Intelligent operation system for pressurized water reactor nuclear power plants

Abstract

The present invention is for controlling a nuclear power plant (1) comprising a reactor (11) including a core (12) and control rods (13), a primary system (10) including a pressurizer (14), a reactor coolant system and a reactor coolant pump (16), and a steam generator (21), a turbine (22), an actuator (25), and a condenser (23). The invention comprises: a collection unit (100) that collects a number of source codes and regulatory documents from a simulator of the nuclear power plant (1) and collects sensor data from the nuclear power plant (1) in real time; a knowledge integration module (110) that trains a large language model by integrating safety regulations corresponding to the source codes within the source codes and regulatory documents collected by the collection unit (100); a digital twin module (120) that synchronizes with the sensor data collected by the collection unit (100) by optimizing a number of uncertainty parameters in real time using a non-linear equation defined by a number of state variables; and a number of prediction models (131) for the sensor data collected by the collection unit (100). It is characterized by including a prediction module (130) that predicts future state variables and calculates the reliability of the future state variables, a control module (140) formed as a dual ensemble structure combining a classical control algorithm (141) and a reinforcement learning algorithm (142) that generates control commands according to future state variable data predicted by the prediction module (130), a verification module (150) that verifies the control command data generated by the control module (140) through a plurality of verification systems, and a guidance generation module (160) that generates driving guidance according to sensor data collected by the collection unit (100), future state variables predicted by the prediction model (131), control commands generated by the control module (140), and abnormality data verified by the verification module (150), using a large language model learned by the knowledge integration module (110).

Description

Technology Field

[0001] The present invention relates to an intelligent operation system for a pressurized water reactor nuclear power plant, and more specifically, to an intelligent operation system for a pressurized water reactor nuclear power plant that improves understanding of a physical model by fine-tuning source code extracted from a simulator and explanatory text corresponding to said source code, and enables real-time referencing of vast regulatory documents by establishing a search augmented generation system. Background Technology

[0002] Pressurized light water reactors, which are the mainstay of modern nuclear power generation, are multiphysics systems in which the dynamics of neutrons, the thermohydraulic behavior of the primary system, the steam cycle of the secondary system, the control system, and the safety system are non-linearly combined.

[0003] In particular, during load-following operation, Xenon-135 poisoning, changes in reactivity coefficients such as the Doppler coefficient and moderator coefficient, and thermo-hydraulic transients occur in combination, so a high degree of precision is required to manage the system stably.

[0004] However, the technologies currently used in the field of nuclear power plant operation support and control face several fundamental limitations.

[0005] First, existing digital twin models rely on low-dimensional modeling that typically adopts only 17 or fewer state variables to ensure computational efficiency.

[0006] As a result, long-term responses such as xenon swings fail to accurately reflect dynamic characteristics or the internal state of the actuator, leading to a problem where prediction accuracy drops sharply during output fluctuations.

[0007] Second, optimization techniques to correct the discrepancy between the model and the actual plant are also limited.

[0008] Existing approaches adjust only five or fewer limited parameters, such as the heat transfer multiplier, and include approximation errors unique to numerical differentiation in the process of calculating derivative values, making precise synchronization difficult.

[0009] This leads to lower reliability of the model, which ultimately results in the absence of guidance functions that provide actual diagnoses and directions for action, while merely offering simple sensor values ​​and alerts to the operator.

[0010] Third, although the Full Scope Simulator, consisting of approximately 450,000 lines of massive source code, possesses a very precise physics engine, it is limited mainly to training or post-analysis purposes and is not organically integrated with real-time driving control.

[0011] In particular, although there have been recent attempts to introduce rapidly advancing large-scale language models into nuclear power plant control, most are limited to learning general regulatory documents and fail to sufficiently understand the deep physical causal relationships inherent in the simulator's source code, which limits the generation of highly reliable operating instructions.

[0012] In conclusion, to efficiently manage nuclear power systems with complex nonlinear behavior, it is urgent to establish an innovative control system that systematically extracts source code knowledge from physical models and integrates it with intelligent models, synchronizes multiple uncertainty parameters in real time based on automatic differentiation, and combines a dual ensemble control system that harmonizes classical control and artificial intelligence with multi-layered verification logic. Prior art literature

[0013] (Patent Document 0001) KR 10-2023-0158432 A(Patent Document 0002) KR 10-2022-0089124 A(Patent Document 0003) KR 10-2023-0098765 A The problem to be solved

[0014] The present invention has been devised to solve the aforementioned problems, and the objective of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant that improves understanding of the physical model by fine-tuning source code extracted from a simulator and the explanatory text corresponding to the source code, and enables real-time referencing of vast regulatory documents by establishing a search augmentation generation system.

