Operating process for a nuclear reactor with on-line RFTC calculation, corresponding nuclear reactor

The present invention generally relates to the protection and / or monitoring of the core of a pressurized water nuclear reactor against boiling crisis.

Maintaining the integrity of the first safety barrier (fuel cladding) is a major objective in the design of nuclear reactors, and is the subject of a safety demonstration carried out, among other things, by simulating using of numerical computer codes the behavior of the reactor during families of accidental transients.

All of this analysis is recorded in the final safety report for the installation, required by the nuclear safety authority to grant the reactor operating license.

The integrity of the fuel cladding is guaranteed by the absence of melting in the center of the fuel pellets, and by the absence of a boiling crisis (Departure from Nucleate Boiling, DNB in ​​English) at the surface of the cladding. These conditions must be satisfied in all points of the heart.

The boiling crisis is characterized by a degradation of the heat exchange between the coolant of the primary circuit and the surface of the fuel cladding, which can lead to a phenomenon of heating and a possible loss of the integrity of the first barrier.

Each nuclear reactor is equipped with a protection system, the role of which is to guarantee the safety of the reactor for a certain number of accident situations. The protection system typically comprises a so-called low RFTC (Critical Heat Flux Ratio) protection chain. In reactors with four primary loops such as the 1300MWe, N4 and EPR, the low RFTC protection chain ensures the protection of the nuclear reactor against the phenomenon of boiling crisis during accidental category 2 transients, characterized by an insertion of reactivity with moderate dynamics (uncontrolled withdrawal of regulation units, untimely dilution of the primary fluid leading to a decrease in the boron concentration, untimely cooling of the primary circuit, etc.).

To this end, the protection system comprises an information processing unit (called the Digital Integrated Protection System or SPIN on the 1300 MWe and N4 reactors), which integrates a functional unit designated by the name “low RFTC algorithm”. This algorithm calculates the safety margins against the boiling crisis online. To do this, it calculates the RFTC, and compares the calculated value with a limit value taking into account the design uncertainties of the reactor and the

uncertainties in reconstructing the value of the RFTC. The low RFTC protection chain causes an automatic shutdown of the reactor if the reconstructed value is below the limit.

The critical heat flow ratio or RFTC is defined as the ratio between the critical heat flow and the local heat flow. The local heat flow corresponds to the heat flow emitted by the fuel rods through the cladding. The critical heat flux is the value of the heat flux leading to damage to the fuel rod cladding, which depends on the local thermohydraulic conditions of the fluid.

The low RFTC algorithm uses a simplified thermo-hydraulic modeling of the core, adjusted during the design by corrective functions (bias curves) acting on the local thermo-hydraulic quantities, in order to ensure a conservative response with respect to the core. RFTC calculated by the 3D thermo-hydraulic reference code.

This simplified thermohydraulic modeling makes it possible to calculate the local thermal flux from the measured quantities characterizing the power distribution of the reactor and the local thermohydraulic conditions making it possible to evaluate the critical thermal flux.

On the other hand, the low RFTC algorithm uses a predetermined critical flux correlation to calculate the critical heat flux. This correlation is determined empirically, from experimental data. The design of this correlation represents a complex and costly development. Its multidimensional polynomial type analytical formula is postulated a priori and manually adjusted by successive stages during the development process in order to meet certain statistical criteria.

The low RFTC calculation algorithm is imprecise, so that it is necessary to provide high margins to ensure safety during nuclear reactor operations. This is very penalizing for the control of the reactor. In addition, although simplified, thermo-hydraulic modeling requires sophisticated digital processing using microprocessor systems making it possible to make the algorithm converge on times compatible with the protection of the nuclear reactor.

In this context, the invention aims to provide a method of operating a nuclear reactor that does not have the above defects.

To this end, the invention relates to a method of operating a nuclear reactor comprising a core, the method comprising the following steps:

- acquisition of a plurality of quantities characterizing the operation of the nuclear reactor;

- calculation of at least one critical heat flow ratio using a deep neural network, the inputs of the deep neural network being determined using the acquired quantities, the deep neural network comprising at least two hidden layers at least five neurons each;

- Calculation of the differences between the at least one calculated critical heat flow ratio and a plurality of predetermined reference threshold values;

- development of a control signal for a reactor control system, using the calculated deviations, the control signal being chosen from the list comprising at least the following values:

* automatic reactor shutdown or alarm;

* do nothing;

- emergency shutdown of the nuclear reactor by the control system when the produced control signal has the value "automatic reactor shutdown", or emission of an alarm signal by the control system when the produced control signal presents the "alarm" value.

The low RFTC algorithm has the major flaw that the local thermohydraulic quantities are reconstituted with a simplified physical modeling: single-channel, single-phase model, without thermo-hydraulic correlation other than the critical flow correlation (pressure drops, void rate, ...).

