Core protection system using DNBR Power Trip Limit evaluation model based on universal approximation theorem

Abstract

The present invention relates to a core protection system using a DNBR PTL evaluation model based on a universal approximation theorem, and to a core protection system that determines whether to shut down a reactor by inferring a Power Trip Limit (PTL) that reaches a limit value (DL; DNBR Limit) of the Departure from Nucleate Boiling Ratio (DNBR) through an artificial neural network model to which the universal approximation theorem is applied. According to the present invention, the DNBR PTL can be predicted quickly and stably under various operating conditions using only a single hidden-layer-based neural network structure, by including a core operation information collection unit (100) that collects values ​​of an ex-core neutron flux or in-core neutron flux having multiple axial levels, a core inlet temperature, pressure, flow rate, output, and radial peak output factor (Fr); a DNBR PTL evaluation unit (200) that calculates the DNBR PTL by an artificial neural network model including one or more hidden layers using the collected core operation information as input; and a shutdown determination unit (300) that generates a reactor shutdown signal when the current operating output value collected from the core operation information collection unit (100) is higher than the calculated PTL.

Description

Technology Field

[0001] The present invention relates to a core protection system using a DNBR PTL evaluation model based on a universal approximation theorem, and more specifically, to a core protection system that determines whether to shut down a reactor by inferring a Power Trip Limit (PTL) that reaches a Departure from Nucleate Boiling Ratio (DNBR) limit value (DL; DNBR Limit) through an artificial neural network model to which the universal approximation theorem is applied. Background Technology

[0002] Domestic core protection systems use a simplified auxiliary reactor analysis model with inputs such as power, temperature, pressure, flow rate, and power distribution to directly calculate the DNBR, and shut down the reactor when the calculated DNBR is smaller than the set value.

[0003] Although such core protection systems can accurately calculate actual DNBR, due to the complexity of the calculation, they cannot use FPGAs based on analog circuits or digital logic circuits and must utilize equipment based on digital computers.

[0004] In particular, regarding the method of operating a reactor by online calculation of DNBR and the corresponding reactor of Public Patent 10-2021-0057134, the basic concept is the same as that of the domestic core protection system, but there is a difference in that the DNBR calculation model is implemented using a deep neural network.

[0005] Since models that predict such DNBR can calculate various DNBR values ​​under various operating conditions, there are limitations in accurately estimating the DNBR at DL required in a core protection system. Furthermore, it is estimated that there are problems with the algorithm becoming somewhat complex due to the use of deep neural network models to accurately simulate it, so there may be limitations in implementing it on platforms such as digital logic circuit-based FPGAs. Prior art literature

[0006] Published Patent No. 2021-0057134 (May 20, 2021) The problem to be solved

[0007] The present invention aims to solve the aforementioned problems and to secure flexibility applicable to various platforms by simplifying the algorithm structure so that it can be applied not only to computer-based PLC equipment but also to digital logic circuit-based FPGAs in order to prevent common cause failures.

[0008] In addition, the purpose is to prevent damage to the nuclear fuel cladding by comparing the current output with the output value (PTL) that reaches the DNBR limit value (DL) in real time, and by generating a reactor shutdown signal when the current output exceeds the PTL, thereby preventing damage to the nuclear fuel from abnormal situations. means of solving the problem

[0009] The present invention relates to a core protection system using a universal approximation theorem-based DNBR PTL evaluation model, comprising: a core operation information collection unit (100) that collects values ​​of ex-core neutron flux or in-core neutron flux having multiple axial levels, core inlet temperature, pressure, flow rate, output, and radial peak output factor (Fr); a DNBR PTL evaluation unit (200) that calculates a DNBR PTL by an artificial neural network model including one or more hidden layers using the collected core operation information as input; and a shutdown determination unit (300) that generates a reactor shutdown signal when the current operating output value collected from the core operation information collection unit (100) is higher than the calculated PTL. The artificial neural network model is optimized (learned or trained) based on data derived from nuclear design and thermohydraulic design codes under various operating conditions of the core protection system and is configured to evaluate the DNBR PTL based on the point in time when the reactor output reaches a DNBR limit value (DL).

[0010] Preferably, the artificial neural network model is pre-trained using training data generated based on operating factors (inlet temperature, pressure, flow rate, output, neutron flux, Fr, etc.) provided by the core operation information collection unit (100), and is configured to evaluate the DNBR PTL based on the time of reaching DL.

[0011] The present invention includes a core operation information collection unit (100) that collects as inputs an ex-core neutron flux or in-core neutron flux having multiple axial levels, a core inlet temperature, pressure, flow rate, output, and a radial peak output factor (Fr) by applying an artificial neural network model based on a universal approximation theorem; a DNBR PTL evaluation unit (200) that calculates in real time a power trip limit (PTL) that reaches a DNBR limit value (DL) based on the inputs; and a shutdown determination unit (300) that generates a reactor shutdown signal when the calculated PTL is compared with the current operating output value and the PTL is exceeded.

[0012] The above artificial neural network model includes one or more hidden layers and is optimized based on data derived from nuclear design codes and thermal-hydraulic design codes, and is configured to evaluate the DNBR PTL based on the time of reaching DL, thereby enabling stable and rapid reactor protection under various operating conditions.

[0013] PTL, which is the output condition at which the DNBR limit value (DL) is reached, is calculated based on one or more computer codes using a simplified numerical model as well as the thermal-hydraulic design code.

[0014] Here, the simplified numerical model is a concept that includes a simplified sub-analysis model, and a simplified computational model can be applied in addition to the thermal-hydraulic design code.

[0015] The PTL, which is the output condition at which the DNBR limit value (DL) is reached, can be calculated not only based on the calculation results of the thermal-hydraulic design code, but also using one or more computer codes utilizing simplified numerical models (including simplified auxiliary furnace analysis models). Through this, by utilizing data derived from various computer models as well as complex thermal-hydraulic analysis codes, it is possible to rapidly and flexibly evaluate the DNBR PTL in response to changes in operating conditions.

[0016] The core operation information collection unit (100) collects some of the input factors, including values ​​derived through calculation from other measured factors as well as values ​​directly measured.

[0017] The core operation information collection unit (100) can collect some of the input factors as values ​​directly measured through sensors, as well as be configured to include values ​​derived through calculation from other measured factors. For example, the core flow rate can be calculated based on the measured values ​​of pump rotational speed and differential pressure, and the core output can be calculated using an energy conservation formula that uses the inlet temperature, outlet temperature, and coolant flow rate. Additionally, the radial peak output factor (Fr) can use values ​​calculated from the synthesis of a three-dimensional output distribution or design-defined conservative values ​​along with real-time measured values.

[0018] The radial peak output factor (Fr) may include not only measured values ​​but also various output ratio values ​​that can be set in response to operating conditions or values ​​that are pre-set according to design maintenance judgments. In the present invention, the Fr value may be set to various output ratios corresponding to operating conditions using not only actual measured values ​​but also output range tables or pre-defined maximum values.