[0015] In addition, the objective of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant that can simulate the entire system of the nuclear power plant through a 24-state nonlinear ordinary differential equation model integrating state variables of the core, primary system, and secondary system, and, in particular, explicitly include the dynamic characteristics of xenon and iodine in the model to precisely reproduce the xenon swing phenomenon that occurs during load-following operation.

[0016] In addition, the objective of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant that can realize an ultra-precision digital twin by introducing automatic differentiation and L-BFGS algorithms to reduce the gap between the actual plant and the model, significantly expanding the optimization range of uncertainty parameters from 5 to 20, and improving state errors.

[0017] In addition, the objective of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant that can significantly improve accuracy compared to a single prediction engine by predicting state variables at future points in time through a triple ensemble system combining a machine learning model, a reinforcement learning model, and a filtering model.

[0018] In addition, the objective of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant that can shorten the time to reach the target output and drastically reduce overshoot by adopting a dual ensemble system that integrates a reinforcement learning algorithm and a classical control algorithm, and generating an optimal control signal through performance-based weights.

[0019] In addition, the objective of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant that can ensure safety through a five-stage multi-layered defense system leading to physical laws, regulatory standards of the final safety analysis report, system integrity, anomaly detection in a large language model, and final approval for the final control command, and generate operation guidance formed in a five-section structure within 800 tokens to provide decision support information with clear logical grounds to the operator. means of solving the problem

[0020] In order to solve the technical problems described above, an intelligent operation system for a pressurized water reactor nuclear power plant according to an embodiment of the present invention is configured to control a nuclear power plant (1) comprising a reactor (11) including a core (12) and control rods (13), a pressurizer (14), a primary system (10) including a reactor coolant system and a reactor coolant pump (16), and a steam generator (21), a turbine (22), an actuator (25), and a condenser (23). The system comprises: a collection unit (100) that collects a number of source codes and regulatory documents from a simulator of the nuclear power plant (1) and collects sensor data from the nuclear power plant (1) in real time; a knowledge integration module (110) that trains a large language model by integrating safety regulations corresponding to the source codes within the source codes and regulatory documents collected by the collection unit (100); and a digital twin that optimizes a number of uncertainty parameters in real time using a non-linear equation defined by a number of state variables and synchronizes them with the sensor data collected by the collection unit (100). A module (120), a prediction module (130) that predicts future state variables through a plurality of prediction models (131) for sensor data collected by the collection unit (100) and calculates the reliability of the future state variables, a control module (140) formed in a dual ensemble structure combining a classical control algorithm (141) and a reinforcement learning algorithm (142) to generate control commands according to the future state variable data predicted by the prediction module (130), a verification module (150) that verifies the control command data generated by the control module (140) through a plurality of verification systems, and using a large language model learned by the knowledge integration module (110), the sensor data collected by the collection unit (100), the future state variables predicted by the prediction model (131), and the control commands generated by the control module (140).It is characterized by including a guidance generation module (160) that generates driving guidance based on abnormality data verified by the verification module (150).

[0021] In addition, the knowledge integration module (110) analyzes the source code collected by the collection unit (100) to extract equations and variables, respectively, and converts the meaning of the equations into explanatory text. The knowledge integration module (110) stores the regulatory documents collected by the collection unit (100) to construct a knowledge database. The knowledge integration module (110) extracts safety regulations corresponding to the source code within the knowledge database. The knowledge integration module (110) maps the source code, the explanatory text corresponding to the source code, and the safety regulations to construct driving guideline data. The knowledge integration module (110) trains the large language model using the driving guideline data.

[0022] In addition, the digital twin module (120) adopts a state model that includes the toxicity dynamics of the core (12) as a variable, the digital twin module (120) generates a governing equation that describes the decay of iodine and the generation and extinction relationship of xenon, and the digital twin module (120) predicts the xenon swing that occurs during load-following operation from the governing equation.

[0023] In addition, the digital twin module (120) is characterized by using an automatic differentiation function and an L-BFGS algorithm to converge uncertainty parameters to correspond to sensor data.

[0024] Additionally, the prediction module (130) is characterized by adjusting the weights of a plurality of prediction models (131) according to the reliability of the future state variable when a defect or transient state occurs in the sensor data.

