation and voltage fluctuation amplitude on the grid side. After spatiotemporal alignment and standardization preprocessing of multi-source heterogeneous data, a time-series dataset integrating the unit's operating status and grid disturbance characteristics is constructed. Step 2, Association Module: This module feeds the time-series dataset into a physical information neural network embedded with the reactor point neutron dynamics equations and thermal-hydraulic conservation equations. The residuals of the physical equations are solved using an automatic differentiation mechanism to constrain network training, forming dynamic response characteristics of the unit that integrate physical laws and data-driven approaches. These dynamic response characteristics are then mapped to a heterogeneous graph neural network topology composed of primary coolant pressure boundary nodes and grid key bus voltage phasors. Spatiotemporal graph convolution operations are used to learn the cross-scale dynamic correlation between reactor power-frequency coupling characteristics and turbine valve opening-voltage fluctuation response, constructing the spatiotemporal correlation characteristics of the unit-grid coupling. Spatiotemporal correlation features are loaded into the variational autoencoder structure to model the latent space distribution of normal and abnormal operating conditions. The probability distribution of potential abnormal modes is inferred by maximizing the lower bound of evidence, and deep feature vectors reflecting the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances are extracted. Step 3, Module Construction: This module is used to construct a dynamic event tree analysis framework with deep feature vectors as evidence variables, taking reactor protection system triggering, turbine digital electro-hydraulic control system fast shutdown action, and grid low-frequency low-pressure load shedding device startup as top events. It generates accident evolution branch paths through adaptive importance sampling, forming a candidate accident sequence set. The candidate accident sequence set is then loaded into a variational inference-based Bayesian network inference engine, updating the posterior distribution of equipment state failure probabilities with deep feature vectors, and calculating the occurrence probability of each accident sequence. Based on the occurrence probability of each accident sequence, the Fousseli-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The real-time calculated core damage frequency and the power grid transient voltage stability margin assessment value are integrated to construct a risk index set that includes the instantaneous value of core damage frequency, the probability of large-scale power outage conditions, and the comprehensive risk importance. Step 4, Output Generation Module: This module uses the risk index set as input to the state space of a multi-agent deep deterministic policy gradient algorithm framework. It configures the power grid dispatch center, nuclear power plant shift supervisor, and emergency command center as agents with independent local observations. It designs a reward function with the goal of minimizing the weighted entropy value of the risk index set. Through multi-scenario game learning in a digital twin environment, it optimizes the communication protocol priority and power adjustment strategy under different risk levels. It outputs automatic generation control and regulation commands for the power grid dispatch center, reactor power regulation rod position suggestions for the nuclear power plant shift supervisor, and load transfer paths for the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions. Step 5, Diagnostic Module: This module encodes the hierarchical communication command sequence and power adjustment suggestions into general object-oriented substation event messages and manufacturing message specification messages that conform to substation communication standards. These messages are transmitted via a heterogeneous redundant transmission architecture consisting of a fifth-generation mobile communication power dedicated network and broadband power line carrier communication. The module enables ultra-reliable low-latency communication mode and configures a dual-transmitter selective reception protection mechanism within the fifth-generation mobile communication network slice. Orthogonal frequency division multiplexing multi-carrier modulation technology is used in the power line carrier channel. Time-sensitive network gateways enable sub-microsecond clock synchronization and seamless redundancy switching between dual channels. Cooperative control commands are then sent to the nuclear power unit reactor power regulation system and the grid automatic generation control device, respectively. The module collects the reactor power response curve and grid frequency recovery trajectory after command execution and feeds them back to Step 1 as input conditions for the next round of dynamic correction of abnormal monitoring data thresholds. The associated module includes: Iterative Unit: Used to feed time-series datasets into a physical information neural network architecture. This network embeds the reactor point reactor neutron dynamics equations and the thermo-hydraulic conservation equations. It calculates the residual loss function of the physical equations through an automatic differentiation mechanism, and then iteratively minimizes the total loss function using the Adam optimizer, forming a dynamic response characteristic of the unit that integrates physical laws and data-driven approaches. The formula for calculating the residual loss function of the physical equations is as follows: In the formula, This represents the physical constraint loss value. Let be the neutron density, and t be the time. For reactivity, To delay the issuance of neutron share 0.0065, The neutron generation time is 0.0001 seconds, and λ is the physical constraint weight of 0.