The use of a deep neural network makes it possible to determine the critical heat flow ratio as a function of the current values ​​of the limit quantities characterizing the operation of the nuclear reactor, in a much more precise and much faster way than with the low algorithm. RFTC implemented today.

The values ​​predicted by the deep neural network are much closer to the values ​​calculated by the reference thermo-hydraulic code. The RFTC reconstruction error is greatly reduced. As a result, the operating margins of the nuclear reactor are significantly better.

The operating process may also have one or more of the characteristics below, considered individually or in any technically possible combination:

said acquired quantities comprise neutron flux measurements carried out by neutron flux detectors, the inputs of the neural network comprising the axial power distribution in the core of the nuclear reactor, reconstructed from the acquired neutron flux measurements;

the axial power distribution is a vector having a plurality of components, each component corresponding to an average power of the nuclear reactor per axial section at a given axial dimension, each component defining one of the inputs of the deep neural network;

- neutron flux detectors are ex-core detectors placed outside the core;

- neutron flux detectors are in-core detectors permanently inserted in the core;

- the nuclear reactor comprises a pressure vessel containing the core and at least one primary loop, the primary loop comprising hot and cold branches fluidly connected directly to the vessel, a pressurizer controlling a pressure in the primary loop, and a primary pump ensuring the circulation of primary fluid in the primary loop, the quantities acquired comprising, for the or each primary loop, one or more of the following quantities: pressure of the pressurizer, flow of primary fluid in the primary loop, temperature of the primary fluid in the hot branches and cold;

the inputs of the neural network include one or more of the following quantities: pressure of the pressurizer, flow rate of primary fluid in the primary loop, temperature of the primary fluid in the cold branch;

- the inputs of the neural network include the thermal power released by the heart;

the thermal power released is determined by calculation, using at least the pressure of the pressurizer, the flow rate of the primary fluid, the temperatures of the primary fluid in the hot and cold branches, acquired for at least one primary loop;

- the thermal power released being determined by calculation, using the acquired neutron flux measurements;

- the inputs of the neural network include the enthalpy elevation factor;

- the quantities acquired include a current insertion position of the organs for controlling the reactivity of the heart, the enthalpy increase factor being determined by carrying out the following operations:

- calculation of a radial peak power factor using the acquired current insertion position;

- calculation of the enthalpy rise factor using the calculated radial peak power factor, the reconstructed axial power distribution and the determined thermal power output;

- the enthalpy rise factor is reconstructed using the axial power distribution;

the nuclear reactor comprises several primary loops, a critical heat flow ratio being calculated for each primary loop using the deep neural network using at least said quantities acquired for said primary loop;

- the detectors are distributed in several sets, the detectors of the same set being staged vertically one above the other, a minimum critical heat flux ratio being calculated for each set of neutron flux detectors in-core at the aid of the deep neural network with the inputs determined using at least the neutron flux measurements made by the in-core neutron flux detectors (31) of said set;

- At the deviation calculation step, a set of deviations is calculated between each calculated critical heat flow ratio and the plurality of predetermined reference values;

- at the step of developing a control signal, producing a provisional control signal for each critical heat flux ratio calculated using the corresponding set of deviations, the provisional control signal being chosen from the list with at least the following values:

* automatic reactor shutdown or alarm;

* do nothing;

the control signal being constructed using at least more of the provisional control signals;

- the critical heat flux ratio is calculated by the deep neural network using the following formula:

where Y is the calculated critical heat flow ratio

X 0 is the vector of the inputs;

nH is the number of hidden layers of the deep neural network;

A, is a predetermined matrix of dimensions N, * N M , N, being the number of neurons of layer i and N M being the number of neurons of layer i-1;

s, is a predetermined nonlinear operator;

- A, is a matrix acting as an affine transformation on the components of the vector X M , the vector X M being the vector whose components are determined by the layer of neurons i-1;

- s, acts as a sigmoid function on all the components of the vector A, .C M , the vector C mi being the vector whose components are determined by the layer of neurons i-1;

the method comprises a design step of the deep neural network, the design step comprising the following sub-steps:

- constitution of a database comprising at least 100,000 reactor core states, each state being defined by a set of values ​​of said quantities characterizing the operation of the nuclear reactor and by a minimum critical heat flux ratio value calculated by a reference 3D thermo-hydraulic calculation code using the set of values, said states being chosen to cover a plurality of predetermined category 2 accident situations, the reactor control system being provided to deal with said accident situations;

- learning of the deep neural network, using at least part of the database;

- validation of the deep neural network using a part of the database independent of that used for the learning phase, and testing of the neural network obtained by coupling to accidental neutronics and thermal-hydraulics codes to simulate accidental transients requesting the intervention of the control system.

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

- a heart ;

- a control system;

- an information processing unit, configured to implement the operating method having the above characteristics.

