Method for operating a nuclear reactor using online calculation of DNBR and corresponding nuclear reactor

Abstract

The method for operating a nuclear reactor with online calculation of DNBR and the corresponding nuclear reactor method include the following steps: ‑ obtaining a plurality of quantities characterizing the operation of the nuclear reactor; ‑ calculating at least one deviation from nucleate boiling ratio by using a deep neural network, the entries of the deep neural network being determined using the obtained quantities, the deep neural network comprising at least two hidden layers, each hidden layer having at least five neurons; ‑ calculating the deviation between at least one calculated deviation from nucleate boiling ratio and a plurality of predetermined reference thresholds; ‑ formulating a control signal for a reactor control system by using the calculated deviation, the control signal being: * automatic reactor shutdown or alarm; * doing nothing; ‑ emergency shutdown of the nuclear reactor or, if relevant, issuing an alarm signal.

Description

[Technical field]

[0001] The present invention generally relates to protecting and / or monitoring the core of a pressurized water nuclear reactor to prevent a boiling crisis. [Background technology]

[0002] Maintaining the integrity of the primary safety barrier (the fuel cladding) is a major goal in nuclear reactor design and is the subject of safety demonstrations performed, among other things, by simulating the behavior of the reactor during various unexpected transients using numerical calculation codes.

[0003] The analysis is all documented in the facility’s final safety report, which is required for nuclear safety authorities to grant authorization for the reactor to operate.

[0004] The integrity of the fuel cladding is ensured by the absence of fusion in the center of the fuel pellet and by the absence of boiling crises (Departure from Nuclear Boiling, DNB) on the surface of the cladding. These conditions must be met at all points in the core.

[0005] The boiling crisis is characterized by a degradation of the heat exchange between the heat transfer fluid of the primary circuit and the surface of the fuel cladding, which degradation may lead to a calefaction phenomenon and a possible loss of integrity of the first barrier.

[0006] Each nuclear reactor is equipped with a protection system whose function is to guarantee the safety of the reactor in a certain number of unexpected situations. The protection system generally includes a protection chain called low DNBR (Departure from Nuclear Boiling Ratio). In reactors with four primary circuits, such as 1300MWe, N4 and EPR, the low DNBR protection chain ensures the protection of the nuclear reactor from boiling crisis phenomena during class 2 unexpected transients, which are characterized by the insertion of reactivity with moderate kinetics (uncontrolled withdrawal of the control unit, untimely dilution of the primary fluid leading to a reduction in the boron concentration, untimely cooling of the primary circuit, etc.).

[0007] To this end, the protection system includes an information processing unit (called Digital Integrated Protection and Control System or DIPC) on 1300MWe and N4 reactors, which integrates a functional unit called "Low DNBR algorithm". This algorithm calculates the safety margin with respect to boiling crises online. To this end, the algorithm calculates the DNBR and compares the calculated value with a limit value that takes into account the uncertainty of the reactor design and the uncertainty of the reconstruction of the DNBR value. If the reconstructed value is below the limit, the low DNBR protection chain causes an automatic shutdown of the reactor.

[0008] The deviation from nucleate boiling ratio or DNBR is defined as the ratio between the critical heat flux and the local heat flux. The local heat flux corresponds to the heat flux emitted by the fuel rod through the cladding. The critical heat flux is the value of the heat flux that causes damage to the fuel rod cladding, which depends on the local thermo-hydraulic conditions of the fluid.

[0009] The low DNBR algorithm uses a simplified thermohydraulic model of the core which is adjusted at the design stage by correction functions (bias curves) acting on the local thermohydraulic quantities in order to ensure a conservative response to the DNBR calculated by the 3D thermohydraulic reference code.

[0010] This simplified thermo-hydraulic modeling enables the calculation of local heat fluxes from measured quantities characterizing the power distribution of the reactor and allowing the evaluation of critical heat fluxes.

[0011] In addition, the low DNBR algorithm uses a predetermined critical flux correlation to calculate the critical heat flux. This correlation is determined empirically based on experimental data. The design of this correlation represents a complex and expensive development. The multi-dimensional polynomial type analytical formula for this correlation is assumed a priori and is manually adjusted in successive steps during the development process in order to meet certain statistical criteria.

[0012] The low DNBR calculation algorithms are not very accurate and therefore large margins must be allowed to guarantee safety during nuclear reactor operation. This is very disadvantageous for reactor control. In addition, despite the simplification, thermal-hydraulic modeling requires complex numerical processing using microprocessor systems that allow the algorithms to converge with a time compatible with nuclear reactor protection.

[0013] Against this background, the object of the present invention is to propose a method for operating a nuclear reactor which does not have the above-mentioned disadvantages. [Summary of the invention]

[0014] To this end, the invention relates to a method for operating a nuclear reactor comprising a core, the method comprising the following steps:

[0015] - acquiring a plurality of quantities characterizing the operation of the nuclear reactor;

[0016] - calculating at least one deviation from nucleate boiling ratio using a deep neural network, entries of the deep neural network being determined using the acquired quantities, the deep neural network comprising at least two hidden layers, each hidden layer having at least five neurons;

[0017] - calculating a deviation between at least one calculated deviation from nucleate boiling ratio and a plurality of predetermined reference threshold values;

[0018] - using the calculated deviation to generate a control signal for a reactor control system, the control signal being selected from a list comprising at least the following values:

[0019] * Reactor automatic shutdown or alarm;

[0020] * Do nothing;

[0021] - When the generated control signal has the value "reactor automatic stop", the nuclear reactor is emergency stopped by the control system, or when the generated control signal has the value "alarm", an alarm signal is issued by the control system.

[0022] The main drawback of the low DNBR algorithm is that the local thermo-hydraulic forces are reconstructed using a simplified physical model (single-channel, single-phase model) with no thermo-hydraulic correlations other than critical flux correlations (pressure drop, porosity, etc.).