[0025] Additionally, the control module (140) determines whether the situation is normal operation or an emergency situation based on future state variable data predicted by the prediction module (130), and the control module (140) increases the weight of the classical control algorithm (141) when the situation is normal operation, and increases the weight of the reinforcement learning algorithm (142) when an emergency situation occurs.

[0026] In addition, the verification module (150) is characterized by blocking the execution of the control command and switching to a safe mode when at least one of the plurality of verification systems fails to verify.

[0027] In addition, the above driving guidance is characterized by including current status diagnosis, future prediction, recommended measures, safety verification, and caution data. Effects of the invention

[0028] In an intelligent operation system for a pressurized water reactor nuclear power plant according to an embodiment of the present invention, source code extracted from a simulator and explanatory text corresponding to the source code are fine-tuned to improve understanding of the physical model, and a search augmented generation system is established to enable real-time referencing of vast regulatory documents.

[0029] In addition, in the intelligent operation system for a pressurized water reactor nuclear power plant according to one embodiment of the present invention, the entire system of the nuclear power plant is simulated through a 24-state nonlinear ordinary differential equation model that integrates state variables of the core, primary system, and secondary system, and in particular, by explicitly including the dynamic characteristics of xenon and iodine in the model, there is an effect of being able to precisely reproduce the xenon swing phenomenon that occurs during load-following operation.

[0030] In addition, in the intelligent operation system for a pressurized water reactor nuclear power plant according to one embodiment of the present invention, automatic differentiation and L-BFGS algorithms are introduced to reduce the gap between the actual plant and the model, thereby significantly expanding the optimization range of uncertainty parameters from 5 to 20 and improving state errors, which has the effect of enabling the realization of an ultra-precision digital twin.

[0031] In addition, in the intelligent operation system of a pressurized water reactor nuclear power plant according to one embodiment of the present invention, a triple ensemble system combining a machine learning model, a reinforcement learning model, and a filtering model is used to predict state variables at future points in time, thereby having the effect of significantly improving accuracy compared to a single prediction engine.

[0032] In addition, in the intelligent operation system of a pressurized water reactor nuclear power plant according to one embodiment of the present invention, a dual ensemble system integrating a reinforcement learning algorithm and a classical control algorithm is adopted to generate an optimal control signal through performance-based weights, thereby having the effect of shortening the time to reach the target output and significantly reducing overshoot.

[0033] In addition, the intelligent operation system for a pressurized water reactor nuclear power plant according to one embodiment of the present invention ensures safety through a five-stage multi-layered defense system for the final control command, which includes physical laws, regulatory standards of the final safety analysis report, system integrity, anomaly detection in a large language model, and final approval. It also generates operation guidance formed in a five-section structure of within 800 tokens, thereby providing the operator with decision support information with clear logical grounds. Brief explanation of the drawing

[0034] Figure 1 is a diagram of a nuclear power plant. FIG. 2 is a configuration diagram of an intelligent operation system for a pressurized water reactor nuclear power plant according to an embodiment of the present invention. Figure 3 is a configuration diagram of the knowledge integration module. Figure 4 is a configuration diagram of a digital twin module. Figure 5 is an uncertainty parameter optimization algorithm. Figure 6 is a configuration diagram of the prediction module. Figure 7 is a configuration diagram of the control module. Figure 8 shows the safety verification process. Fig. 9 is an example diagram of driving guidance generation. Specific details for implementing the invention

[0035] Hereinafter, in order to explain in detail enough so that a person skilled in the art to which the present invention pertains can easily implement the technical concept of the present invention, embodiments of the present invention will be described with reference to the attached drawings.

[0036] However, the following examples are merely illustrative to aid in understanding the present invention and do not reduce or limit the scope of the present invention.

[0037] Furthermore, the present invention may be implemented in various different forms and is not limited to the embodiments described herein.

[0038] Figure 1 is a diagram of the configuration of a nuclear power plant (1).

[0039] The nuclear power plant (1) is configured to include a primary system (10) and a secondary system (20).

[0040] The primary system (10) converts a low-temperature coolant into a high-temperature coolant using heat generated through nuclear fission, and pressurizes the high-temperature coolant to a high pressure.

[0041] The secondary system (20) is connected to the primary system (10) to generate steam using the high-temperature coolant converted from the primary system (10), and to generate electricity using the steam.

[0042] The primary system (10) is composed of a reactor (11), a pressurizer (14), and a reactor coolant pump (16).

[0043] The reactor (11) converts a low-temperature coolant into a high-temperature coolant using heat generated through nuclear fission.