[0019] The radial peak output factor (Fr) can use values ​​measured in real time, as well as apply predefined maximum values ​​or table values ​​for each output range based on design conservative judgment. For example, Fr, which represents the ratio of average output to high-temperature rod output, can be used directly if actual measured values ​​exist, and if measurement uncertainty or safety margins are required, conservative values ​​or various output ratio values ​​corresponding to operating conditions can be selectively applied to maintain conservativeness at all times when evaluating DNBR PTL.

[0020] The DNBR PTL evaluation unit (200) receives input values ​​collected from the core operation information collection unit (100) in real time during reactor operation, infers the DNBR PTL in real time based on the input values, and the shutdown determination unit (300) includes a real-time response structure that outputs a reactor shutdown signal immediately when the current operation output exceeds the inferred PTL.

[0021] The DNBR PTL evaluation unit (200) receives input values ​​in real time from the core operation information collection unit (100) during reactor operation and infers the DNBR PTL in real time, which is the output condition that reaches the DNBR limit value (DL) under the current operating conditions, based on the input values. Additionally, the shutdown judgment unit (300) includes a real-time response structure that immediately compares the inferred PTL with the current operating output value and generates a reactor shutdown signal immediately if the current output exceeds the PTL. Through this, rapid judgment and control are possible even under rapidly changing operating conditions, and an immediate protection function can be performed to prevent nuclear fuel damage.

[0022] The DNBR PTL evaluation unit (200) includes an uncertainty correction module (240) that applies a correction coefficient or a repair coefficient to the DNBR PTL value calculated by an artificial neural network model in response to the generalization error of the prediction model, the measurement error of the operating data, or the manufacturing tolerance, and the shutdown determination unit (300) is configured to determine whether to shut down the reactor based on the DNBR PTL value corrected through the uncertainty correction module (240).

[0023] The DNBR PTL evaluation unit (200) may include an uncertainty correction module (240) that applies a correction factor or a repair factor to the DNBR PTL value calculated by the artificial neural network model, taking into account the generalization error of the prediction model, the measurement error of the operation data, and the manufacturing tolerance.

[0024] The uncertainty correction module (240) performs a two-step correction procedure on the inference value of the neural network. First, it corrects the model's predicted value by applying a 'correction factor' that takes into account measurement error and model generalization error. Second, it calculates a more conservative DNBR PTL value by additionally applying a 'conservative factor' to secure measurement uncertainty and design margin. This dual correction structure satisfies regulatory requirements and safety design standards, and enables protection that always assumes the worst-case condition in determining whether the DL is reached.

[0025] The corrected DNBR PTL value calculated by the above uncertainty correction module (240) is provided as a comparison standard for the shutdown determination unit (300), and the shutdown determination unit (300) generates a reactor shutdown signal when the current output exceeds this corrected DNBR PTL value, thereby enabling conservative core protection that takes into account measurement uncertainty and model deviation.

[0026] A method using a core protection system utilizing a universal approximation theorem-based DNBR PTL evaluation model according to the present invention comprises the steps of: (a) a core protection system utilizing a universal approximation theorem-based DNBR PTL evaluation model (hereinafter referred to as the "system") collecting ex-core neutron flux or in-core neutron flux having multiple axial levels, core inlet temperature, pressure, flow rate, power information, and radial peak power factor (Fr) values; (b) inputting the collected operating information into an artificial neural network model comprising one or more hidden layers to calculate a power value (DNBR PTL) at the point in time when the DNBR reaches a limit value (DL) under current operating conditions; and (c) immediately generating a reactor shutdown signal if the current operating power is higher than the calculated DNBR PTL. The artificial neural network model is pre-optimized based on learning data derived from nuclear design codes and thermohydraulic design codes under various operating conditions and is configured to evaluate the DNBR PTL based on the point in time when the DL is reached. Effects of the invention

[0027] According to the present invention, the output value (PTL) at which the DNBR reaches a limit value (DL) under current operating conditions is predicted in real time through an artificial neural network, and the reactor is shut down when the current output is higher than the PTL, thereby preventing damage to the nuclear fuel cladding.

[0028] In addition, the present invention can rapidly and stably predict DNBR PTL under various operating conditions using only a single hidden layer-based neural network structure.

[0029] It can be applied to FPGA or PLC-based systems through a lightweight structure that facilitates hardware implementation, and enables real-time control.

[0030] In addition, the safety of the system can be ensured by applying uncertainty correction and compensation factors to the inference results of the neural network. Brief explanation of the drawing

[0031] Figure 1 shows the configuration of a core protection system using a DNBR PTL evaluation model based on a universal approximation theorem according to one embodiment of the present invention. Figure 2 shows the detailed configuration of the core operation information collection unit according to the present embodiment. Figure 3 shows the detailed configuration of the DNBR PTL evaluation unit according to the present embodiment. Figure 4 shows the detailed configuration of the stop judgment unit according to the present embodiment. Figure 5 shows an example of applying Fr by output of a core protection system using a DNBR PTL evaluation model based on a universal approximation theorem according to one embodiment of the present invention. FIG. 6 is a model for calculating the Power Trip Limit (DNBR PTL) of a core protection system having one hidden layer using a universal approximation theorem-based DNBR PTL evaluation model according to one embodiment of the present invention. FIG. 7 is a flowchart illustrating a method using a core protection system based on a universal approximation theorem-based DNBR PTL evaluation model according to one embodiment of the present invention. Specific details for implementing the invention

[0032] The present invention relates to a core protection system configured to enable simple and stable real-time evaluation of the Departure from Nucleate Boiling Ratio (DNBR) PTL for the safe operation of a nuclear reactor and prevention of nuclear fuel damage.

[0033] In particular, the present invention is configured to accurately estimate DNBR PTL according to operating conditions using a simple model by utilizing an artificial neural network based on the general approximation theorem, without complex thermal hydraulic by-furnace analysis.

[0034] This allows for implementation on hardware platforms such as FPGAs or PLCs even in situations requiring high-speed real-time judgment, and is characterized by significantly improved hardware applicability compared to existing complex calculation models in that it is designed to enable conservative reactor shutdown judgments for various operating conditions.

[0035] The universal approximation theorem in this embodiment is a theory that mathematically guarantees that an artificial neural network having one hidden layer and a sufficient number of hidden nodes can approximate any continuous function and non-linear relationship with desired accuracy.

[0036] Figure 1 shows the configuration of a core protection system using a DNBR PTL evaluation model based on a universal approximation theorem according to one embodiment of the present invention.

[0037] As illustrated in FIG. 1, the core protection system (10) using a universal approximation theorem-based DNBR PTL evaluation model according to the present embodiment includes a core operation information collection unit (100), a DNBR PTL evaluation unit (200), and a shutdown determination unit (300).

[0038] The core operation information collection unit (100) performs the function of collecting or calculating the following key operation variables in real time to configure the input values ​​of an artificial neural network model for DNBR PTL evaluation.

[0039] This core operation information collection unit collects values ​​for the radial peak power factor (Fr) based on the core output, the core inlet temperature, pressure, flow rate, and core output, as well as the ex-core neutron flux or in-core neutron flux having two or more axial levels.

[0040] In relation to an ex-core neutron flux or an in-core neutron flux having two or more axial levels, the reactor core is used as an input representing the axial (Z-axis) power distribution.