[0023] The use of deep neural networks enables the determination of the departure from nucleate boiling ratio from current values ​​of limiting quantities characterizing nuclear reactor operation, more accurately and faster than low DNBR algorithms implemented today.

[0024] The values ​​predicted by the deep neural network are much closer to those calculated by the reference thermal-hydraulic code. The DNBR reconstruction error is significantly reduced. As a result, the operating margin of the nuclear reactor is much better.

[0025] The operating method may also have one or more of the following characteristics, considered alone or in any technically possible combination:

[0026] - the acquired quantities include neutron flux measurements made by a neutron flux detector, the neural network entries comprising an axial power distribution in a core of the nuclear reactor reconstructed from the acquired neutron flux measurements;

[0027] - the axial power distribution is a vector having a plurality of components, each component corresponding to an average power of the nuclear reactor for each axial slice at a given axial position, each component defining one of the entries of the deep neural network;

[0028] - Neutron flux detectors are ex-core detectors arranged outside the core;

[0029] - Neutron flux detectors are in-core detectors permanently inserted in the core;

[0030] - a nuclear reactor comprising: a pressure vessel containing a core and at least one primary circuit, the primary circuit comprising a hot branch and a cold branch directly fluidically connected to the vessel; a pressurizer controlling the pressure in the primary circuit; and a main pump ensuring the circulation of a primary fluid in the primary circuit, the quantities acquired comprising, for the primary circuit or each primary circuit, one or more of the following quantities: pressurizer pressure, primary fluid flow in the primary circuit, temperature of the primary fluid in the hot branch and in the cold branch;

[0031] - the neural network entries include one or more of the following quantities: pressure of the pressurizer, flow rate of the primary fluid in the primary circuit, temperature of the primary fluid in the cold branch;

[0032] - The neural network entry includes the thermal power released by the core;

[0033] - determining the released thermal power by calculation using at least the pressurizer pressure, the primary fluid flow rate, the temperature of the hot fluid and the primary fluid in the cold branch obtained for at least one primary circuit;

[0034] - Determine the released thermal power by calculation using the acquired neutron flux measurements;

[0035] - The entries of the neural network include enthalpy rise factors;

[0036] - The acquired quantities include the current insertion position of the core reactivity control device, and the enthalpy rise factor is determined by performing the following operations:

[0037] - using the acquired current insertion position to calculate the radial power crest factor;

[0038] - calculating an enthalpy rise factor using the calculated radial power crest factor, the reconstructed axial power distribution and the determined thermal power output;

[0039] - Use the axial power distribution to reconstruct the enthalpy rise factor;

[0040] - the nuclear reactor comprises a plurality of primary circuits, and a deviation from nucleate boiling ratio is calculated for each primary circuit using a deep neural network using at least the quantity obtained for each primary circuit;

[0041] - the detectors are distributed in a number of groups, the detectors of the same group being staggered vertically above each other, a minimum deviation from nucleate boiling ratio being calculated for each group of in-core neutron flux detectors using a deep neural network, the entries of the deep neural network being determined using at least neutron flux measurements made by the in-core neutron flux detectors (31) of said group;

[0042] - in a deviation calculation step, calculating a set of deviations between each calculated deviation from nucleate boiling ratio and a plurality of predetermined reference values;

[0043] - in the step of generating a control signal, a temporary control signal is generated for each calculated deviation from nucleate boiling ratio using a corresponding set of deviations, the temporary control signal being selected from a list comprising at least the following values:

[0044] * Reactor automatic shutdown or alarm;

[0045] * Do nothing;

[0046] elaborate the control signal by using at least several temporary control signals;

[0047] -The deviation from nucleate boiling ratio is calculated by a deep neural network using the following formula:

[0048]

[0049] where Y is the calculated deviation from nucleate boiling ratio;

[0050] X0 is the entry vector;

[0051] nH is the number of hidden layers of the deep neural network;

[0052] A i It is scheduled N i *N i-1 N-dimensional matrix i , where N i is the number of neurons in layer i, and N i-1 is the number of neurons in layer i-1;

[0053] σ i is a predetermined nonlinear operator;

[0054] -A i is the vector X i-1 The components of the matrix act as an affine transformation, the vector X i-1 is a vector whose components are determined by neuron layer i-1;

[0055] -σ i Vector A i .X i-1 All components of act as sigmoid functions, and the vector X i-1 is a vector whose components are determined by neuron layer i-1;

[0056] -The method comprises a deep neural network design step, the design step comprising the following sub-steps:

[0057] - constructing a database comprising at least 100,000 reactor core states, each state being defined by a set of values ​​of said magnitude characterizing the operation of the nuclear reactor and a minimum deviation from nucleate boiling ratio calculated by a reference 3D thermal-hydraulic calculation code using the set of values, said states being selected to cover a plurality of predetermined contingencies of category 2, the reactor control system being provided to handle said contingencies;

[0058] - using at least a portion of the database to learn a deep neural network;

[0059] - Validation of the deep neural network by using a part of the database independent of that used for the learning phase, and testing of the neural network obtained by coupling with the accidental neutron and thermal-hydraulic codes, which make it possible to simulate unexpected transients requiring the intervention of the control system.

[0060] According to a second aspect, the invention relates to a nuclear reactor comprising:

[0061] - Reactor core;

[0062] - control systems;

[0063] - an information processing unit configured to implement an operating method having the above-mentioned characteristics.

Brief Description of the Drawings

[0064] Other features and advantages will emerge from the following detailed description given as an indication and not as a limitation, with reference to the accompanying drawings, in which:

[0065] Figure 1 is a diagram showing the steps of the method of the present invention;

[0066] Figure 2 is intended for implementation Figure 1 A simplified schematic diagram of a nuclear reactor according to the method;

[0067] Figure 3 The first embodiment of the present invention Figure 1 Schematic diagram of the acquisition and calculation steps of the method in;

[0068] Figure 4 is Figure 1 A schematic diagram of the deep neural network used in the method; and

[0069] Figure 5 The second embodiment of the present invention Figure 1 Schematic diagram of the acquisition and calculation steps of the method. [Specific implementation method]

[0070] Figure 1 The operating method schematically shown in Figure 2A nuclear reactor of the type shown is implemented.