[0044] The pressurizer (14) is connected to the reactor (11) and maintains the high-temperature coolant discharged from the reactor (11) in a high-pressure state so that the high-temperature coolant does not boil.

[0045] The reactor coolant pump (16) is connected to the reactor (11) and pumps low-temperature coolant into the reactor (11).

[0046] The secondary system (20) is configured to include a steam generator (21), a turbine (22), a condenser (23), and an actuator (25).

[0047] The steam generator (21) is connected to the pressurizer (14) and heat exchanges the high-temperature coolant that has passed through the pressurizer (14) with the low-temperature feed water, thereby generating steam from the low-temperature feed water.

[0048] The turbine (22) is connected to the steam generator (21) and rotates by the steam generated by the steam generator (21).

[0049] A condenser (23) is provided at the bottom of the turbine (22) and cools the steam passing through the turbine (22) to convert it into low-temperature feed water.

[0050] The actuator (25) may include a water supply valve (27), a governor valve (28), and a steam dump (29) (not shown).

[0051] The main feedwater pump (26) is connected to the condenser (23) and the steam generator (21) and pumps the low-temperature feedwater converted by the condenser (23) to the steam generator (21).

[0052] The water supply valve (27) is provided between the condenser (23) and the steam generator (21) and is formed to be openable and closable.

[0053] The governor valve (28) is provided between the steam generator (21) and the turbine (22) and is formed to be openable and closable.

[0054] The steam dump (29) is connected to the steam generator (21) and, in the event of an emergency, releases the steam generated by the steam generator (21) into the atmosphere.

[0055] The reactor (11) is composed of a core (12) and control rods (13).

[0056] The core (12) is installed inside the reactor (11) and generates heat through nuclear fission.

[0057] The control rod (13) is inserted into the reactor (11) and provided in the core (12) to reduce the nuclear chain reaction occurring in the core (12) by absorbing neutrons.

[0058] A number of nuclear fuels that generate heat through nuclear fission are provided inside the core (12).

[0059] A cladding material surrounding the nuclear fuel is provided on the outer side of the nuclear fuel.

[0060] FIG. 2 is a configuration diagram of an intelligent operation system for a pressurized water reactor nuclear power plant according to an embodiment of the present invention.

[0061] Referring to FIG. 2, an intelligent operation system for a pressurized water reactor nuclear power plant according to an embodiment of the present invention comprises a collection unit (100), a knowledge integration module (110), a digital twin module (120), a prediction module (130), a control module (140), a verification module (150), and a guidance generation module (160).

[0062] The collection unit (100) collects a number of source codes and regulatory documents from the simulator of the nuclear power plant (1) and collects sensor data from the nuclear power plant (1) in real time.

[0063] The knowledge integration module (110) integrates the source code collected by the collection unit (100) and the safety regulations corresponding to the source code within the regulatory documents to train a large language model.

[0064] The digital twin module (120) optimizes multiple uncertainty parameters in real time using a non-linear equation defined by multiple state variables and synchronizes with the sensor data collected by the collection unit (100).

[0065] The prediction module (130) predicts future state variables for each of the sensor data collected by the collection unit (100) through a plurality of prediction models (131) and calculates the reliability of the future state variables.

[0066] The control module (140) is formed as a dual ensemble structure combining a classical control algorithm (141) and a reinforcement learning algorithm (142) to generate control commands according to future state variable data predicted by the prediction module (130).

[0067] The verification module (150) verifies the control command data generated by the control module (140) through a plurality of verification systems.

[0068] The guidance generation module (160) generates driving guidance based on sensor data collected by the collection unit (100), future state variables predicted by the prediction model (131), control commands generated by the control module (140), and abnormality data verified by the verification module (150), using a large language model learned by the knowledge integration module (110).

[0069] The above source code may be source code generated by Fortran or the C language.

[0070] The above regulatory documents may include a final safety analysis report, emergency operation procedures, abnormal operation procedures, and technical guidelines for the nuclear power plant (1).

[0071] The above multiple verification systems may include compliance with physical laws, compliance with the final safety analysis report, the status of the safety system, and the presence or absence of abnormalities and approval of the large language model.

[0072] Sensor data may include core sensor data, primary system sensor data, and secondary system sensor data.

[0073] Core sensor data may include neutron flux, core power, control rod positions, and boron concentration.

[0074] Primary system sensor data may include high-temperature tube temperature, low-temperature tube temperature, reactor coolant system pressure, pressurizer water level, reactor coolant system flow rate, and reactor coolant pump speed.