Other characteristics and advantages will emerge from the detailed description which is given below, by way of indication and in no way limiting, with reference to the appended figures, among which:

- Figure 1 is a step diagram representing the method of the invention;

- Figure 2 is a simplified schematic representation of a nuclear reactor provided for implementing the method of Figure 1;

FIG. 3 is a schematic representation of the acquisition and calculation steps of the method of FIG. 1, for a first embodiment of the invention;

FIG. 4 is a schematic representation of the deep neural network used in the method of FIG. 1; and

FIG. 5 is a schematic representation of the acquisition and calculation steps of the method of FIG. 1, for a second embodiment of the invention.

The operating method represented schematically in FIG. 1 is intended to be implemented in a nuclear reactor of the type represented in FIG. 2.

This nuclear reactor is typically a pressurized water reactor (PWR).

As a variant, it is intended to be implemented on any other type of suitable nuclear reactor.

The nuclear reactor 1 has a core 3.

The core 3 typically comprises a plurality of nuclear fuel assemblies, of prismatic shape, placed against each other. Each nuclear fuel assembly comprises a plurality of nuclear fuel rods, held in position relative to one another by grids.

Each rod has a tubular cladding, made of a metal such as zircaloy, nuclear fuel pellets being stacked inside the cladding.

The nuclear reactor 1 also comprises a pressure vessel 5 containing the core 3 and at least one primary loop 7.

Typically, the nuclear reactor has several primary loops 7, for example three or four primary loops 7.

The or each primary loop 7 comprises hot and cold branches 9, 1 1 fluidly connected directly to the tank 5, a pressurizer 13 controlling the pressure in the primary loop 7, and a primary pump 15 ensuring the circulation of the primary fluid in the loop primary 7. The pressurizer 13 is generally common to the various primary loops.

Typically, the primary loop 7 also includes a steam generator 17, placed upstream of the primary pump 15.

The hot branch 9 fluidly connects a primary fluid outlet pipe from the tank 5 to the primary fluid inlet water box of the steam generator 17. The pressurizer 13 is pricked on the hot branch 9.

An intermediate branch 19 of the primary circuit, frequently referred to as a “U-shaped branch”, connects the primary fluid outlet water box of the steam generator 17 to the suction of the primary pump 15. The cold branch 1 1 connects a discharge from the primary pump 15 to a primary fluid inlet pipe of the tank 5.

The nuclear reactor 1 also includes organs for controlling the reactivity of the core 21.

These control units 21 are typically designated by the name “control clusters” and comprise a plurality of rods made of a material absorbing neutrons.

The nuclear reactor 1 also comprises mechanisms, not shown, configured to selectively move the control members 21 inside the vessel 5, so as to insert them or extract them outside the core 3.

The nuclear reactor 1 also includes a control system 23.

This control system 23 is typically a protection system, configured to cause an emergency shutdown of the nuclear reactor in the event of an accident situation.

To do this, the protection system 23 is configured to send an automatic shutdown order of the reactor to the drive mechanism of the control units 21. Such an order causes the rapid insertion of the control units into the core 3 of the reactor. nuclear.

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

According to another variant, the control system 23 plays the role of both a monitoring system and a protection system.

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.

Thus, the nuclear reactor 1 comprises temperature sensors 23, 25 respectively measuring the temperature of the primary fluid in the hot and cold branches 9, 1 1.

The nuclear reactor 1 comprises a sensor 27, measuring the pressure in the pressurizer 13.

The nuclear reactor also includes a sensor 29 measuring the speed of rotation of the primary pump 15.

The nuclear reactor 1 also includes neutron flux detectors 31.

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

The detectors 31 are distributed around the tank 5.

Typically, these are multi-stage ex-core chambers, with for example 6 sections arranged one above the other along the central axis of the tank.

The detectors 31 make it possible to measure the neutron flux emitted by the core 3 at different axial levels.

The nuclear reactor 1 comprises a sensor 33 for the position of the organs for monitoring the reactivity of the heart 21.

The nuclear reactor also comprises an information processing unit 35, configured to implement an operating method which will be described below.

The information processing unit is for example an element of the control system 23. It is typically formed of a processor and of a memory associated with the processor. As a variant, the information processing unit is produced in the form of programmable logic components, such as FPGAs (standing for Field-Programmable Gâte Array), or else in the form of dedicated integrated circuits, such as ASICs ( from English Application-Specific Integrated Circuit).

The various sensors 23, 25, 27, 29, 31, 33 provide information to the information processing unit 35.

The method of operating the nuclear reactor 1 according to the first embodiment will now be described, with reference to FIGS. 1, 3 and 4.

The process comprises the following steps:

- step S10: acquisition of a plurality of quantities characterizing the operation of nuclear reactor 1;

- step S12: calculation of at least one critical heat flow ratio (RFTC) using a deep neural network 37, shown in FIG. 4, the inputs of the deep neural network 37 being determined using the acquired quantities, the deep neural network 37 comprising at least two hidden layers of at least five neurons each;

- step S14: calculation of the differences between the at least one critical heat flow ratio (RFTC) calculated and a plurality of predetermined reference values;

- step S16: production of a control signal for the reactor control system 23, using the calculated deviations, the control signal being chosen from the list comprising at least the following values:

* automatic reactor shutdown or alarm;

* do nothing;

- step S18: emergency shutdown of the reactor by the control system 23, when the control signal produced has the value "automatic shutdown of the reactor", or emission of an alarm signal by the control system 23 when the The generated control signal has the value "alarm".