[0071] The nuclear reactor is typically a pressurized water reactor (PWR).

[0072] Alternatively, the reactor is intended for use in any other suitable nuclear reactor.

[0073] The nuclear reactor 1 has a core 3 .

[0074] The core 3 is generally composed of a plurality of prismatic nuclear fuel assemblies arranged one next to the other. Each nuclear fuel assembly comprises a plurality of nuclear fuel rods which are held in position relative to each other by a grid.

[0075] Each rod has a tubular cladding made of a metal such as zirconium-tin alloy, within which nuclear fuel pellets are stacked.

[0076] The nuclear reactor 1 further comprises a pressure vessel 5 containing a core 3 and at least one primary circuit 7 .

[0077] Typically, a nuclear reactor has several primary circuits 7 , for example three or four primary circuits 7 .

[0078] The primary circuit 7 or each primary circuit 7 comprises a hot branch and a cold branch 9, 11 which are directly fluidically connected to the pressure vessel 5, a pressurizer 13 which controls the pressure in the primary circuit 7, and a main pump 15 which provides circulation of the primary fluid in the primary circuit 7. The pressurizer 13 is usually common to different primary circuits.

[0079] Typically, the primary circuit 7 also has a steam generator 17 upstream of the main pump 15 .

[0080] The hot branch 9 fluidly connects the primary fluid outlet manifold from the pressure vessel 5 to the primary fluid inlet tank of the steam generator 17. The pressurizer 13 is connected to the hot branch 9.

[0081] The intermediate branch 19 of the primary circuit, usually called "U branch", connects the primary fluid outlet tank of the steam generator 17 to the suction of the main pump 15. The cold branch 11 connects the outlet of the main pump 15 to the primary fluid inlet of the pressure vessel 5.

[0082] The nuclear reactor 1 further includes a core reactivity control device 21 .

[0083] These control means 21 are generally called "control beams" and comprise a plurality of rods made of neutron absorbing material.

[0084] The nuclear reactor 1 further comprises mechanisms, not shown, which are configured to selectively move the control devices 21 within the pressure vessel 5 in order to insert them into the core 3 or remove them from the core.

[0085] The nuclear reactor 1 also has a control system 23 .

[0086] The control system 23 is typically a protection system configured to cause an emergency stop of the nuclear reactor in unexpected circumstances.

[0087] To do this, the control system 23 is configured to send an automatic reactor stop command to the drive mechanism of the control device 21. Such a command enables the control device to be quickly inserted into the core 3 of the nuclear reactor.

[0088] Alternatively, the control system 23 is a monitoring system configured to generate an alarm when the nuclear reactor approaches its operating limits.

[0089] According to another variant, the control system 23 acts both as a monitoring system and as a protection system.

[0090] The nuclear reactor 1 is also equipped with a plurality of sensors configured to measure a plurality of quantities characterizing the operation of the nuclear reactor 1 .

[0091] The nuclear reactor 1 thus has temperature sensors 23 , 25 which measure the temperature of the primary fluid in the hot branch 9 and in the cold branch 11 , respectively.

[0092] The nuclear reactor 1 comprises a sensor 27 which measures the pressure in the pressurizer 13 .

[0093] The nuclear reactor also has a sensor 29 which measures the rotation speed of the main pump 15 .

[0094] The nuclear reactor 1 also has a neutron flux detector 31 .

[0095] According to a first embodiment of the invention, these detectors are ex-core detectors arranged outside the core 3 .

[0096] The detectors 31 are distributed around the pressure vessel 5 .

[0097] Typically, they are multi-excore chambers, for example having 6 sections arranged one above the other along the central axis of the vessel.

[0098] The detector 31 is used to measure the neutron flux emitted by the core 3 at different axial levels.

[0099] The nuclear reactor 1 comprises a sensor 33 for the position of the core reactivity control device 21 .

[0100] The nuclear reactor further comprises an information processing unit 35 configured to implement the operating method to be described below.

[0101] The information processing unit is, for example, an element of the control system 23. It is usually composed of a processor and a memory associated with the processor. Alternatively, the data processing unit is implemented in the form of a programmable logic component such as an FPGA (field programmable gate array) or in the form of a dedicated integrated circuit such as an ASIC (application-specific integrated circuit).

[0102] The various sensors 23 , 25 , 27 , 29 , 31 , 33 provide information to a data processing unit 35 .

[0103] Now refer to Figure 1 , Figure 3 and Figure 4 A method for operating a nuclear reactor 1 according to a first embodiment is described.

[0104] The method comprises the following steps:

[0105] - Step S10: Acquiring a plurality of quantities characterizing the operation of the nuclear reactor 1;

[0106] -Step S12: Use Figure 4 The deep neural network 37 shown calculates at least one deviation from nucleate boiling ratio (DNBR), the entries of the deep neural network 37 being determined using the acquired quantities, the deep neural network 37 comprising at least two hidden layers, each hidden layer having at least five neurons;

[0107] - Step S14: calculating a deviation between at least one calculated deviation from nucleate boiling ratio (DNBR) and a plurality of predetermined reference values;

[0108] - Step S16: Using the calculated deviation, formulate a control signal for the control system of the reactor 23, the control signal being selected from a list comprising at least the following values:

[0109] * Reactor automatic shutdown or alarm;

[0110] * Do nothing;

[0111] - Step S18: When the elaborated control signal has the value "reactor automatic stop", the control system 23 emergency stops the reactor, or when the elaborated control signal has the value "alarm", the control system 23 issues an alarm signal.

[0112] The acquired quantities include at least neutron flux measurements made by the neutron flux detector 31, which allow the axial power distribution P(z) in the core 3 to be reconstructed, as described below.