[0075] Secondary system sensor data may include steam generator pressure, water level, feedwater flow rate, main steam flow rate, turbine speed, turbine output, and condenser pressure.

[0076] Figure 3 is a configuration diagram of the knowledge integration module (110).

[0077] Referring to FIG. 3, the knowledge integration module (110) analyzes the source code collected by the collection unit (100), extracts equations and variables respectively, and converts the meaning of the equations into explanatory text.

[0078] The knowledge integration module (110) stores regulatory documents collected by the collection unit (100) to build a knowledge database.

[0079] The knowledge integration module (110) extracts safety regulations corresponding to the source code within the knowledge database.

[0080] The knowledge integration module (110) maps the source code and the description and safety regulations corresponding to the source code to construct driving guideline data.

[0081] The knowledge integration module (110) trains the large language model using the driving instruction data.

[0082] Meanwhile, the digital twin module (120) adopts a state model that includes the dynamic characteristics of the core (12) as variables.

[0083] The above digital twin module (120) generates a governing equation that describes the decay of iodine and the generation and extinction relationships of xenon.

[0084] The digital twin module (120) can predict the xenon swing that occurs during load-following operation from the governing equation.

[0085] Through this, the control rod (13) can be operated or the concentration of boron can be adjusted.

[0086] FIG. 4 is a configuration diagram of a digital twin module (120), and FIG. 5 is an uncertainty parameter optimization algorithm.

[0087] Referring to FIGS. 4 and 5, the digital twin module (120) can optimize the uncertainty parameter in real time so that the difference between the uncertainty parameter and the sensor data can be minimized.

[0088] Specifically, the digital twin module (120) uses an automatic differentiation function and an L-BFGS algorithm to converge uncertainty parameters to correspond to sensor data.

[0089] Through this, the reliability of the data can be ensured even if the nuclear power plant (1) experiences aging deterioration or sensor errors.

[0090] The digital twin module (120) is configured to include a physical model section (121) and an optimization section (122).

[0091] The physical model section (121) models the behavior of a nuclear power plant (1), including a core (12), a primary system (10), and a secondary system (20), using a nonlinear ordinary differential equation defined by a number of state variables.

[0092] The optimization unit (122) optimizes the uncertainty parameter so that the weighted squared error between the sensor data collected by the collection unit (100) and the output value of the physical model unit (121) can be minimized.

[0093] The optimization unit (122) defines the uncertainty parameter.

[0094] The optimization unit (122) defines an objective function that includes the weighted squared error between the sensor data collected by the collection unit (100) and the state variables of the physical model.

[0095] The optimization unit (122) performs automatic differentiation on the nonlinear ordinary differential equation modeled by the physical model unit to calculate sensitivity for each uncertainty parameter.

[0096] The above sensitivity refers to the value obtained by taking the partial derivative of the nonlinear ordinary differential equation with respect to the uncertainty parameter.

[0097] The optimization unit (122) optimizes the uncertainty parameter in a direction that minimizes the objective function through the L-BFGS algorithm.

[0098] The optimization unit (122) can perform the process of optimizing the sensitivity calculation and uncertainty parameters through parallel computation within a period of 60 seconds.

[0099] The automatic differentiation function refers to the ability to calculate the sensitivity of a state variable to changes in uncertainty parameters in real time using the chain rule, without numerical approximation error.

[0100] L-BFGS stands for Limited-memory Broyden-Fletcher-Goldfarb-Shanno.

[0101] The L-BFGS algorithm can determine the optimization direction and magnitude of uncertainty parameters, respectively, using a quasi-Newtonian method based on the influence calculated by the automatic differentiation function above.

[0102] The uncertainty parameter optimized by the above optimization unit (122) is transmitted to the knowledge integration module (110) based on a large language model.

[0103] Through this, the cause of the abnormal behavior of the nuclear power plant (1) can be physically analyzed.

[0104] Meanwhile, the above state variables are composed of core, primary system, and secondary system state variables.

[0105] The number of the above state variables can be set to 24.

[0106] Core state variables include neutron flux, delayed neutron leading nucleus, fuel temperature, moderator temperature, iodine concentration, and xenon concentration.

[0107] Primary system state variables include hot tube temperature, cold tube temperature, reactor coolant system pressure, pressurizer water level, reactor coolant system flow rate, reactor coolant system stock, boron concentration, and reactor coolant pump speed.

[0108] Secondary system state variables include steam generator pressure, water level, internal energy, feedwater valve opening, main steam pressure, governor valve opening, turbine speed, condenser pressure, hot well water level, and steam dump valve opening.