[0041] In addition, the core inlet temperature measures the temperature of the coolant flowing into the core in real time via sensors and is a key factor in determining thermohydraulic conditions.

[0042] In relation to core flow rate, it can be directly measured or derived from auxiliary variables such as differential pressure and pump rotational speed, and is a key factor in determining thermo-hydraulic conditions.

[0043] Regarding core pressure, the pressure inside the core or pressure vessel is measured via sensors and is a key factor in determining thermohydraulic conditions.

[0044] Regarding the radial peak power factor (Fr) value (ratio of average power to hot rod power), it is derived based on the core power and is a key factor in determining thermohydraulic conditions. Fr is the ratio of hot rod power to core average power. Core power can be calculated using the measured core inlet and outlet temperatures and core flow rates, or alternatively, using the ex-core neutron flux.

[0045] These input information is used to calculate the DNBR PTL that serves as a standard for determining whether DL is reached and generating a reactor shutdown signal, by transmitting the DNBR PTL (output when DL is reached) based on an artificial neural network to the DNBR PTL evaluation unit (200).

[0046] The detailed configuration of the core operation information collection unit (100) according to the present embodiment is as follows.

[0047] Figure 2 shows the detailed configuration of the core operation information collection unit according to the present embodiment.

[0048] As illustrated in FIG. 2, the core operation information collection unit (100) is composed of a plurality of sensors and calculation modules for collecting or calculating neural network input values ​​for DNBR PTL evaluation, and includes a neutron flux measurement unit (110), an inlet temperature sensor unit (120), a flow rate calculation unit (130), a pressure sensor unit (140), and a radial peak output factor calculation unit (150).

[0049] The neutron flux measuring unit (110) is configured to measure the ex-core neutron flux or the in-core neutron flux at two or more levels in the axial direction (Z-axis), and the measured neutron flux distribution is used as an input representing the output distribution in the core axial direction (Z-axis).

[0050] The inlet temperature sensor unit (120) is a sensor that measures the temperature of the coolant in the lower core inlet in real time and is used as an input for DNBR PTL production.

[0051] The flow rate calculation unit (130) is a module that directly measures the core coolant flow rate or calculates the flow rate using auxiliary information such as differential pressure and pump rotation speed, and is used as an input for DNBR PTL production.

[0052] The pressure sensor unit (140) measures the pressure inside the core or pressure vessel in real time and is used as an input for DNBR PTL production.

[0053] The radial peak output factor calculation unit (150) calculates the radial peak output factor (Fr) based on the input output, and may apply a predefined maximum value or a conservative estimate if necessary.

[0054] For example, the core operation information collection unit (100) configures neural network input values ​​for DNBR PTL evaluation by reflecting the real-time operating status of the reactor. For example, an operating situation in which the reactor gradually increases from a 70% output level to 100% can be assumed.

[0055] At this time, the neutron flux measuring unit (110) measures the ex-core neutron flux from a plurality of detectors installed in the axial direction and obtains neutron flux distribution data at each level. This value is used to determine the distribution in the axial direction (Z-axis) of the core.

[0056] The radial peak output factor (Fr) calculation unit (150) determines an Fr value indicating how high the output at the high-temperature rod location is compared to the overall average output. If Fr is 1.75 at 100% output, this value is used as an input for the DNBR PTL calculation. This value may be a value measured by an external system, or a value determined conservatively.

[0057] Meanwhile, the inlet temperature sensor unit (120) measures that the temperature of the coolant flowing into the core is 290°C, which is a key factor in determining the thermal hydraulic conditions. As the core inlet temperature decreases, the cooling efficiency increases, the DNBR characteristics improve, and accordingly, the DNBR PTL tends to increase.

[0058] In other words, the lower the temperature, the better the heat transfer performance of the coolant becomes, leading to an increase in DNBR; consequently, the DNBR PTL is calculated to be larger. That is, the lower the core inlet temperature, the more safely the reactor can operate even at higher power outputs, so the PTL tends to increase.

[0059] The flow rate calculation unit (130) directly receives the measured flow rate or calculates, based on the pump rotation speed and differential pressure signal, that the flow rate of the coolant circulating in the core is, for example, 4500 kg / s. As the flow rate increases, the heat removal performance of the coolant improves, thereby improving the DNBR characteristics, and accordingly, the DNBR PTL increases. Conversely, when the flow rate decreases, the cooling efficiency decreases, and the DNBR PTL tends to decrease.

[0060] The pressure sensor unit (140) measures the pressure of the coolant in real time, for example, measuring a pressure of 15.5 MPa. The coolant pressure acts as a key factor in the calculation of the DNBR PTL, and as the pressure increases, the density of the coolant increases, reducing the likelihood of boiling and improving the DNBR characteristics, thereby increasing the PTL. On the other hand, if the pressure is low, the DNBR characteristics deteriorate, and the PTL may decrease.

[0061] Some of the factors input by the core operation information collection unit (100) in this embodiment include not only directly measured values ​​but also values ​​derived through calculation from other measured factors.

[0062] For reference, the core flow rate can be derived through a flow rate calculation formula combined with the flow coefficient and coolant density based on measurements from pump rotational speed and differential pressure sensors, or directly measured values ​​may be used. Core output can be calculated according to an energy conservation formula using the inlet temperature, coolant specific heat, and flow rate, and the core output can be calculated using neutron flux. The radial peak power factor (Fr) can be calculated by real-time synthesis of the three-dimensional power distribution, or a value conservatively set in the design may be used.

[0063] The method of deriving and using some of these input parameters from other measurements rather than directly measuring them can reduce sensor dependency and operating costs, improve input reliability by ensuring the possibility of correction under various operating conditions, and provide a lightweight evaluation system suitable for hardware implementation by simplifying the computational structure.

[0064] In addition, the Fr value includes the measured value, as well as a value set based on conservative design judgment or all configurable output ratio values.

[0065] For reference, Fr, calculated from neutron flux measurement results with a core average power of 100% and a hot rod position power of 113%, is 113 / 100=1.13. Using this value, the thermal load condition corresponding to Fr=1.13 is applied in the DNBR PTL calculation.

[0066] Accordingly, assuming a situation where the high-temperature rod can momentarily have a higher output, it is possible to design it to use a predefined maximum value such as Fr = 1.18 or Fr = 1.20 even if the operating condition is measured at 113%. This is a measure to secure an additional safety margin and is consistent with ensuring conservatism required by regulatory agencies or design standards.

[0067] That is, even if Fr is calculated to be 1.13 from real-time neutron flux measurements, a conservative Fr value such as 1.15 or 1.20 may be applied depending on operating conditions or design criteria.

[0068] In addition, by using a preset Fr table for each output range, Fr = 1.15 is automatically applied when the current output is 85%, making it possible to perform a stable and conservative DNBR PTL evaluation in preparation for an increase in the thermal load of the high-temperature rod.

[0069] As in this embodiment, by setting the Fr value not only to the actual value but also to a design-conservative maximum output ratio value or a predefined value, it is possible to always evaluate under the assumption of the worst-case conditions when calculating the DNBR PTL, thereby providing enhanced safety in preventing nuclear fuel damage and enabling conservative judgment even in situations of uncertainty or missing values ​​in the measurement.