[0113] For a single or each primary circuit 7, the acquired quantities also include one or more of the following quantities: pressure p of the pressurizer, primary fluid flow Q in the primary circuit, primary fluid temperature T in the hot branch and the cold branchbc and T bf .

[0114] Typically, all of the above quantities are obtained.

[0115] The pressure of the pressurizer p and the primary fluid temperature T in the hot branch and the cold branch bc and T bf Provided directly from sensors 27, 23, 25.

[0116] The primary fluid flow rate Q is reconstructed using the rotational speed of the main pump measured by the sensor 29. For example, it is determined by calculation using a relationship that indicates the primary fluid flow rate directly from the rotational speed.

[0117] The acquired quantities also include the current insertion position of the core reactivity control device 21 provided by the sensor 33 .

[0118] Therefore, if Figure 1 and Figure 3 By way of example, the acquisition step S10 generally comprises a sub-step S20 of reading the values ​​provided by sensors 23, 25, 27, 29, 31, 33, wherein certain quantities characterizing the operation of the reactor are read directly from these sensors. These quantities are in particular the pressure p of the pressurizer and the temperature T of the hot branch. bc and cold branch temperature T bf .

[0119] The acquisition step S10 also comprises a sub-step S22 during which at least one other quantity characterizing the operation of the nuclear reactor is determined by calculation from the measured values. This generally involves reconstructing the primary fluid flow Q, or applying a filter to the quantity acquired in step S20.

[0120] The step S12 of calculating at least one DNBR includes a sub-step S24 of determining entries of the deep neural network 37 using the acquired quantities, and a sub-step S25 of calculating at least one DNBR using the deep neural network 37 by using the entries determined in the sub-step S24.

[0121] The entries of the neural network 37 include at least the axial power distribution P(z) in the nuclear reactor core 3 reconstructed from the acquired neutron flux measurements.

[0122] The axial power distribution P(z) is a vector with multiple components, each component corresponding to the average power level of the nuclear reactor at a given axial position. Each component defines an entry of the deep neural network.

[0123] The number of components depends on the accuracy required to reconstruct the power distribution and thus on the desired accuracy of the DNBR calculation. For example, the axial power distribution has as many components as there are parts in the various detectors 31: the signals from the different detector parts define the deep neural network entries.

[0124] Alternatively, the axial power distribution has more components than the number of sections in the individual detectors 31. The module for reconstructing the axial power distribution from the values ​​from the different sections of the detector provides a finer gridded axial power distribution according to a dedicated algorithmic processing (such as polynomial interpolation), the individual components of which define a deep neural network entry.

[0125] Thus, in sub-step S24, the power of the nuclear reactor 1 at a given axial level is calculated from the measured neutron flux, using, for example, an analytical relationship that directly gives the power level as a function of the reconstructed neutron flux and the thermal power Pth released by the core as described below.

[0126] The entries of the neural network 37 also include the thermal power Pth released by the core.

[0127] In sub-step S24, by using at least the pressurizer pressure p, the primary fluid flow Q, and the primary fluid temperature T in the hot branch and the cold branch bc and T bf A calculation is performed to determine the thermal power output Pth.

[0128] The calculation method is known and will not be described in detail here.

[0129] Typically, the calculation of the thermal power Pth released for each primary circuit 7 is performed using the quantities obtained for the primary circuit 7 .

[0130] The entries of the neural network 37 also include the enthalpy rise factor FΔH.

[0131] In sub-step S24, the enthalpy rise factor FΔH is determined by performing the following operations:

[0132] - using the acquired current insertion position of the reactivity control device of the core 21, calculate the radial power peak factor Fxy(z);

[0133] - Calculation of the enthalpy rise factor FΔH using the calculated radial power crest factor Fxy(z), the reconstructed axial power distribution P(z) and the determined thermal power release Pth.

[0134] The radial power crest factor Fxy(z) and the enthalpy rise factor FΔH are calculated using known formulas and will not be described in detail here.

[0135] Preferably, using the released heat power Pth calculated for each primary circuit 7 , enthalpy rise factor calculation FΔH is performed for the primary circuit 7 .

[0136] When the nuclear reactor 1 includes a plurality of primary circuits 7, preferably, in sub-step S25, the deviation from nucleate boiling ratio DNBR is calculated for each primary circuit 7 using the deep neural network 37. To this end, the deep neural network 37 uses at least an entry determined using the acquired amount of the corresponding primary circuit 7.

[0137] Alternatively, a single deviation from nucleate boiling ratio DNBR is calculated using the deep neural network 37. For example, it is calculated using entries determined by any processing of the acquired quantities for the primary circuit suitable for DNBR evaluation.

[0138] like Figure 4 As can be seen in the figure, the deep neural network 37 includes Figure 4 The input layer 39 on the left is located Figure 4 The output layer 41 on the right, and a plurality of intermediate layers 43, also referred to as hidden layers. The number nH of the intermediate layers 43 is greater than or equal to two, preferably greater than or equal to five, and preferably greater than or equal to ten.

[0139] This allows a large number of different situations in nuclear reactors to be considered.

[0140] Each hidden layer 43 includes at least five neurons 45 , preferably at least seven neurons 45 , and preferably at least ten neurons 45 .

[0141] The input layer 39 has as many input neurons as entries.

[0142] For example, the input layer 39 has a bf , Pth, FΔH. It also has an input neuron for each component of the axial power distribution P(z).

[0143] For example, if the axial power distribution P(z) is a six-component vector, the input layer 39 has eleven input neurons.

[0144] The output layer 41 has a single output neuron corresponding to the minimum value of the PCR calculated for the corresponding primary loop.

[0145] Advantageously, in Figure 1 In step S22, the entry is subjected to a first post-processing before being introduced into the neural network 37.

[0146] The first post-processing consists, for example, in applying a dynamic compensation term, typically an advance-delay filter, which allows each acquired quantity to be an image of the corresponding physical parameter. This post-processing compensates for instrument response times (cold and hot branch temperature measurements of the vessel), fall times of the core reactivity control devices, and response times of the process and treatment algorithms. The functional characteristics of the information processing performed by the neural network 37 allow a significant reduction in the need to use filters compared to conventional algorithms.