[0109] Meanwhile, the above uncertainty parameter may include a core parameter and an extension parameter.

[0110] The number of the above uncertainty parameters can be set to N. The above N can be set to 20.

[0111] The above key parameters may include heat transfer multiplier, reactivity bias, flow resistance, thermal efficiency, and pressure loss.

[0112] The above extension parameters may include pump efficiency, control rod (13) value deviation, xenon decay constant deviation, and device-specific time constant error.

[0113] FIG. 6 is a configuration diagram of a prediction module (130). Referring to FIG. 6, a plurality of prediction models (131) may include a machine learning model (132), a reinforcement learning model (133), and a filtering model (134).

[0114] A machine learning model (132) learns past operating data for a nuclear power plant (1) and predicts future state variables.

[0115] The reinforcement learning model (133) predicts future state variables based on driving data derived through trial and error within a virtual simulation environment.

[0116] The filtering model (134) predicts future state variables based on the equations extracted by the knowledge integration module (110).

[0117] The filtering model (134) can be implemented in the form of a Kalman filter.

[0118] The prediction module (130) adjusts the weights of a plurality of prediction models (131) according to the reliability of the future state variable when a defect or transient state occurs in the sensor data.

[0119] Here, the weight of the prediction model (131) represents the importance assigned to each prediction model (131).

[0120] The weights of the above prediction model (131) can be set to 0 or more and 1 or less. The sum of the weights of the above prediction model (131) can be set to 1.

[0121] Specifically, when the prediction model (131) is in normal operation, the weights of the machine learning model (132), reinforcement learning model (133), and filtering model (134) are set to 0.4, 0.3, and 0.3, respectively.

[0122] When a defect occurs in the sensor data, the prediction module (130) lowers the weight of the machine learning model (132) and raises the weights of the reinforcement learning model (133) and the filtering model (134), respectively.

[0123] When a defect occurs in the sensor data, the prediction module (130) adjusts the weight of the machine learning model (132) to 0.18 and the weights of the reinforcement learning model (133) and the filtering model (134) to 0.35 and 0.47, respectively.

[0124] The prediction module (130) can ensure the stability of the prediction by increasing the weight of the filtering model (134) based on physical laws when a defect occurs in the sensor data.

[0125] The prediction module (130) can increase the weights of the reinforcement learning model (133) when a transient state occurs.

[0126] FIG. 7 is a configuration diagram of a control module (140). Referring to FIG. 7, a classical control algorithm (141) generates control commands based on mathematical models and physical laws.

[0127] The reinforcement learning algorithm (142) generates control commands for non-linear states through trial and error learning.

[0128] The control module (140) determines whether the operation is on schedule or an emergency situation based on future state variable data predicted by the prediction module (130).

[0129] The control module (140) increases the weight of the classical control algorithm (141) during normal operation, and increases the weight of the reinforcement learning algorithm (142) when an emergency situation occurs.

[0130] Here, the weight of the classical control algorithm (141) is the importance assigned to the classical control algorithm (141), and the weight of the reinforcement learning algorithm (142) is the importance assigned to the reinforcement learning algorithm (142).

[0131] The sum of the weights of the classical control algorithm (141) and the weights of the reinforcement learning algorithm (142) can be set to 1.

[0132] When the control module (140) is in normal operation, it sets the weight of the classical control algorithm (141) to 0.65 to 0.75.

[0133] The control module (140) increases the weight of the reinforcement learning algorithm (142) to 0.65 to 0.75 when an emergency situation occurs.

[0134] The control module (140) increases the weight of the reinforcement learning algorithm (142) to 0.55 to 0.65 when responding to a xenon swing.

[0135] By combining the damping characteristics of the classical control algorithm (141) and the non-linear adaptability of the reinforcement learning algorithm (142), the time to reach the target output can be shortened and overshoot can be suppressed.

[0136] Meanwhile, the verification module (150) blocks the execution of the control command and switches to a safe mode when at least one of the plurality of verification systems fails to verify.

[0137] Figure 8 is a safety verification process.

[0138] Referring to FIG. 8, the verification module (150) may be configured to include a law verification unit (151), a reference verification unit (152), a system verification unit (153), a model verification unit (154), and an approval unit (155).

[0139] The law verification unit (151) applies the control command generated by the control module (140) to the digital twin module (120) to simulate the state after a preset reference time and verifies compliance with the laws of conservation of energy, mass, and momentum.

[0140] The above standard time can be set to 8 to 12 seconds.