5. The third building unit is used to map the dynamic response characteristics of the unit to the set of primary coolant pressure boundary nodes and the set of voltage phasors of the key bus of the power grid, and to build a heterogeneous graph neural network topology. The node feature matrix contains the reactor power level and the frequency deviation of the power grid, and the edge weight matrix represents the thermal-hydraulic impedance and the transmission line admittance. Through spatiotemporal graph convolution operation, the cross-scale dynamic correlation between the reactor power-frequency coupling characteristics and the turbine valve opening-voltage fluctuation response is learned, and the spatiotemporal correlation characteristics of the unit-grid coupling are constructed. Extraction unit: used to load spatiotemporal correlation features into the variational autoencoder structure. The encoder network compresses the high-dimensional spatiotemporal correlation features into a low-dimensional latent variable space. The decoder network reconstructs the distribution of normal operating conditions features, infers the probability distribution of potential abnormal modes by maximizing the lower bound of evidence, and extracts deep feature vectors that reflect the coupling relationship between equipment-level thermal-hydraulic anomalies and system-level electromagnetic transient disturbances. The building module includes: The fourth building unit is used to construct a dynamic event tree analysis framework using deep feature vectors as evidence variables. It sets the reactor protection system trigger, the turbine digital electro-hydraulic control system fast shutdown action, and the grid low-frequency low-pressure load shedding device as the top events. It adopts an adaptive importance sampling algorithm to adjust the branch sampling weights according to the probability of abnormal modes in the deep feature vectors. The accident evolution branch path is generated through sequence Monte Carlo simulation to form a set of candidate accident sequences. Among them, the candidate accident sequence set contains multiple accident evolution branch paths. Each path is composed of multiple branch nodes connected in series according to the state transition relationship. Each branch node includes the discrete state level of physical quantities such as reactor power, coolant temperature and pressure, grid frequency, and voltage, as well as the trigger state of automatic protection actions and human factors operation. The transition probability between nodes is jointly determined by equipment reliability data, human factors reliability data and physical process dynamic model. The fifth building unit is used to load the candidate accident sequence set into the variational inference-based Bayesian network inference engine, construct a directed acyclic graph with nodes representing device status, human actions, and environmental parameters, update the conditional probability table using deep feature vectors as evidence input, and calculate the occurrence probability of each accident sequence. The formula for calculating the probability of an accident sequence is as follows: In the formula, The probability of the updated accident sequence occurring. For prior probability, , where k is the first principal component score after dimensionality reduction by principal component analysis of the deep feature vector, and e is the natural constant. Fusion Unit: Based on the occurrence probability of each accident sequence, the Fuseli-Vesely importance of each basic event is calculated using a risk importance comprehensive ranking algorithm. The unit then merges the real-time calculated core damage frequency with the power grid transient voltage stability margin assessment value to construct a risk index set.

5. The intelligent diagnostic system for abnormal operating conditions of nuclear power units according to claim 4, characterized in that, The acquisition module includes: The first processing unit is used to install piezoresistive pressure sensors and platinum resistance temperature sensors in the hot and cold sections of the reactor coolant system of the nuclear power unit, install a self-powered neutron detector at the core outlet, and install a synchronous phasor measurement unit on the high-voltage bus of the power grid side. It collects the original analog signals of pressure, temperature, neutron flux density, frequency deviation, and voltage fluctuation amplitude at a sampling rate of 100 frames per second. After quantization and anti-aliasing filtering by a 16-bit analog-to-digital converter, the original digital sampling sequence is formed. The first building unit is used to input the original digital sampling sequence into the clock synchronization engine based on the core period characteristics. It uses the IEEE 1588 precise time protocol to align the clock between the nuclear power unit's distributed control system and the power grid's wide-area measurement system. It uses a cubic spline interpolation algorithm to compensate for the sampling time differences of heterogeneous systems and reconstructs a unified timestamp across systems with a period of 10 milliseconds to build a unified timestamp dataset. The second building block is used to load the unified timestamp dataset into a data cleaning engine based on sliding window Z-score anomaly detection and Min-Max normalization. The threshold for neutron flux density signal spike detection is set to 3 times the standard deviation, and the threshold for grid frequency deviation step detection is set to 0.5 Hz. After removing outlier samples, Kalman filtering is used to compensate for missing values ​​through state estimation. A 60-second sliding window is used for moving average smoothing to build a time series dataset that integrates unit operating status and grid disturbance characteristics.

6. The intelligent diagnostic system for abnormal operating conditions of nuclear power units according to claim 4, characterized in that, The output generation module includes: Input Unit: Used to input the risk index set as the state space into the multi-agent deep deterministic policy gradient algorithm framework, configuring the power grid dispatch center, nuclear power plant shift leader and emergency command center as agents with independent local observations, and designing a reward function with the goal of minimizing the weighted entropy value of the risk index set; Evaluation Strategy Unit: Used to optimize communication protocol priority and power adjustment strategies through multi-scenario game learning in a digital twin environment. The digital twin environment includes a reactor physics and thermal model, a turbine speed regulation model, and a power grid transient stability model. Three training scenarios are set: normal operating condition, design baseline accident condition, and above-design baseline accident condition. A centralized training and decentralized execution mechanism is adopted. Each agent generates actions independently through the policy network. The network parameters are updated through policy gradient ascent after evaluation by a global value function to optimize the communication protocol priority and power adjustment strategies under different risk levels. Generation Unit: Used to output automatic generation control and regulation commands to the power grid dispatch center, reactor power regulation rod position suggestions to the nuclear power plant operating shift leader, and load transfer paths to the emergency command center, generating hierarchical communication command sequences and power adjustment suggestions.

Images