[0147] These entries are preferably the subject of a second post-processing by normalizing the individual quantities or components before being introduced into the neural network.

[0148] The entries form a vector, called the input vector and annotated as X0. Each hidden layer i of the neural network processes the vector value X from the previous layer i-1. i-1 To generate the vector value X i .

[0149] Thus, for each hidden layer i, i is between 1 and nH,

[0150] X i =σ i .A i .X i-1

[0151] Among them, A i It is scheduled N i *N i-1 N-dimensional matrix i , where N i is the number of neurons in layer i, and N i-1 is the number of neurons in layer i-1;

[0152] σ i is a predetermined nonlinear operator.

[0153] In other words, the deviation from nucleate boiling ratio is calculated by the deep neural network 37 using the following formula:

[0154]

[0155] where Y is the calculated deviation from nucleate boiling ratio.

[0156] Each matrix A i Use as vector X i-1 In other words,

[0157] A i X i-1 =W i X i-1 +b i

[0158] Among them, W i is the N of layer i i The neurons are connected to the N neurons in layer i-1 i-1 The synaptic weight matrix of neurons, and b i is the N of layer i i dimensional deviation vector.

[0159] For the last layer, we note that the matrix AnH+1 has dimension 1xNnH and σnH+1=1: no nonlinear transfer function is applied to build the Y signal of the last layer.

[0160] Nonlinear operator σ i Each is used as a nonlinear function, such as the vector A i X i-1 The sigmoid function is a function expressed as follows:

[0161]

[0162] Here, l is a specially parameterized constant.

[0163] In step S14, for each primary circuit, a set of deviations between the deviation from nucleate boiling ratio calculated for the primary circuit and a plurality of predetermined reference values ​​is calculated. Thus, as many deviation calculations as primary circuits are performed.

[0164] These values ​​are operational reference thresholds, called protection (or monitoring) thresholds, which incorporate methodological uncertainties, namely measurement uncertainty and calculation uncertainty.

[0165] Typically, the calculated deviations are displayed in alphanumeric or graphical form on at least one screen for real-time monitoring of the protection and operating margins of the units of the respective primary circuit.

[0166] When calculating a single DNBR, a single group deviation is calculated in step S14.

[0167] Step S16 is for interpreting the deviation group found in step S14.

[0168] To do this, a temporary control signal is generated for each primary loop using the set of deviations calculated for that loop.

[0169] When the control system is a protection system, the temporary control signal is selected from a list containing at least the following values: reactor automatically stops, do nothing.

[0170] When the control system is a monitoring system, the temporary control signal is selected from a list containing at least the following values: alarm, do nothing.

[0171] When the control system is a protection and monitoring system, the temporary control signal is selected from a list containing at least the following values: reactor automatic stop, alarm, do nothing.

[0172] The temporary control signals of all primary circuits are used to generate control signals sent to the control system of the reactor 23 .

[0173] After accounting for the deviations noted in step S14, step S16 comprises applying voting logic to the signals generated for the various loops of the reactor. For example, if the majority of temporary control signals have the value "reactor automatic stop", then the control signal sent to the control system 23 is "reactor automatic stop".

[0174] In step S18, the control system 23 receives the control signal formed in step 16 and acts accordingly. For example, if the control signal is "reactor automatic stop", the control system triggers the automatic stop of the reactor 1 by inserting the reactivity control device of the core 21, or if the generated control signal has the value "alarm", the control system triggers the emission of an alarm signal.

[0175] Preferably, the method further comprises a step S26 of designing a deep neural network 37. The design step S26 comprises the following sub-steps:

[0176] -S28: Create a database containing at least 100,000 reactor states;

[0177] - S30: Use at least a portion of the database to learn a deep neural network.

[0178] Advantageously, the database comprises at least 500,000 reactor states, preferably at least 1,000,000 reactor states.

[0179] Each state is defined by a set of values ​​of quantities characterizing nuclear reactor operation and a deviation from nucleate boiling ratio value. The deviation from nucleate boiling ratio value is calculated by a 3D thermal hydraulic reference code using a corresponding set of values ​​of quantities characterizing nuclear reactor operation.

[0180] The 3D core thermal hydraulic calculation code is a high-precision code, such as the FLICA code. Typically, the Monte Carlo sampling technique is used to generate the database.

[0181] The states are selected to cover a number of predetermined contingency situations leading to triggering of intervention of the generic low-level DNBR protection chain of the reactor.

[0182] Predetermined unexpected situations are for example those referred to as Category 2 in the Nuclear Reactor Safety Reports, which require intervention of the generic low-level DNBR protection chain.

[0183] The ranges of variation of the quantities characterizing the operation of the nuclear reactor are also selected in such a way that they cover the effective range of deviations from the nucleate boiling ratio.

[0184] The data are generated in such a way that they optimize the detection of physical symmetries by the neural network during a learning phase.

[0185] Preferably, the input variables are normalized.

[0186] The learning sub-step S30 is performed on only a portion of the database created in sub-step S28. For example, this sub-step uses 50% to 90% (typically 80%) of the database.

[0187] The learning phase includes the following operations:

[0188] i) determine the best neuronal structure by performing sensitivity studies dedicated to the choice of model hyperparameters, a step that may require the use of an optimization algorithm (of the genetic algorithm type): the number of hidden layers, the number of neurons in each layer, the choice of the transfer function imposed by the phenomenology of the boiling crisis phenomenon (derivable, for example, due to the analytical expression of the critical flux correlation and which must exhibit a strong nonlinear nature);

[0189] ii) Develop a learning algorithm specifically optimized for the generated database, which allows to accelerate and stabilize the convergence process and improve the convergence towards a more robust solution by avoiding local minima. This improves the generalization ability of the neural network, thereby improving its ability to respond positively to the validation and testing phases described below.