[0141] The standard verification unit (152) verifies whether the final safety analysis report complies with the standards.

[0142] The criteria for the final safety analysis report may include DNBR (Departure from Nucleate Boiling Ratio), peak cladding temperature, reactor coolant system pressure, and nuclear fuel temperature.

[0143] The system verification unit (153) verifies the availability of the emergency core cooling system, the integrity of the containment building, and the possibility of tripping the reactor protection system, respectively.

[0144] The model verification unit (154) detects abnormal patterns in control commands and future state variables through the large language model, calculates the reliability of the control commands and future state variables, and determines that it passes when the reliability of the control commands and future state variables exceeds a preset threshold.

[0145] The above threshold can be set to 0.75 or higher and 0.85 or lower.

[0146] The approval unit (155) approves the control command based on the verification results of the above-mentioned law verification unit (151), reference verification unit (152), system verification unit (153), and model verification unit (154).

[0147] The approval unit (155) classifies the control command into one of approval, waiting for approval, or rejection based on the verification results of the above-mentioned law verification unit (151), reference verification unit (152), system verification unit (153), and model verification unit (154).

[0148] Figure 9 is an example diagram of driving guidance generation.

[0149] Referring to FIG. 9, the driving guidance may include current state diagnosis, future prediction, recommended measures, safety verification, and caution data.

[0150] Meanwhile, the current state diagnosis data refers to data that diagnoses the current operating mode by comparing the sensor data collected by the collection unit (100) with a preset normal range.

[0151] Future state prediction data includes prediction reliability data including the variability of state variables, causes, and 95% confidence intervals.

[0152] Recommended action data includes details of control commands, expected effects, and time required.

[0153] Safety verification data includes compliance with the final safety analysis report and summary data on the verification results.

[0154] The caution data may include output change rate limits and data on the necessity of entering emergency procedures.

[0155] In an intelligent operation system for a pressurized water reactor nuclear power plant according to an embodiment of the present invention, source code extracted from a simulator and explanatory text corresponding to the source code are fine-tuned to improve understanding of the physical model, and a search augmented generation system is established to enable real-time referencing of vast regulatory documents.

[0156] In addition, in the intelligent operation system for a pressurized water reactor nuclear power plant according to one embodiment of the present invention, the entire system of the nuclear power plant (1) is simulated through a 24-state nonlinear ordinary differential equation model that integrates the core, primary system, and secondary system state variables, and in particular, the dynamic characteristics of xenon and iodine are explicitly included in the model, thereby having the effect of precisely reproducing the xenon swing phenomenon that occurs during load-following operation.

[0157] In addition, in the intelligent operation system for a pressurized water reactor nuclear power plant according to one embodiment of the present invention, automatic differentiation and L-BFGS algorithms are introduced to reduce the gap between the actual plant and the model, thereby significantly expanding the optimization range of uncertainty parameters from 5 to 20 and improving state errors, which has the effect of enabling the realization of an ultra-precision digital twin.

[0158] In addition, in the intelligent operation system of a pressurized water reactor nuclear power plant according to one embodiment of the present invention, a triple ensemble system combining a machine learning model, a reinforcement learning model (133), and a filtering model (134) is used to predict state variables at a future point in time, thereby having the effect of significantly improving accuracy compared to a single prediction engine.

[0159] In addition, in the intelligent operation system of a pressurized water reactor nuclear power plant according to one embodiment of the present invention, a dual ensemble system integrating a reinforcement learning algorithm (142) and a classical control algorithm (141) is adopted to generate an optimal control signal through performance-based weights, thereby having the effect of shortening the time to reach the target output and drastically reducing overshoot.

[0160] In addition, the intelligent operation system for a pressurized water reactor nuclear power plant according to one embodiment of the present invention ensures safety through a five-stage multi-layered defense system for the final control command, which includes physical laws, regulatory standards of the final safety analysis report, system integrity, anomaly detection in a large language model, and final approval. It also generates operation guidance formed in a five-section structure of within 800 tokens, thereby providing the operator with decision support information with clear logical grounds.