[0070] The DNBR PTL evaluation unit (200) is configured to calculate the allowable power trip limit (PTL) until the DNBR limit value (DL) is reached in the current operating state using an artificial neural network model based on operating data input from the core operating information collection unit (100).

[0071] This configuration is designed as a fast and simple inference-based method using pre-trained neural networks instead of complex numerical analysis based on traditional thermal-hydraulic analysis codes, and is lightweight enough to be implemented even on FPGA / PLC-based hardware platforms.

[0072] The DNBR PTL evaluation unit (200) normalizes the input parameters transmitted from the core operation information collection unit (100) and inputs them into a pre-trained artificial neural network to infer the DNBR PTL. Subsequently, a correction value that considers the uncertainty of the model is applied to the inferred PTL to calculate a more stable predicted value for reaching DL.

[0073] Figure 3 shows the detailed configuration of the DNBR PTL evaluation unit according to the present embodiment.

[0074] As shown in FIG. 3, the DNBR PTL evaluation unit (200) configuration may consist of an input normalization unit (210), an artificial neural network operation unit (220), an output post-processing unit (230), and an uncertainty correction module (240), and has a structure that can be implemented in real time even on a hardware-based system such as an FPGA, taking into account inference speed and simplification.

[0075] The input normalization unit (210) is configured to scale or normalize input values ​​(pressure, flow rate, temperature, Fr, etc.) received from the core operation information collection unit (100) into a format that can be processed by an artificial neural network.

[0076] That is, the input normalization unit (210) performs the function of scaling or normalizing various input values, such as pressure, flow rate, temperature, and Fr transmitted from the core operation information collection unit (100), into a predefined range (e.g., 0 to 1) or a standard normal distribution (e.g., mean 0, variance 1) so that the artificial neural network model can process them stably. This process prevents the influence of scale differences between input values ​​or convergence delays during the neural network learning and inference process, and ensures the generalization performance of the model.

[0077] For reference, generalization performance refers to the ability of a model to accurately predict new inputs (such as driving conditions) that were not used in training. It refers to the ability of an artificial neural network to reliably predict DNBR PTLs under various conditions encountered in actual driving environments, rather than overfitting to the training data.

[0078] In other words, the generalization performance of the model referred to here means the ability of the artificial neural network to reliably predict DNBR PTL not only for the driving conditions used for training but also for various input conditions that may be encountered during actual driving.

[0079] Input normalization is a key preprocessing step for securing such generalization performance; it corrects for distortion in the relative importance of input values ​​and can improve the convergence stability and prediction reliability of the neural network.

[0080] The artificial neural network computation unit (220) is composed of an artificial neural network including one or more hidden layers and performs the function of inferring the DNBR PTL corresponding to the current operating condition based on the input value transmitted from the input normalization unit (210). This artificial neural network computation unit learns operating condition-output (DNBR) data generated from nuclear design codes and thermohydraulic design codes, and is trained into a model capable of inferring the DNBR PTL under various operating conditions in real time.

[0081] The artificial neural network may include non-linear functions such as ReLU and sigmoid or linear functions as activation functions, and is pre-trained based on vast training data related to core operating conditions.

[0082] The data used for training includes DNBR analysis results for various operating conditions generated from nuclear design codes and thermohydraulic analysis codes, and is optimized to learn the relationship between input values ​​such as hot rod output ratio, flow rate, pressure, and temperature and the DNBR response. As a result, the artificial neural network can effectively predict the Power Trip Limit (PTL) at which the DNBR limit value (DL) is reached under given conditions.

[0083] In addition, the present artificial neural network model includes linear or non-linear activation functions, and at least one optimization technique among simulated annealing, genetic algorithm, back propagation, ADAM, or ADAMAX may be applied to optimize the learning of the neural network. The term "optimization technique" refers to all available methods and is not limited to the techniques described above.

[0084] For reference, Simulated Annealing is a stochastic optimization technique and a metaheuristic global search algorithm. It gradually updates the solution in a direction that lowers the value of the energy function (error function), initially exploring various solutions with high probability before gradually converging. This has the advantage of allowing the algorithm to be designed to avoid getting stuck in a local optimum.

[0085] Simulated Annealing is a stochastic global optimization technique that can be applied during the optimization process of the artificial neural network operation unit (220). This technique is an algorithm that mathematically simulates the physical principle of the crystal structure converging to a stable state during the slow cooling process of a metal. It considers a set of parameters, such as the weights and bias values ​​of the artificial neural network, as an energy state, and generally uses the mean squared error (MSE) between the predicted DNBR PTL value and the true value as the energy function.

[0086] For example, in the initial stage, a randomly set of parameters is applied to the neural network to calculate the error value, and the energy of the current state is defined based on this. From this state, some parameters are randomly adjusted to generate new candidate solutions (new parameter sets), and the error at that point is calculated to produce a new energy. If a new solution has a lower error (a lower energy state), it is accepted, and the system transitions to the next state; however, even if the error increases slightly, the solution can be accepted with a certain probability. This probability is adjusted according to the current "temperature" parameter; initially, a high temperature allows for the acceptance of various solutions, but as iterations progress, the temperature is gradually lowered to prevent falling into a local optimum.

[0087] For example, if the error in the current state is 0.020 and increases slightly to 0.023 after applying a new parameter set, the algorithm can accept this if the temperature is still high. However, if the temperature has decreased as training progresses into the later stages, the algorithm does not accept this increase in error and maintains the previous state. In this way, the algorithm gradually explores the solution space and finds the optimal parameter set through a process of continuously reducing the error over hundreds to thousands of iterations.

[0088] In the present invention, by applying this simulated annealing technique, the artificial neural network is configured to converge to a generalized DNBR PTL prediction model for various operating conditions without falling into a local optimal value. In particular, when the input combination of operating conditions is complex and the interactions are non-linear, this technique can contribute to improving learning stability and reliability.

[0089] The genetic algorithm is applied to the optimization process of artificial neural networks, and its primary objective is to search for parameters such as weights and biases to converge to a global optimal solution.

[0090] For example, in the learning process of the artificial neural network computation unit (220), a genetic algorithm is applied to generate 100 sets of initial weight and bias parameters, and then the fitness is evaluated based on the mean squared error (MSE) between the predicted DNBR PTL and the DNBR PTL based on the thermal-hydraulic analysis code for each set. Subsequently, a top parameter set with high fitness is selected, and new generations are repeatedly generated through crossover and mutation processes to explore the globally optimal combination of parameters. The neural network optimized through hundreds of generations of iterative learning is configured to infer the DNBR PTL in real time under various operating conditions.

[0091] Here, for reference, the parameter set is a numerical set containing all weights and biases that constitute the learning of an artificial neural network, and is an element that numerically defines the connection relationships between the input layer and the hidden layer, and between the hidden layer and the output layer. In the present invention, by iteratively evolving this parameter set through a genetic algorithm, an optimal neural network model capable of stably predicting DNBR PTL under various operating conditions can be derived.