[0190] This is achieved, for example, by developing an adaptive batch stochastic gradient descent algorithm. The updates of the synaptic weight and bias tensor learning are performed for each cycle on data batches of evolving size by scanning the entire learning library parameterized by the number of cycles.

[0191] Furthermore, an inertial method with regard to the adaptation step of the learning algorithm is advantageously used in order to increase the efficiency and robustness of the learning phase.

[0192] Preferably, step S26 also comprises a sub-step S32 of validation of the deep neural network determined in sub-step S30. This validation phase is performed on a second part of the database not used in sub-step S30. Typically, the entire part of the database not used in step S30 is used in step S32. This makes it possible to validate the deep neural network predictions developed at the end of step S30 by comparison with predictions generated using a reference code, and to characterize the generality of the predictive model thus developed. Sensitivity studies are preferably carried out during sub-step S32. If the validation phase is not satisfactory, return to sub-step S30 and then modify the neuronal structure (i.e. the number of hidden layers, the number of neurons per layer) or the hyperparameters of the learning algorithm in order to restart the learning step aimed at determining the matrix of synaptic weights and bias vectors of the neural network, thereby defining the previously defined matrix A i .

[0193] Preferably, step S26 further includes sub-step S34, which performs phenomenon verification (testing) on ​​the deep neural network on the database-independent data determined in sub-step S28.

[0194] During sub-step S34, a deep neural network is implemented in conjunction with a computational code that simulates unexpected transients. These unexpected transients are typically those defined in nuclear reactor safety reports.

[0195] If this final testing phase is not satisfactory, sub-step S30 is repeated.

[0196] The method described above is therefore particularly suitable for being implemented with a reactor of the type described above.

[0197] In contrast, the nuclear reactor 1 described above is particularly suitable for carrying out the method just described.

[0198] Specifically, Figure 2 As shown, the information processing unit 35 includes:

[0199] - a module 47 for acquiring a plurality of quantities characteristic of the operation of the nuclear reactor;

[0200] - a module 49 for calculating at least one deviation from nucleate boiling ratio using a deep neural network 37, configured to determine entries of the deep neural network 37 using the acquired and determined quantities, the deep neural network 37 comprising at least two hidden layers, each having at least five neurons;

[0201] - a module 51 for calculating the deviation between at least one calculated deviation from nucleate boiling ratio and a plurality of predetermined reference values;

[0202] - A module 53 for using the calculated deviation to generate a control signal for a reactor control system, the control signal being selected from a list comprising at least the following values:

[0203] * Reactor automatic shutdown or alarm;

[0204] *Do nothing.

[0205] The reactor control system 23 is configured to automatically stop the reactor when the generated control signal has a value of "reactor automatic stop". The reactor control system 23 is configured to generate an alarm when the processed control signal has a value of "alarm".

[0206] In other words, module 47 is configured to implement step S10 of the above method. Module 49 is configured to perform step S12 of the above method. Module 51 is configured to perform step S14 of the above method, and module 53 is configured to perform step S16 of the above method.

[0207] Now refer to Figure 5 A second embodiment of the present invention is described. Only the differences between the second embodiment and the first embodiment will be described in detail below. The same elements or elements performing the same function will be indicated by the same reference numerals.

[0208] In the second embodiment, the neutron flux detector 31 is not an ex-core detector disposed outside the core 3 , but an in-core detector permanently inserted into the core 3 .

[0209] They are arranged in groups, with detectors of the same group staggered vertically above each other.

[0210] The detectors are, for example, neutron detectors on rods arranged vertically in the core (colons). Each electron rod allows the neutron flux to be measured at several axial heights in the core, such as six axial heights for a vertical arrangement of 6 electrons. A nuclear reactor is usually equipped with 12 rods, each with 6 electrons, which are distributed in the core.

[0211] The released thermal power Pth is determined by calculation using the acquired neutron flux measurements. The calculation no longer uses the pressurizer pressure p, the primary fluid flow rate Q, the temperature T of the primary fluid in the hot and cold branches bc , T bf The calculation formula used is known and will not be described in detail here.

[0212] The enthalpy rise factor FΔH is reconstructed using the axial power distribution P(z). It is not necessary to reconstruct the factor Fxy(z) from the measurements of the positions of the reactivity control devices of the core 21 provided by the sensors 33 .

[0213] In step S12 , a deviation from nucleate boiling ratio DNBR is calculated for each set of in-core neutron flux detectors 31 using a deep neural network 37 , wherein entries are determined using neutron flux measurements made by the in-core neutron flux detectors 31 .

[0214] From the neutron flux measurements provided by the detector group at several axial heights, a set of parameters P(z), Pth and FΔH specific to the detector group is determined. These parameters are used as entries of the neural network 37.

[0215] Other terms are also used, such as the pressurizer pressure p, the primary fluid flow Q, the primary fluid temperature T in the cold leg bc 、T bf These input quantities are the same for all sensor groups.

[0216] In the deviation calculation step S14, a set of deviations between each previously determined deviation from nucleate boiling ratio and a plurality of predetermined reference values ​​are calculated.

[0217] In the control signal generating step S16, a temporary control signal is generated for each calculated deviation from nucleate boiling ratio using a corresponding set of deviations.

[0218] According to the above-mentioned voting logic, at least several temporary control signals are used to generate the control signal.

[0219] A third embodiment of the present invention will now be briefly described. Only the differences between the third embodiment and the second embodiment will be described in detail below. The same elements or elements performing the same function will be indicated by the same reference numerals.

[0220] In a third embodiment, the data processing unit 35 has a continuous 3-dimensional map of the power in the reactor core.

[0221] This mapping is a quantity that characterizes the operation of the reactor. As in the first and second modes, other quantities such as T bf , p and Q.

[0222] Some entries of the deep neural network 37 are determined computationally using a 3D map of the power distribution in the core, thereby replacing the quantities P(z), pth and FΔH described previously for the first embodiment of the invention.