[0161] As described above, the main technical concept of the present invention is to provide an intelligent operation system for a pressurized water reactor nuclear power plant. The embodiments described above with reference to the drawings are merely one example, and the true scope of the present invention is based on the patent claims, but extends to various equivalent embodiments that may exist. Explanation of the symbols

[0162] 1: Nuclear power plant 10: Primary system 11: Nuclear reactor 12: Core 13: Control rods 14: Pressurizer 15: Reactor Coolant System 16: Coolant pump 20: Secondary system 21: Steam generator 22: Turbine 23: Revenge 25: Actuator 26: Main water supply pump 27: Water supply valve 28: Governor valve 29: Steam Dump 30: Generator 100: Collection Department 110: Knowledge Integration Module 120: Digital Twin Module 121: Physical Model Department 122: Optimization Section 130: Prediction Module 131: Prediction Model 132: Machine learning models 133: Reinforcement Learning Model 134: Filtering Model 140: Control Module 141: Classical Control Algorithms 142: Reinforcement Learning Algorithms 150: Verification Module 151: Law Verification Department 152: Standard Verification Department 153: System Verification Unit 154: Model Validation Unit 155: Approval Department 160: Guidance Generation Module

Claims

Claim 1 An intelligent operating system for a pressurized light water reactor nuclear power plant for controlling a nuclear power plant (1) comprising a reactor (11) including a core (12) and control rods (13), a primary system (10) including a pressurizer (14), a reactor coolant system and a reactor coolant pump (16), and a steam generator (21), a turbine (22), an actuator (25), and a condenser (23), comprising: a collection unit (100) that collects a number of source codes and regulatory documents from a simulator of the nuclear power plant (1) and collects sensor data from the nuclear power plant (1) in real time; a knowledge integration module (110) that trains a large language model by integrating safety regulations corresponding to the source codes within the source codes and regulatory documents collected by the collection unit (100); and a digital twin module (120) that optimizes a number of uncertainty parameters in real time using a non-linear equation defined by a number of state variables and synchronizes them with the sensor data collected by the collection unit (100). A prediction module (130) that predicts future state variables through a plurality of prediction models (131) for sensor data collected by the collection unit (100) and calculates the reliability of the future state variables; a control module (140) formed as a dual ensemble structure combining a classical control algorithm (141) and a reinforcement learning algorithm (142) to generate control commands according to the future state variable data predicted by the prediction module (130); a verification module (150) that verifies the control command data generated by the control module (140) through a plurality of verification systems; and a guidance generation module (160) that generates driving guidance according to sensor data collected by the collection unit (100), future state variables predicted by the prediction model (131), control commands generated by the control module (140), and abnormality data verified by the verification module (150), using a large language model learned by the knowledge integration module (110).An intelligent operation system for a pressurized water reactor nuclear power plant, comprising: a knowledge integration module (110) that analyzes source code collected by the collection unit (100) to extract equations and variables, respectively, and converts the meaning of the equations into explanatory text; a knowledge integration module (110) that stores regulatory documents collected by the collection unit (100) to construct a knowledge database; a knowledge integration module (110) that extracts safety regulations corresponding to the source code within the knowledge database; a knowledge integration module (110) that maps the source code and the explanatory text and safety regulations corresponding to the source code to construct operation guideline data; and a knowledge integration module (110) that trains a large language model using the operation guideline data. Claim 2 delete Claim 3 In claim 1, the digital twin module (120) adopts a state model that includes the toxicity dynamics of the core (12) as a variable, the digital twin module (120) generates a governing equation that describes the decay of iodine and the generation and extinction relationship of xenon, and the digital twin module (120) predicts the xenon swing that occurs during load-following operation from the governing equation, an intelligent operation system for a pressurized water reactor nuclear power plant. Claim 4 In claim 1, the digital twin module (120) is an intelligent operation system for a pressurized water reactor nuclear power plant that uses an automatic differentiation function and an L-BFGS algorithm to converge uncertainty parameters to correspond to sensor data. Claim 5 In claim 1, the prediction module (130) adjusts the weights of a plurality of prediction models (131) according to the reliability of the future state variable when a defect or transient state occurs in the sensor data, in an intelligent operation system for a pressurized water reactor nuclear power plant. Claim 6 An intelligent operation system for a pressurized water reactor nuclear power plant according to claim 1, wherein the control module (140) determines whether the operation is on schedule or an emergency situation based on future state variable data predicted by the prediction module (130), and the control module (140) increases the weight of the classical control algorithm (141) during normal operation and increases the weight of the reinforcement learning algorithm (142) during an emergency situation. Claim 7 In claim 1, the verification module (150) is an intelligent operation system for a pressurized water reactor nuclear power plant that blocks the execution of the control command and switches to a safe mode when at least one of the plurality of verification systems fails to verify. Claim 8 In claim 1, the operation guidance is an intelligent operation system for a pressurized water reactor nuclear power plant that includes current status diagnosis, future prediction, recommended measures, safety verification, and precaution data.

Images