[0092] Backpropagation is a traditional neural network learning technique that calculates the error in the output layer and modifies the weights of each layer by backpropagating this error. Based on Gradient Descent, it is important to strike a balance between learning speed and error reduction. It is widely used in Multilayer Perceptron (MLP) structures.

[0093] For reference, the artificial neural network computation unit (220) of the present invention is a learning model for inferring DNBR PTL from the operating conditions (input values) of the core, and is configured based on a Multi-Layer Perceptron (MLP) structure including a hidden layer. At this time, during the learning process of the artificial neural network, a backpropagation algorithm is applied so that the learning is performed in a direction that minimizes the prediction error by repeatedly modifying the weights and bias values ​​of each layer.

[0094] For example, assume that one of the training data generated through nuclear design and thermohydraulic analysis codes is as follows. The input values ​​are pressure 15.5 MPa, flow rate 4800 kg / s, inlet temperature 290℃, and Fr value 1.18. The correct answer (DNBR reference value) is 1.31.

[0095] When this input value is fed into the neural network, an error may exist in the output DNBR prediction value, such as 1.22, because the initial weight state is random. In this case, the error in the output layer is calculated as follows.

[0096] Loss=(1.31-1.22) 2 =0.0081

[0097] The backpropagation algorithm calculates the contribution (partial derivative) of each neuron from the output layer to the input layer based on this error, and uses this to adjust each weight. Gradient Descent is applied at this stage, so that the weights are updated to reduce the error according to the learning rate.

[0098] By repeatedly applying this process to thousands of training datasets, the neural network optimizes its weights to accurately approximate the actual DNBR PTL value according to input conditions.

[0099] As a result of repeated training, the neural network becomes able to make predictions very close to the actual correct answer for the same input conditions, such as DNBR PTL = 1.30~1.31.

[0100] In the present invention, learning utilizing the backpropagation algorithm has the advantage of rapidly optimizing a relatively simple network structure. Considering computational efficiency and applicability on hardware platforms, the learning can be performed in an external learning environment (GPU / server, etc.), and the implementation can be achieved by loading only the final weights into the inference system.

[0101] ADAM (Adaptive Moment Estimation) is an optimization algorithm that dynamically adjusts the learning rate by considering both the first and second moments (moving average). Although it is based on gradient descent, it is effective for complex networks such as CNNs and RNNs due to improved learning stability and convergence speed.

[0102] The artificial neural network computation unit (220) of the present invention may apply the ADAM (Adaptive Moment Estimation) algorithm as an optimization technique during the learning process for inferring DNBR PTL from core operating conditions. While ADAM is based on gradient descent, it has the characteristic of automatically adjusting the learning rate (adaptive learning rate) for each parameter by simultaneously considering the first moment (moving average) and the second moment (variance) of the error gradient that occurs during the learning process.

[0103] For example, when an artificial neural network performs training with input values ​​such as pressure, flow rate, temperature, and Fr and corresponding DNBR target values, a loss function, that is, the difference between the predicted DNBR and the actual reference value, is calculated at each training epoch.

[0104] ADAM does not simply apply the gradient (∇L) of this loss function all at once, but calculates the average gradient (m) and average squared gradient (v) up to that point as follows.

[0105] m t =β1m t-1 +(1-β1)g t

[0106] v t =β2v t-1 +(1-β2)g t 2

[0107] At this time g t is the gradient at the current time point, and β1 and β2 represent the momentum weights, typically using 0.9 and 0.999.

[0108] Subsequently, by correcting these two moment information and applying different learning rates to each parameter, the neural network's learning speed can be fast while suppressing error oscillations, enabling stable convergence.

[0109] In particular, as in the present invention, when there are input values ​​for various operating conditions and the range of change or sensitivity of some variables such as Fr is large, learning efficiency and prediction accuracy are simultaneously improved because an appropriate learning rate is applied to each variable through ADAM.

[0110] In the present invention, through this ADAM-based learning, a stable DNBR prediction model can be constructed without overfitting even if the entire learning data is repeated for tens or hundreds of epochs, and the learned parameters are applied to the real-time inference structure of the DNBR PTL evaluation unit (200) and used to determine the shutdown of the reactor.

[0111] Regarding ADAMAX, the explanation is as follows. It is a variation of ADAM that uses the L∞ norm for optimization. In particular, it demonstrates stable learning rate control performance in sparse data or very large vector spaces.

[0112] ADAMAX is an extended form of the ADAM (Adaptive Moment Estimation) algorithm and is an optimization method that uses the L∞ norm (maximum value standard) instead of the L2 norm (mean square) as the denominator term for adjusting the learning rate. ADAMAX can be applied to the learning process for predicting DNBR PTL in the artificial neural network operation unit (220) of the present invention, which enables effective learning, particularly for sparse feature spaces where the range of change of input parameters is large or only some variables are sensitive.

[0113] For example, if the input values ​​such as flow rate, pressure, and Fr exhibit relatively large fluctuations, but under certain operating conditions a specific factor (e.g., inlet temperature) remains nearly constant or output changes occur only at certain levels of neutron flux, the parameter space used for learning may be partially sparse or unevenly distributed.

[0114] ADAMAX can stably maintain the learning rate based on the maximum gradient information of the variable, even for parameters where the average change is small or converges to zero in such situations, so that global parameter optimization of the neural network is possible without the phenomenon of learning stopping or convergence slowing down.

[0115] In the present invention, ADAMAX can be selectively applied according to the characteristics of the input space during the process of learning the relationship between various combinations of driving conditions and DNBR output values, thereby ensuring numerical stability of parameter updates and enabling the implementation of a DNBR PTL inference model capable of high-precision prediction even in complex input structures.

[0116] By utilizing various optimization techniques as described above, the artificial neural network model can rapidly and stably predict DNBR PTL according to changes in operating conditions, and improve error convergence speed and generalization ability, thereby increasing the reliability of the reactor protection system.

[0117] Although optimization techniques have been mentioned in this embodiment, the present invention is not limited thereto, and it is understood that various optimization techniques can be utilized even if not mentioned in the embodiment.

[0118] The output post-processing unit (230) converts the DNBR PTL prediction value calculated by the artificial neural network computation unit (220) into a real value applicable to the actual operation environment, and performs the function of generating a highly reliable DNBR PTL value by reflecting a correction coefficient or uncertainty correction value therein.

[0119] The generated corrected DNBR PTL value is compared with the core output value entered in advance and used as a criterion for determining whether to shut down the reactor.

[0120] As such, the output post-processing unit does not stop at simple denormalization (inverse scaling) but acts as an important computational module that derives key information for actual control decisions.

[0121] For example, even if a DNBR PTL value for a specific operating condition is derived from the neural network computation unit during reactor operation, it is difficult to use it directly for control decisions because it is output in a normalized form.

[0122] The output post-processing unit converts this value into a DNBR PTL value in actual physical units, and then applies a statistical correction value or a design consolidation factor in consideration of errors or uncertainties that may be included in the predicted value.

[0123] The final calculated corrected DNBR PTL value is compared with the core output, and based on the result, it is possible to determine whether to continue operating the reactor or to induce an immediate shutdown.