The quantities acquired include at least neutron flux measurements carried out by the neutron flux detectors 31, making it possible to reconstruct the axial power distribution P (z) in the core 3, as described below.

The quantities acquired also include, for the or each primary loop 7, one or more of the following quantities: pressure p of the pressurizer, flow rate Q of primary fluid in the primary loop, temperature of primary fluid in the hot and cold branches Tbc and Tbf.

Typically, all of the above quantities are acquired.

The pressure of the pressurizer p and the temperatures of the primary fluid in the hot and cold branches Tbc and Tbf are supplied directly by the sensors 27, 23, 25.

The flow rate of primary fluid Q is reconstructed by using the speed of rotation of the primary pumps, measured by the sensor 29. It is determined by calculation, for example using a relation directly indicating the flow of primary fluid from the speed of rotation .

The quantities acquired also include the current insertion position of the organs for monitoring the reactivity of the heart 21, supplied by the detector 33

Thus, as illustrated in FIGS. 1 and 3, the acquisition step S10 typically comprises a sub-step S20 for reading the value supplied by the sensors 23, 25, 27, 29, 31, 33, certain quantities characterizing the operation of the reactor being read directly on these sensors. These are in particular the pressure p of the pressurizer, and the hot and cold branch temperatures Tbc and Tbf.

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 typically concerns the reconstruction of the flow rate Q of primary fluid, or the application of filters on the quantities acquired in step S20.

The step of calculating at least one RFTC S12 comprises a substep S24 of determining the inputs of the deep neural network 37 using the acquired quantities, and a substep S25 of calculating the at least one RFTC at using the deep neural network 37 using the inputs determined in substep S24.

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

The axial power distribution P (z) is a vector having a plurality of components, each component corresponding to an average power level of the nuclear reactor at a given axial dimension. Each component defines one of the inputs of the deep neural network.

The number of components depends on the degree of precision desired for the reconstruction of the power distribution, and therefore on the desired precision of the calculation of the RFTC. For example, the axial power distribution has as many components as there are sections in each detector 31: the signals coming from the different sections of the detectors in this case define inputs of the deep neural network.

As a variant, the axial power distribution has more components than the number of sections in each detector 31. A module for reconstructing the axial power from the values ​​coming from the different sections of the detectors provides from a dedicated algorithmic processing ( for example a polynomial interpolation) the more finely meshed axial power distribution of which each component defines one of the inputs of the deep neural network.

In the sub-step S24, the power of the nuclear reactor 1 at a given axial level is therefore calculated from the measured neutron fluxes, for example using analytical relationships directly giving the power level as a function of the neutron flux and of the power thermal released by the Pth core, reconstructed as described below.

The inputs of the neural network 37 also include the thermal power released by the core Pth.

In sub-step S24, the thermal power released Pth is determined by calculation, using at least the pressure of the pressurizer p, the flow rate of primary fluid Q, and the temperatures of the primary fluid in the hot and cold branches Tbc and Tbf.

The calculation method is known and will not be detailed here.

Typically, a calculation of the thermal power Pth released for each primary loop 7 is carried out, using the quantities acquired for said primary loop 7.

The inputs to the neural network 37 also include the enthalpy elevation factor FAH.

In substep S24, the enthalpy elevation factor FAH is determined by carrying out the following operations:

- Calculation of a radial peak power factor Fxy (z), using the current insertion position of the organs for controlling the reactivity of the heart 21 acquired;

- Calculation of the enthalpy elevation factor FAH using the radial peak power factor Fxy (z) calculated, the reconstructed axial power distribution P (z) and the determined thermal power Pth.

The radial peak power factor Fxy (z) and the enthalpy rise factor FAH are calculated using known formulas, which will not be detailed here.

Preferably, an enthalpy rise factor FAH is calculated for each primary loop 7, using the released thermal power Pth calculated for said primary loop 7.

When the nuclear reactor 1 comprises several primary loops 7, a critical heat flux ratio RFTC is preferably calculated in the sub-step S25 for each primary loop 7, using the deep neural network 37. To do this, the deep neural network 37 uses the inputs determined with at least the magnitudes acquired for the corresponding primary loop 7.

As a variant, a single RFTC critical heat flux ratio is calculated using the deep neural network 37. It is for example calculated using inputs determined by any processing of the quantities acquired for the primary loops, suitable for the evaluation. of the RFTC.

As visible in Figure 4, the deep neural network 37 comprises an input layer 39 located to the left of Figure 4, an output layer 41 located to the right of Figure 4, and a plurality of intermediate layers 43, called still hidden layers. The number nH of intermediate layers 43 is greater than or equal to two, preferably greater than or equal to five, and more preferably greater than or equal to ten.