[0223] Advantageously, the components of the 3D power distribution constitute entries of a dedicated neural structure called a convolutional layer, which enables the extraction of the characteristics of the power distribution to be optimized in order to generate a deep neural network, and in particular optimize its structure for hardware implementation. The output of the convolutional layer defines a partial entry of the deep neural network 37. In particular, the output replaces the vector of the components of the axial power distribution at the input of the deep neural network 37, and the size FΔH. This neural structure suitable for processing spatially structured data replaces matrix multiplication by a mathematical convolution operation between the input power distribution of the structure and a series of kernels (or filters) adjusted during the training of step S30 with a series of degrees of freedom. The convolutional layer also includes a nonlinear processing of the convolution output by a nonlinear transfer function, and a third step that allows the introduction of invariants under local geometric transformations (such as translations). These steps can be repeated in the generation of the optimal neuronal structure.

[0224] The method for utilizing the invention may present several variations.

[0225] The quantities characterizing the operation of the nuclear reactor and the entries of the deep neural network obtained in step S10 may not exactly correspond to the above list. Other quantities may be obtained. Certain quantities may not be obtained. Certain entries may not be used. Other entries may be added.

[0226] In any case, the power distribution has an important influence on the location in the core and the value of the minimum margin of the DNBR.

[0227] Taking this into account improves the accuracy of the DNBR reconstruction. Figure 2 , Figure 3 and Figure 5 Some quantities may not be directly read by sensors equipped with the nuclear reactor. They can be derived from other measured quantities or values ​​from the operating system of the nuclear reactor.

[0228] The convolutional layer described in the third embodiment can also be implemented in the first two implementations to process the axial power distribution P(z) in the neuron structure. The other items of the neural network remain unchanged.

[0229] The above-described method has several advantages.

[0230] The response time is very short, for example in the millisecond range. This is obtained in particular because the method does not require the implementation of a convergence cycle, only parameters specific to deep neural networks are necessary, and the unit calculation operations are simple and easily parallelizable if the hardware solution allows.

[0231] Therefore, this approach frees up computing power for other parts of the nuclear reactor control and instrumentation system.

[0232] This performance is obtained by the fact that the complexity of the underlying physics is encoded through the parameterization of the deep neural network from the database. This allows the deep neural network to reconstruct the DNBR values ​​of the thermal-hydraulic reference code by simply using the synaptic weights and transfer functions of the neural structure.

[0233] The DNBR calculations provided by the deep neural network are very reliable and the responses are close to those provided by a 3D thermal-hydraulic reference code (here FLICA). The conservatism of the response is adjustable.

[0234] The quality of the response provided by the deep neural network is significantly attributed to the fact that no reconstruction of intermediate local thermal-hydraulic variables is performed without fine physical modeling. The reliability of the response provided is a function of the fineness of the grid of the database used to learn the neural network. This only takes the computational time before commissioning in a nuclear reactor. The method is also more robust to uncertainty propagation on the input data (initiated by random fluctuations in measurements from core instruments), which improves the verification and hardware qualification phases.

Claims

1. A method for operating a nuclear reactor (1) comprising a core (3), the method comprising the following steps: - acquiring a plurality of quantities characterizing said operation of said nuclear reactor (1); - calculating at least one deviation from nucleate boiling ratio using a deep neural network (37), the entries of the deep neural network (37) being determined using the plurality of quantities obtained, the deep neural network comprising at least two hidden layers, each hidden layer having at least five neurons; - calculating a deviation between said at least one calculated deviation from nucleate boiling ratio and a plurality of predetermined reference threshold values; - using the calculated deviation to generate a control signal for a control system (23) of the reactor, the control signal being selected from a list comprising at least the following values: * Reactor automatic shutdown or alarm; * Do nothing; -When the generated control signal has the value "reactor automatic stop", the control system (23) urgently stops the nuclear reactor (1), or when the generated control signal has the value "alarm", the control system (23) sends an alarm signal.

2. The method according to claim 1, wherein: The plurality of quantities acquired include neutron flux measurements made by a neutron flux detector (31), and the entries of the deep neural network (37) include an axial power distribution (P(z)) in the core of the nuclear reactor reconstructed from the neutron flux measurements.

3. The method according to claim 2, wherein: The axial power distribution (P(z)) is a vector having a plurality of components, each component corresponding to the average power of the nuclear reactor for each axial slice at a given axial position, each component defining one of the entries of the deep neural network (37).

4. The method according to claim 2 or 3, wherein: The neutron flux detector (31) is an ex-core detector arranged outside the core (3).

5. The method according to claim 2 or 3, wherein: The neutron flux detector (31) is an in-core detector permanently inserted into the core (3).

6. The method according to any one of claims 1 to 3, wherein: The nuclear reactor (1) comprises: a pressure vessel (5) containing the core (3) and at least one primary circuit (7), the primary circuit (7) comprising a hot branch and a cold branch (9, 11) directly connected to the pressure vessel (5); a pressurizer (13) controlling the pressure in the primary circuit (7); and a main pump (15) ensuring the circulation of the primary fluid in the primary circuit (7), wherein for the primary circuit (7) or each primary circuit (7), the plurality of quantities obtained comprise one or more of the following quantities: the pressure (p) of the pressurizer, the flow rate (Q) of the primary fluid in the primary circuit (7), the temperature (T) of the primary fluid in the hot branch, bc ) and the temperature of the primary fluid in the cold branch (T bf ).

7. The method according to claim 6, wherein: The neural network entry has one or more of the following variables: pressurizer pressure (p), flow rate (Q) of the primary fluid in the primary circuit (7), temperature (T bf ).

8. The method according to any one of claims 1 to 3, wherein: The entries of the deep neural network (37) include the thermal power (Pth) released by the core.

9. The method according to claim 6, wherein: The entries of the deep neural network (37) include the thermal power (Pth) released by the core, and are obtained by using at least the pressurizer pressure (p), the primary fluid flow (Q), the temperature (T bc ) and the temperature of the primary fluid in the cold branch (T bf ) is calculated to determine the thermal power (Pth) released.