[0124] As such, the output post-processing unit goes beyond the simple restoration of predicted values ​​and plays a key role in ensuring the reliability and safety of reactor protection judgments.

[0125] The uncertainty correction module (240) performs the function of calculating a highly reliable DNBR PTL value by reflecting uncertainties such as prediction model uncertainty error, measurement uncertainty, and manufacturing tolerance to the DNBR PTL value output from the output post-processing unit (230) to ensure sufficient conservatism in core protection.

[0126] The DNBR PTL determined through this correction procedure is set sufficiently conservatively so that it satisfies the requirements required in the event of a reactor shutdown when compared to the output, thereby establishing a more conservative and reliable reactor protection standard.

[0127] For example, even if a neural network calculates the DNBR PTL under specific operating conditions, there may be generalization errors or data deviations in the value. In this case, the uncertainty correction module applies a standard deviation-based correction coefficient by referring to the error distribution under similar past conditions, and additionally reflects a compensation coefficient to adjust the DNBR PTL in a safer direction.

[0128] This process increases the reliability of neural network-based reasoning logic and can simultaneously satisfy the safety and regulatory responsiveness of actual reactor protection systems.

[0129] The shutdown decision unit (300) is configured to perform the function of ultimately determining whether to continue operating the reactor or to perform an immediate shutdown based on the comparison result of the current core output and the DNBR PTL transmitted through the uncertainty correction module (240).

[0130] This configuration compares the current output with the DNBR PTL value corrected in the uncertainty correction module (240), and if the current output is equal to or greater than the DNBR PTL value, it determines that the DNBR value has reached or is lower than the DNBR limit value (DL), and serves as a reference point for determining this as a condition for reactor shutdown.

[0131] Accordingly, the shutdown judgment unit (300) is configured to output a control command for shutting down the reactor as a final judgment step to ensure the safety of the reactor protection system when the current operating output is higher than the calculated DNBR PTL, that is, when there is a possibility that the DNBR will reach the limit value (DL).

[0132] That is, the shutdown determination unit (300) generates a reactor shutdown signal when the current operating output value collected from the core operation information collection unit (100) is higher than the calculated PTL.

[0133] This configuration performs a final logic decision function that determines whether to stop based on the possibility of reaching DL by comparing the PTL value predicted by the neural network model with the current operating output and generating a stop signal in situations where there is a risk of reaching DL.

[0134] Figure 4 shows the detailed configuration of the stop judgment unit according to the present embodiment.

[0135] As shown in FIG. 4, the stop judgment unit (300) is configured with a PTL judgment comparison unit (310) and a stop command generation unit (320).

[0136] The PTL judgment comparison unit (310) compares the corrected DNBR PTL value transmitted from the uncertainty correction module (240) with the current core output and logically determines whether the criteria are satisfied. If the DNBR PTL value is greater than or equal to the current output, it is considered normal operation, and if it is less than or equal to the current output, it is determined as a shutdown condition and performs the function of generating a control command for shutting down the reactor.

[0137] That is, the PTL judgment comparison unit (310) compares the corrected DNBR PTL value transmitted from the uncertainty correction module (240) with the current core output, and determines normal operation if the PTL value is greater than the current output, and determines a shutdown condition if the current output exceeds the PTL.

[0138] In this case, the PTL judgment comparison unit (310) recognizes that the stop judgment condition is satisfied and performs the function of transmitting a signal to the stop command generation unit (320) for generating a control command for stopping the reactor.

[0139] The stop command generation unit (320) performs the function of generating and outputting a control command for reactor shutdown when the shutdown condition is finally satisfied according to the judgment result of the PTL judgment comparison unit (310). The generated control command can be interfaced in the form of a digital signal, flag setting, status register switching, etc., and the command is configured to be linked with the actual reactor control system to perform a system shutdown operation.

[0140] The means for constituting the present invention are described as follows.

[0141] To implement the present invention, the core inlet temperature, pressure, flow rate, output, and at least two levels of in-core or out-core neutron flux must be measured. The core flow rate can be measured by an instrument, and separate correction may be performed if necessary. The core output can be calculated using the measured core inlet temperature, outlet temperature, core pressure, and flow rate; alternatively, it can be measured or calculated using the out-core neutron flux. The Fr value may vary depending on the output and is designed to take a value proportional to the measured or calculated output.

[0142] For the development of the model, design values ​​for model optimization are required. The design values ​​are the respective DNBR PTLs calculated using nuclear design and thermal-hydraulic design codes under various operating conditions of the core protection system.

[0143] The DNBR PTL model uses core inlet temperature, pressure, Fr, and in-core or ex-core neutron flux signals of at least two levels as inputs, and is composed of an artificial neural network comprising one or more hidden layers; the output of the artificial neural network becomes the DNBR PTL value. Linear or non-linear activation functions can be used for the artificial neural network. Optimization of the artificial neural network is performed using various artificial neural network optimization techniques used in AI, ensuring that the difference between the nuclear design and the thermohydraulic design code is minimized.

[0144] The overall details regarding the configuration and operation according to the detailed description of the invention are as follows.

[0145] To calculate the DNBR PTL (i.e., the output when the DNBR value reaches DL), the core power distribution, core inlet temperature, core flow rate, core pressure, and output are used.

[0146] In the case of the innovative SMR, since the core power distribution is not directly synthesized, the overall average power distribution of the core reflects 2 or more levels of in-core neutron flux or ex-core neutron flux, and the power ratio (Fr) of the hot rod relative to the core average is taken from the pre-set Fr ratio for each power.

[0147] In addition, when calculating direct DNBR values, the DNBR values ​​appear in various ranges depending on the output, but this increases the uncertainty of the calculated DNBR values.

[0148] Since the core protection system must cause a reactor shutdown before the DNBR value reaches DL, the output is not used as a direct DNBR calculation model, but the output is modified to set the output (PTL) at which the DNBR value reaches DL as the target value.

[0149] That is, the system according to the present invention calculates the output when DL is reached using the output distribution, core inlet temperature, core flow rate, and Fr value, and establishes a methodology to prevent damage to the nuclear fuel cladding by generating a reactor shutdown signal when the current output reaches the output at which DL is reached.

[0150] FIG. 5 is an example of applying Fr by output of a core protection system using a DNBR PTL evaluation model based on a universal approximation theorem according to an embodiment of the present invention, wherein a maximum Fr is set for each section and expanded using a linear interpolation method.

[0151] In Figure 5, the x-axis represents output (%) and the y-axis represents the Fr value (ratio of average output to high-temperature rod output), and the graph shows a trend in which the Fr value gradually decreases as output increases. This reflects the phenomenon in actual reactors where the difference in output between the high-temperature rod and the average rod is large during low-power operation, while conversely, as output increases, the output distribution becomes flatter and the Fr value decreases.

[0152] In the present invention, the radial peak output factor (Fr) value used for DNBR PTL evaluation is not used as a single fixed value, but is extended and applied using a linear interpolation method based on a maximum Fr table defined for each section according to output conditions.