This makes it possible to account for a large number of different situations in the nuclear reactor.

Each hidden layer 43 comprises at least five neurons 45, preferably at least seven neurons 45, and more preferably at least ten neurons 45.

The input layer 39 has as many input neurons as there are inputs.

For example, the input layer 39 comprises an input neuron for each of the inputs p, Q, Tbf, Pth, FAH. It also includes an input neuron for each of the components of the axial power distribution P (z).

Thus, the input layer 39 comprises for example eleven input neurons, if the axial power distribution P (z) is a vector with six components.

The output layer 41 comprises a single output neuron, corresponding to the minimum value of the RFTC calculated for the corresponding primary loop.

Advantageously, the inputs are the subject of a first post-processing before being introduced into the neural network 37 during step S22 of FIG. 1.

The first post-processing consists for example of an application of dynamic compensation terms, typically lead-delay filters, allowing each quantity acquired to be the image of the corresponding physical parameter. This post-processing makes it possible to compensate for the response time of the instrumentation (temperature measurements of the cold and hot branches of the vessel), the fall time of the control organs of the reactivity of the heart, and the response time of the processes and algorithms. treatment. The functional nature of the processing of information by the neural network 37 makes it possible to considerably reduce the need to use filters, compared with conventional algorithms.

The inputs are preferably the subject of a second post-processing before being introduced into the neural network, by normalization of each of the quantities or components.

The inputs constitute a vector, called the input vector and noted X 0 . Each hidden layer i of the neural network develops a vector of values ​​X i by processing the vector of values ​​X M originating from the immediately preceding layer i-1.

Thus, for each hidden layer i, i being between 1 and nH,

Xi = Gi- Ai- Xi- !

where A, is a predetermined matrix of dimensions N, * N, _i, N, being the number of neurons of layer i and N M being the number of neurons of layer i-1;

s, is a predetermined nonlinear operator.

In other words, the critical heat flux ratio is calculated by the deep neural network 37 using the following formula:

where Y is the calculated critical heat flow ratio.

Each matrix A, acts as an affine transformation on the vector components X M . In other words:

where W, is the matrix of synaptic weights which connect the N, neurons of layer i to N, _i neurons of layer i-1, and b, is the vector of N-dimensional biases, of layer i.

For the last layer it will be noted that the matrix A nH + i is of dimension 1 x N nH and that s hH + i = 1: ü there is no application of a nonlinear transfer function to establish the signal Y of the last coat.

S nonlinear operators, each act as a nonlinear function, e.g. sigmoid on all components of the vector A, C M . A sigmoid function is a function expressed as follows:

where I is a specially parameterized constant.

In step S14, for each primary loop, a set of deviations is calculated between the critical heat flux ratio calculated for said primary loop and the plurality of predetermined reference values. As many difference calculations are thus carried out as there are primary loops.

These values ​​are operating reference thresholds, called protection (or monitoring) thresholds, which integrate process uncertainties, namely measurement uncertainties and calculation uncertainties.

Typically, the calculated deviations are displayed in alphanumeric or graphic form on at least one screen, for real-time monitoring of the protection and operating margins of the unit, for each primary loop.

When a single RFTC is calculated, a single set of deviations is calculated in step S14.

Step S16 makes it possible to interpret the sets of deviations observed in step S14.

To do this, a provisional control signal is produced for each primary loop, using the deviation set calculated for said loop.

When the control system is a protection system, the provisional control signal is chosen from the list comprising at least the following values: automatic shutdown of the reactor, do nothing.

When the control system is a surveillance system, the temporary control signal is chosen from the list comprising at least the following values: alarm, do nothing.

When the control system is a protection and surveillance system, the provisional control signal is chosen from the list comprising at least the following values: automatic reactor shutdown, alarm, do nothing.

The control signal sent to the reactor control system 23 is developed using the provisional control signals of all the primary loops.

After having interpreted the differences observed in step S14, step S16 consists in applying vote logics for the signals produced for each loop of the reactor. For example, if the majority of the temporary control signals have the value “automatic reactor shutdown”, then the control signal sent to the control system 23 is “automatic reactor shutdown”.

In step S18, the control system 23 receives the control signal generated in step 16, and acts accordingly. For example, it triggers an automatic shutdown of the reactor 1 by inserting the reactivity control members of the core 21 if the control signal is “automatic reactor shutdown”, or the emission of an alarm signal if the signal elaborated command has the value "alarm".

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

- S28: creation of a database comprising at least 100,000 reactor states,

- S30: learning of the deep neural network, using at least part of the database.

Advantageously, the database comprises at least 500,000 states of the reactor, preferably at least 1,000,000 states of the reactor.