10. The method according to claim 5, wherein: The entries of the deep neural network (37) include thermal power (Pth) released by the core, and the thermal power (Pth) released is determined by calculation using the neutron flux measurement.

11. The method according to any one of claims 1 to 3, wherein: The entries of the deep neural network (37) include enthalpy rise factors (FΔH).

12. The method according to claim 2 or 3, in, The nuclear reactor (1) comprises: a pressure vessel (5) containing the core (3) and at least one primary circuit (7), the primary circuit (7) comprising a hot branch and a cold branch (9, 11) directly connected to the pressure vessel (5); a pressurizer (13) controlling the pressure in the primary circuit (7); and a main pump (15) ensuring the circulation of the primary fluid in the primary circuit (7), wherein for the primary circuit (7) or each primary circuit (7), the plurality of quantities obtained comprise one or more of the following quantities: the pressure (p) of the pressurizer, the flow rate (Q) of the primary fluid in the primary circuit (7), the temperature (T) of the primary fluid in the hot branch, bc ) and the temperature of the primary fluid in the cold branch (T bf ), wherein the entries of the deep neural network (37) include enthalpy rise factors (FΔH), wherein the neutron flux detector (31) is an off-core detector arranged outside the core (3), the entries of the deep neural network (37) include the thermal power (Pth) released by the core, the pressurizer pressure (p), the primary fluid flow (Q), the temperature (T bc ) and the temperature of the primary fluid in the cold branch (T bf ) is calculated to determine the released thermal power (Pth), and the plurality of quantities obtained include a current insertion position of a core reactivity control device (21), and the enthalpy rise factor (FΔH) is determined by performing the following operations: - calculating a radial power peak factor (Fxy(z)) using the acquired current insertion position; and - calculating the enthalpy rise factor (FΔH) using the calculated radial power crest factor (Fxy(z)), the reconstructed axial power distribution (P(z)) and the determined output of the thermal power (Pth).

13. The method according to claim 2 or 3, wherein: The entries of the deep neural network (37) include an enthalpy rise factor (FΔH), wherein the neutron flux detector (31) is an in-core detector permanently inserted into the core (3), and the axial power distribution (P(z)) is used to reconstruct the enthalpy rise factor (FΔH).

14. The method according to claim 6, wherein: The nuclear reactor (1) comprises several primary circuits (7), and the deviation from nucleate boiling ratio is calculated for each primary circuit (7) by means of the deep neural network (37) using at least the acquired quantity for each primary circuit (7).

15. The method according to claim 5, wherein: The neutron flux detectors (31) are distributed in a plurality of groups, wherein the neutron flux detectors (31) of the same group are vertically staggered above and below each other, and the minimum deviation from nucleate boiling ratio is calculated for each group of in-core neutron flux detectors (31) by means of the deep neural network (37), wherein the entries of the deep neural network are determined using at least the neutron flux measurements performed by the in-core neutron flux detectors (31) of the group.

16. The method according to claim 14, wherein: - in the deviation calculation step, calculating a set of deviations between each calculated deviation from nucleate boiling ratio and the plurality of predetermined reference thresholds; - in said step of generating a control signal, generating for each calculated deviation from nucleate boiling ratio a temporary control signal using a corresponding set of deviations, said temporary control signal being selected from a list comprising at least the following values: * Reactor automatic shutdown or alarm; * Do nothing; At least several of the temporary control signals are used to generate the control signal.

17. The method according to claim 15, wherein: - in the deviation calculation step, calculating a set of deviations between each calculated deviation from nucleate boiling ratio and the plurality of predetermined reference thresholds; - in said step of generating a control signal, generating for each calculated deviation from nucleate boiling ratio a temporary control signal using a corresponding set of deviations, said temporary control signal being selected from a list comprising at least the following values: * Reactor automatic shutdown or alarm; * Do nothing; At least several of the temporary control signals are used to generate the control signal.

18. The method according to any one of claims 1 to 3, wherein: The deviation from nucleate boiling ratio is calculated by the deep neural network (37) using the following formula: Wherein, Y is the calculated deviation from nucleate boiling ratio; X0 is the vector of entries; nH is the number of hidden layers of the deep neural network; A i It is scheduled N i *N i-1 N-dimensional matrix i , where N i is the number of neurons in layer i, and N i-1 is the number of neurons in layer i-1; σ i is a predetermined nonlinear operator.

19. The method according to claim 18, wherein: A i is the vector X i-1 The components of the matrix act as an affine transformation, the vector X i-1 is a vector whose components are determined by the neuron layer i-1.

20. The method according to claim 18, wherein: σ i Vector A i .X i-1 All components of the vector X act as sigmoid functions. i-1 is a vector whose components are determined by the neuron layer i-1.

21. The method according to any one of claims 1 to 3, wherein: The method comprises a deep neural network (37) design step, wherein the design step comprises the following sub-steps: - constructing a database containing at least 100,000 reactor core states, each state being defined by a set of values ​​of said quantities characterizing said operation of said nuclear reactor (1) and by a minimum deviation from nucleate boiling ratio value calculated by a reference 3D thermal-hydraulic calculation code using said set of values, said states being selected to cover a plurality of contingencies of a predetermined category 2, characterized by the insertion of a reactivity with moderate kinetics, requiring the intervention of a protection chain with a low deviation from nucleate boiling ratio to ensure freedom from boiling crisis phenomena, said control system (23) of said reactor being provided to handle said contingencies; - using at least a portion of the database to learn the deep neural network (37); as well as - Validating the deep neural network using a portion of the database independent of that used for the learning phase and testing the neural network obtained by coupling with an unexpected neutron and thermal hydraulic code, which makes it possible to simulate unexpected transients requiring intervention of the control system (23).

22. A nuclear reactor (1), comprising: - a core (3); - control system (23); - an information processing unit (35) configured to implement a method for operating a nuclear reactor (1) comprising a core (3) according to any one of claims 1 to 3.

Images