[0153] This means that when the actual operating output is located between specific predefined output points (e.g., 20%, 40%, 60%, etc.), a suitable Fr value for the operating condition is calculated by linearly interpolating based on the Fr value between two adjacent points in the table.

[0154] In other words, it is interpreted as a structure where the relative output of the high-temperature rod is higher as the output decreases, and as the output increases, a relatively equilibrium output distribution is formed, causing the Fr value to decrease.

[0155] In the present invention, a maximum Fr value for each section is defined based on such an output-frame relationship, and the Fr value is calculated using a linear interpolation method for the actual output value according to operating conditions, thereby enabling a conservative yet flexible DNBR PTL evaluation.

[0156] This allows Fr values ​​to be effectively applied across the entire output range using only a few defined representative output conditions.

[0157] An artificial neural network model is utilized to apply the general approximation theorem technique for calculating the DNBR Power Trip Limit (PTL) of the core protection system. The artificial neural network model requires one or more hidden layers.

[0158] FIG. 6 is a model for calculating a DNBR PTL having one hidden layer of a core protection system using a universal approximation theorem-based DNBR PTL evaluation model according to one embodiment of the present invention.

[0159] FIG. 6 illustrates the structure of a DNBR PTL evaluation model based on the universal approximation theorem according to one embodiment of the present invention, which is a multilayer perceptron-based artificial neural network structure including one hidden layer.

[0160] The input layer consists of major operating variables of the core (ex-core neutron flux or in-core neutron flux of 2 levels or more, core inlet temperature, core flow rate, core pressure, Fr, etc.) and is transmitted after being normalized through the neural network input normalization unit (210), and in the hidden layer, the input values ​​and weight α ij The results computed through this are used to approximate complex relationships between variables using non-linear activation functions (ReLu, Sigmoid, etc.), and in the output layer, the intermediate results transmitted from the hidden layer are processed by the output layer weight β j The final DNBR PTL is calculated by performing a weighted sum process. In this diagram, eight hidden nodes are configured as an example, but the actual structure can be varied depending on the optimization process, and the activation function and optimization method can also be varied depending on the situation.

[0161] This model (an artificial neural network model for DNBR PTL calculation) features a simple computational structure that enables real-time prediction under various operating conditions and is configured to be applied to hardware platforms in a lightweight form.

[0162] This diagram illustrates the basic structure of an artificial neural network model for calculating DNBR PTL (Power Trip Limit), showing a multilayer perceptron (MLP) based structure consisting of an input layer, a hidden layer, and an output layer.

[0163] In particular, this model is characterized by a structure with one hidden layer, designed to enable rapid inference of DNBR PTL under various operating conditions based on the Universal Approximation Theorem.

[0164] The Universal Approximation Theorem is a theory that states that even an artificial neural network with only one hidden layer can approximate any continuous non-linear function with desired accuracy if it has a sufficient number of hidden nodes and appropriate weights.

[0165] Therefore, the artificial neural network computation unit of the present invention is configured to effectively learn and predict the non-linear relationship between complex thermal-hydraulic conditions and DNBR using only a single hidden layer structure.

[0166] This prediction model can be expressed by the following formula.

[0167]

[0168] This formula mathematically represents the neural network-based function approximation structure for DNBR PTL prediction of the present invention. It is a computational form of a Multilayer Perceptron (MLP) model having one hidden layer.

[0169] x i is the input value (e.g., ex-core neutron flux or in-core neutron flux of 2 levels or higher, core inlet temperature, core flow rate, core pressure, Fr, etc.), α ij is the weight between the input layer and the hidden layer (input → hidden node connection weight), b ij ε is the bias value for hidden layer node j, f() is the activation function (e.g., ReLU, Sigmoid, etc.), β j is the weight between the hidden layer and the output layer (hidden node → output connection weight), c j ε is the output layer bias, and DNBR PTL is the output reference value inferred by the neural network, representing the reactor output at the point when DNBR reaches the limit value (DL) under current operating conditions.

[0170] Each input value x of the input layer i is weight α ijIt is combined with and passed to the hidden node. In the hidden node, the result of their linear combination is biased b ij We add and apply the activation function f to perform non-linear processing. The result calculated in the hidden layer is then applied again to the weight β j It is combined with and passed to the output node, where the activation function f is applied once again and the output bias C j The final DNBR PTL (Power Trip Limit) is calculated by adding the above.

[0171] Here, various linear and non-linear functions can be used for the activation function f(x), and generally, linear functions of the form ax+b or non-linear functions such as Relu or sigmoid can be used for the hidden layer and output layer, and the form of the function is not limited in the present invention.

[0172] In safety analysis, the analysis is performed based on nuclear and thermohydraulic design models. Therefore, for the target value (DNBR PTL) of this model, the input and target values ​​are constructed using the calculation results of the nuclear and thermohydraulic design codes produced under the conditions (ex-core neutron flux or in-core neutron flux, core inlet temperature, core flow rate, core pressure, Fr) determined within the safety analysis range.

[0173] In addition, core flow rate, core inlet temperature, and core pressure are derived by dividing them evenly within the operating range. The target value (DNBR PTL) is selected as the output value when the DL calculated in the thermal-hydraulic design code is reached by executing the thermal-hydraulic design code using the output distribution of the core design code and the derived core flow rate, core inlet temperature, and core pressure.

[0174] The formula optimization method of DNBR PTL utilizes optimization methods used in artificial neural network optimization learning, and uses the same conditions used as input to the thermal-hydraulic design code to minimize computational uncertainty so that a ij , β ij , bij and c ij Optimization is performed so that it is determined.

[0175] Regarding the reactor shutdown method of the core protection system, the trip condition of the system is defined as when the value of the DNBR PTL corrected by the uncertainty factor is lower than the current output, as follows.

[0176] Trip condition: (DNBR PTL / (Uncertainty 1 - Uncertainty 2)) < Current Output

[0177] The DNBR PTL of the core protection system is the output when the DNBR value reaches DL, and uncertainty is a penalty to make the core protection system operate conservatively under various conditions.

[0178] The above method is applied to design a conservative DNBR reactor shutdown of the core protection system.

[0179] FIG. 7 is a flowchart illustrating a method using a core protection system based on a universal approximation theorem-based DNBR PTL evaluation model according to one embodiment of the present invention.

[0180] First, a core protection system using a universal approximation theorem-based DNBR PTL evaluation model (hereinafter referred to as the system) collects information on ex-core neutron flux or in-core neutron flux having multiple axial levels, core inlet temperature, pressure, flow rate and power, and radial peak power factor (Fr) (a).

[0181] Next, the system inputs the collected core operation information into an artificial neural network model and calculates the DNBR PTL through an artificial neural network including one or more hidden layers (b).

[0182] Here, in step (b), the step of inputting collected core operation information into an artificial neural network model including one or more hidden layers to calculate the output value (DNBR PTL) at the time when the DNBR reaches the limit value (DL) under current operating conditions is a process of inferring the output value (Power Trip Limit) at the time when the DL is reached in real time by inputting the collected operation variables (ex-core neutron flux or in-core neutron flux, inlet temperature, pressure, flow rate, output, radial peak power factor (Fr)) into a pre-trained artificial neural network and non-linearly approximating the complex thermo-hydraulic correlation.