Each state is defined by a set of values ​​of the quantities characterizing the operation of the nuclear reactor, and by a critical heat flow ratio value. Said critical heat flux value is calculated by a reference 3D thermo-hydraulic calculation code using the corresponding set of values ​​of the quantities characterizing the operation of the nuclear reactor.

The 3D core thermo-hydraulic computer code is a high precision code, for example the FLICA code. Typically, Monte Carlo sampling techniques are used to generate the database.

The states are chosen to cover a plurality of predetermined accident situations leading to triggering the intervention of the generic low RFTC protection chain of the reactor.

The predetermined accident situations are, for example, those which are said to be category 2 in the nuclear reactor safety report, which require the intervention of the generic low RFTC protection chain.

The ranges of variations of the quantities characterizing the operation of the nuclear reactor are also chosen so that they cover the range of validity of the critical heat flux correlation.

The data is generated in such a way that it optimizes the detection of physical symmetry during the learning phase by the neural network.

Preferably, the input variables are normalized.

The learning sub-step S30 is performed on only part of the database constituted in the sub-step S28. For example, it uses between 50% and 90% of the database, typically 80%.

The learning phase includes the following operations:

i) Determination of the optimal neuronal structure by carrying out a sensitivity study dedicated to the choice of model hyperparameters, this step may require the use of an optimization algorithm (of the genetic algorithm type): number of hidden layers, number of neurons by layer, choice of the transfer function imposed by the phenomenology of the phenomenon of boiling crisis (for example differentiable and having to present a strong non-linear character because of the analytical expression of the correlation of critical flux);

ii) Development of a learning algorithm optimized specifically for the generated database, allowing to accelerate and stabilize the convergence process, and to improve convergence towards more robust solutions by avoiding local minima. This improves the generalization capacities of the neural network and therefore its capacity to respond positively to the validation and testing phases described below.

This is achieved for example by developing an adaptive batch stochastic gradient descent algorithm. The update at each cycle of the training of the tensor of the synaptic weights and of the biases is carried out on batches of data of evolving size by scanning all the training base, parameterized by the number of the cycle.

Moreover, methods of inertia on the adaptation step of the learning algorithm are advantageously used, in order to improve the efficiency and the robustness of the learning phase.

Preferably, step S26 further comprises a substep S32 for validating the deep neural network determined in substep S30. This validation phase is performed on a second part of the database, which was not used in sub-step S30. Typically, all the part of the database which has not been used in step S30 is used in step S32. This makes it possible to validate the predictions of the deep neural network developed at the end of step S30 by comparison with the predictions generated with the reference code, and to characterize the capacity for generalization of the predictive model thus developed. Sensitivity studies are preferably performed during substep S32. If the validation phase is not satisfactory, we return to sub-step S30, then we modify the neuronal structure (ie

Preferably, step S26 also comprises a sub-step S34 of phenomenological validation (test) of the deep neural network on data independent of the database determined in sub-step S28.

During the sub-step S34, the deep neural network is implemented in coupling with calculation codes simulating accidental transients. These accidental transients are typically those defined in the nuclear reactor safety report.

If this final test phase is not satisfactory, the sub-step S30 is repeated.

Thus, the process described above is particularly suitable for being implemented by a reactor of the type described above.

Conversely, the nuclear reactor 1 described above is particularly suitable for implementing the process which has just been described.

In particular, the information processing unit 35 comprises, as illustrated in FIG. 2:

a module 47 for acquiring a plurality of quantities characterizing the operation of the nuclear reactor;

a module 49 for calculating at least one critical heat flow ratio using the deep neural network 37, configured to determine the inputs of the deep neural network 37 using the acquired and determined quantities, the neural network deep 37 comprising at least two hidden layers of at least five neurons each;

a module 51 for calculating the differences between the at least one calculated critical heat flow ratio and a plurality of predetermined reference values;

a module 53 for developing a control signal for the reactor control system using the calculated deviations, the control signal being chosen from the list comprising at least the following values:

* automatic reactor shutdown or alarm;

* do nothing.

The reactor control system 23 is configured to cause an automatic shutdown of the reactor when the generated control signal has the value “automatic reactor shutdown”. The reactor control system 23 is configured to cause the emission of an alarm when the generated control signal has the value “alarm”.

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

A second embodiment of the invention will now be described, with reference to FIG. 5. Only the points by which the second embodiment differs from the first will be detailed below. Elements that are identical or perform the same functions will be designated by the same references.

In the second embodiment, the neutron flux detectors 31 are not ex-core detectors placed outside the core 3. The neutron flux detectors 31 are in-core detectors, permanently inserted in the core 3.

They are distributed in several sets, the detectors of the same set being staggered vertically one above the other.

The detectors are for example collectrons, arranged on rods vertically in the heart. Each collectron rod makes it possible to measure the neutron flux at several axial heights in the core, for example six axial heights for a vertical arrangement of 6 collectrons. The nuclear reactor is typically equipped with 12 rods of 6 collectrons each, distributed in the core.