[0183] This step enables fast predictions that can be implemented on FPGA or PLC-based hardware by using a lightweight neural network model instead of direct calculation of the thermal-hydraulic design code, and plays a key role in determining whether DL is reached in real time.

[0184] Next, the system generates a reactor shutdown signal when the current operating output is higher than the calculated DNBR PTL, that is, when it is determined that there is a risk that the DNBR will reach the limit value (DL) (c).

[0185] Here, the artificial neural network model is pre-trained based on learning data derived from nuclear design codes and thermal-hydraulic analysis codes under various operating conditions of the core protection system, and is configured to infer the output value at the time when the reactor output reaches the DNBR limit value (DL) as the DNBR PTL.

[0186] The phrase 'when the current operating output is high' in step (c) above means that the current operating output is greater than or equal to the DNBR PTL value inferred by the artificial neural network model, and this is a condition indicating that the DNBR has reached or is about to reach the limit value (DL), and serves as a criterion for generating a reactor shutdown signal.

[0187] The artificial neural network model in this embodiment is pre-optimized based on learning data derived from nuclear design codes and thermal-hydraulic design codes under various operating conditions of the core protection system, and is configured to evaluate the DNBR PTL based on the condition that the reactor power reaches the DNBR limit value (DL).

[0188] Here, 'pre-optimization' refers to the process of training the weights and biases of a neural network in advance by utilizing the simulation results of thermohydraulic analysis codes and nuclear design codes for various operating conditions (pressure, flow rate, inlet temperature, power distribution, radial peak power factor (Fr), etc.) as training data before directly using actual reactor operation data.

[0189] Such prior learning provides the ability to generalize and predict DNBR PTL for various input conditions in actual driving environments, and through learning based on the time of reaching DL, ensures that DNBR PTL calculation is always performed in accordance with reference values ​​that consider safety margins.

[0190] Therefore, the model is designed to evaluate the possibility of reaching DL in real time even in situations of sudden changes in driving conditions and to provide immediate control judgment criteria to the stop judgment unit (300).

[0191] According to the present invention, in order to prevent damage to the nuclear fuel cladding, the DNBR PTL (Power Trip Limit), which is the output value at the time when the DNBR reaches the limit value (DL) under current operating conditions, is predicted, and if a state of output higher than the PTL is confirmed, the reactor is shut down, thereby providing an effect of preventing damage to the nuclear fuel cladding.

[0192] That is, the core protection system using the DNBR PTL evaluation model based on the universal approximation theorem according to the present embodiment determines whether safe operation is possible by comparing the current operating output with the PTL, which is the output value at the time of reaching DL, and shuts down the reactor if the PTL is exceeded.

[0193] By applying a single hidden layer-based neural network structure using the present invention, DNBR PTL can be predicted quickly and stably even under various operating conditions.

[0194] Through a lightweight structure that facilitates hardware implementation, it can be applied to FPGA or PLC-based systems, enabling real-time control.

[0195] In addition, safety can be ensured and the reliability of stop judgments can be improved simultaneously through uncertainty correction and the application of compensation factors.

Claims

Claim 1 A core protection system using a universal approximation theorem-based DNBR PTL evaluation model comprises: a core operation information collection unit (100) that collects values ​​of ex-core neutron flux or in-core neutron flux having multiple axial levels, core inlet temperature, pressure, flow rate, output, and radial peak power factor (Fr); a DNBR PTL evaluation unit (200) that takes the collected core operation information as input, directly calculates a DNBR PTL—which is an output condition reaching a DNBR limit value (DL)—using a single hidden-layer-based universal approximation theorem model, and performs conservatism reflection to further restrict the operable range by correcting the calculated DNBR PTL value with an uncertainty factor; and a shutdown determination unit (300) that generates a reactor shutdown signal when the current operating output value collected from the core operation information collection unit (100) is higher than the DNBR PTL reflected in conservatism. The universal approximation theorem-based DNBR PTL evaluation model is from nuclear design and thermohydraulic design codes under various operating conditions of the core protection system. A core protection system using a universal approximation theorem-based DNBR PTL evaluation model, characterized by being optimized based on derived data and configured to evaluate the DNBR PTL based on the point in time when the reactor output reaches the DNBR limit value (DL). Claim 2 A core protection system using a universal approximation theorem-based DNBR PTL evaluation model, characterized in that, in claim 1, the PTL, which is the output condition for reaching the DNBR limit value (DL), is calculated based on one or more computer codes utilizing a simplified numerical model as well as a thermo-hydraulic design code. Claim 3 A core protection system using a universal approximation theorem-based DNBR PTL evaluation model, wherein, in claim 1, the core operation information collection unit (100) collects some of the input factors including not only directly measured values ​​but also values ​​derived through calculation from other measured factors. Claim 4 A core protection system using a universal approximation theorem-based DNBR PTL evaluation model according to claim 1, characterized in that the radial peak power factor (Fr) may include not only a measured value but also a value pre-set according to design maintenance judgment or various power ratio values ​​that can be set in response to operating conditions. Claim 5 A core protection system using a universal approximation theorem-based DNBR PTL evaluation model according to claim 1, wherein the DNBR PTL evaluation unit (200) receives input values ​​collected from the core operation information collection unit (100) in real time during reactor operation, infers the DNBR PTL in real time based on the input values, and the shutdown determination unit (300) includes a real-time response structure that outputs a reactor shutdown signal immediately when the current operating output exceeds the inferred PTL. Claim 6 A core protection system using a universal approximation theorem-based DNBR PTL evaluation model according to claim 1, wherein the DNBR PTL evaluation unit (200) includes an uncertainty correction module (240) that applies a correction coefficient or a repair coefficient to the DNBR PTL value calculated by the universal approximation theorem-based DNBR PTL evaluation model in response to the generalization error of the prediction model, the measurement error of the operating data, or the manufacturing tolerance, and the shutdown determination unit (300) is configured to determine whether to shut down the reactor based on the DNBR PTL value corrected through the uncertainty correction module (240). Claim 7 A method using a core protection system utilizing a universal approximation theorem-based DNBR PTL evaluation model comprises: (a) collecting ex-core neutron flux or in-core neutron flux having multiple axial levels, core inlet temperature, pressure, flow rate, power information, and radial peak power factor (Fr) values; (b) using the collected core operation information as input, directly calculating the DNBR PTL, which is a power condition reaching the DNBR limit value (DL), by a single hidden-layer-based universal approximation theorem model, and performing conservatism by correcting the calculated DNBR PTL value with an uncertainty factor to further restrict the operable range; and (c) generating a reactor shutdown signal when the current operating power collected in step (a) is higher than the DNBR PTL with conservatism applied, wherein the universal approximation theorem-based DNBR PTL evaluation model is pre-optimized based on learning data derived from nuclear design codes and thermohydraulic design codes under various operating conditions of the core protection system, and when the reactor power reaches the DNBR limit value (DL) A core protection method using a universal approximation theorem-based DNBR PTL evaluation model characterized by being configured to evaluate DNBR PTL based on conditions reached.

Images