The thermal power Pth released is determined by calculation, using the acquired neutron flux measurements. This calculation no longer uses the pressure of the pressurizer p, the flow rate Q of the primary fluid, the temperatures of the primary fluid in the hot and cold branches Tbc, Tbf. The calculation formulas used are known and will not be detailed here.

The enthalpy elevation factor FAH is reconstructed using the axial power distribution P (z). It is not necessary to reconstitute the factor Fxy (z) from the measurement provided by the position sensor 33 of the organs for monitoring the reactivity of the heart 21.

In step S12, an RFTC critical heat flux ratio is calculated for each set of 31 in-core neutron flux detectors, using the deep neural network 37, with the inputs determined using the neutron flux measurements performed by said 31 in-core neutron flux detectors.

From the neutron flux measurements at several axial heights provided by the set of detectors, a set of parameters P (z), Pth and FAH, specific for this set of detectors, is determined. These parameters are used as inputs for the neural network 37.

Other inputs are also used, for example the pressure of the pressurizer p, the flow rate Q of the primary fluid, the temperature of the primary fluid in the cold branch Tbc, Tbf. These input quantities are the same for all sets of detectors.

In the step of calculating the deviations S14, a set of deviations is calculated between each previously determined critical heat flow ratio and the plurality of predetermined reference values.

In the step of developing a control signal S16, a provisional control signal is produced for each calculated critical heat flow ratio, using the corresponding set of deviations.

The control signal is produced using at least several of the provisional control signals, according to the voting logic described above.

A third embodiment of the invention will now be briefly described. Only the points by which the third embodiment differs from the second will be detailed below. Elements that are identical or perform the same functions will be designated by the same references.

In the third embodiment, the information processing unit 35 has a map of the power in the reactor core, in 3 dimensions, continuously.

This mapping constitutes one of the quantities characterizing the operation of the reactor. The other quantities, for example Tbf, p and Q, are acquired as in the first and second embodiments.

Certain inputs of the deep neural network 37 are determined by calculation, using the power distribution map in the heart in 3D, thus replacing the quantities P (z), Pth and FAH previously described for the first embodiment of the invention.

Advantageously, the components of the 3D power distribution constitute the inputs of a dedicated neural structure, called the convolution layer which makes it possible to optimize the extraction of characteristics of the power distribution to develop the deep neural network, and in particular of '' optimize its structure for a hardware implementation. The output of this convolutional layer defines part of the inputs of the deep neural network 37. This output notably replaces the vector of the components of the axial power distribution at the input of the deep neural network 37, as well as the magnitude FAH. This neural structure adapted to process spatially structured data, replaces the matrix multiplication by a mathematical operation of convolution between the power distribution at the input of the structure and a series of cores (or filters) whose degrees of freedom are adjusted during the learning of step S30. The convolution layer also includes a nonlinear processing of the outputs of the convolution through nonlinear transfer functions, as well as a third step allowing to introduce invariants under local geometric transformations (for example translations). These steps can be repeated when developing the optimal neural structure. The convolution layer also includes a nonlinear processing of the outputs of the convolution through nonlinear transfer functions, as well as a third step allowing to introduce invariants under local geometric transformations (for example translations). These steps can be repeated when developing the optimal neural structure. The convolution layer also includes a nonlinear processing of the outputs of the convolution through nonlinear transfer functions, as well as a third step allowing to introduce invariants under local geometric transformations (for example translations). These steps can be repeated when developing the optimal neural structure.

The operating method of the invention can have multiple variants.

The quantities characterizing the operation of the nuclear reactor, acquired in step S10, and the inputs of the deep neural network, may not exactly correspond to the list described above. Other quantities can be acquired. Some quantities may not be acquired. Some inputs may not be used. Other entries can be added.

In any case, the power distribution significantly impacts the location in the core and the value of the minimum RFTC margin.

Taking it into account makes it possible to improve the precision of the reconstruction of the RFTC. The quantities characterizing the operation of the nuclear reactor are not necessarily acquired in the manner described above and shown in FIGS. 2, 3 and 5. Certain quantities could not be read directly by the sensors fitted to the nuclear reactor. They could be developed from other measured quantities or from values ​​obtained from the operating system of the nuclear reactor.

The convolutional layer described in the context of the third embodiment could also be implemented in the first two embodiments in order to process the axial distribution of power P (z) in the neuronal structure. The other inputs of the neural network being unchanged.

The process described above has multiple advantages.

The response time is extremely short, and is for example of the order of a millisecond. This is obtained in particular from the fact that the method does not require the setting

using convergence loops, only the parameters specific to the deep neuron network being necessary, the unit computation operations are simple and can be easily parallelized if the hardware solution allows it.

As a result, it frees up computing capacity for other parts of the nuclear reactor control and instrumentation system.

These performances are obtained from the fact that the complexity of the underlying physics is encoded by the parameterization of the deep neural network from the database. This allows the deep neural network to reconstruct the RFTC value of the reference thermo-hydraulic code, simply using the synaptic weights and transfer functions of the neural structure.

The calculation of RFTC provided by the deep neural network is very reliable, and the answer is close to that provided by the reference 3D thermo-hydraulic code (here FLICA). The conservatism of the response is adjustable.

The quality of the response provided by the deep neural network is notably due to the fact that no reconstruction of intermediate local thermohydraulic variables is carried out without fine physical modeling. The reliability of the response provided depends on the fineness of the mesh of the database used for training the neural network. This only costs calculation time, before commissioning in the nuclear reactor. The method is also more robust to the propagation of uncertainties on the input data (initiated by random fluctuations in measurements from core instrumentation) which improves the validation and material qualification step.

CLAIMS

1.- A method of operating a nuclear reactor (1) comprising a core (3), the method comprising the following steps:

- acquisition of a plurality of quantities characterizing the operation of the nuclear reactor (1);

- calculation of at least one critical heat flux ratio using a deep neural network (37), the inputs of the deep neural network (37) being determined using the acquired quantities, the deep neural network comprising at least two hidden layers of at least five neurons each;

- Calculation of the differences between the at least one calculated critical heat flow ratio and a plurality of predetermined reference threshold values;

- Development of a control signal for a control system (23) of the reactor, using the calculated deviations, the control signal being chosen from the list comprising at least the following values:

* automatic reactor shutdown or alarm;

* do nothing;

- emergency shutdown of the nuclear reactor (1) by the control system (23) when the control signal produced has the value "automatic reactor shutdown", or emission of an alarm signal by the control system ( 23) when the generated control signal has the value “alarm”.

2.- The method of claim 1, wherein said acquired quantities comprise neutron flux measurements performed by neutron flux detectors (31), the inputs of the neural network (37) comprising the axial power distribution (P (z )) in the core of the nuclear reactor, reconstructed from the acquired neutron flux measurements.

3.- The method of claim 2, wherein the axial power distribution (P (z)) is a vector having a plurality of components, each component corresponding to an average power of the nuclear reactor per axial section at a given axial dimension, each component defining one of the inputs of the deep neural network (37).

4. A method according to claim 2 or 3, wherein the neutron flux detectors (31) are ex-core detectors arranged outside the core (3).

5. A method according to claim 2 or 3, wherein the neutron flux detectors (31) are in-core detectors permanently inserted in the heart (3).

6.- Method according to any one of the preceding claims, wherein the nuclear reactor (1) comprises a pressure vessel (5) containing the core (3) and at least one primary loop (7), the primary loop (7). ) comprising hot and cold branches (9, 1 1) fluidly connected directly to the tank (5), a pressurizer (13) controlling a pressure in the primary loop (7), and a primary pump (15) ensuring the circulation of primary fluid in the primary loop (7), the quantities acquired comprising, for the or each primary loop (7), one or more of the following quantities: pressure of the pressurizer (p), flow rate (Q) of primary fluid in the primary loop (7), temperature of the primary fluid in the hot and cold branches (Tbc, Tbf).

7. A method according to claim 6, in which the inputs of the neural network (37) comprise one or more of the following quantities: pressure of the pressurizer (p), flow rate (Q) of primary fluid in the primary loop (7), temperature of the primary fluid in the cold leg (Tbf).

8. A method according to any preceding claim, wherein the inputs of the neural network (37) include the thermal power released by the heart (Pth).

9. A method according to claim 8 combined with claim 6 or 7, wherein the thermal power (Pth) released is determined by calculation, using at least the pressure of the pressurizer (p), the primary fluid flow (Q) , the temperatures of the primary fluid in the hot and cold branches (Tbc, Tbf), acquired for at least one primary loop (7).

10. A method according to claim 8 combined with claim 5, wherein, the thermal power (Pth) released being determined by calculation, using the acquired neutron flux measurements.

1 1. A method according to any preceding claim, wherein the inputs of the neural network (37) include the enthalpy rise factor (HDR).

12.- The method of claim 1 1 combined with claim 4 and claim 9, wherein the quantities acquired comprise a current insertion position of the organs for controlling the reactivity of the heart (21), the elevation factor. enthalpy (FAH) being determined by carrying out the following operations:

- calculation of a radial peak power factor (Fxy (z)) using the current acquired insertion position;

- calculation of the enthalpy rise factor (FAH) using the radial peak power factor (Fxy (z)) calculated, the reconstructed axial power distribution (P (z)) and the thermal power released (Pth) determined.

13. A method according to claim 11 combined with claim 5, wherein the enthalpy elevation factor (FAH) is reconstructed using the axial power distribution (P (z)).

14.- Method according to any one of claims 6 to 7, wherein the nuclear reactor (1) comprises several primary loops (7), a critical heat flow ratio being calculated for each primary loop (7) using of the deep neural network (37) using at least said quantities acquired for said primary loop (7).

15.- The method of claim 5, wherein the detectors (31) are distributed in several sets, the detectors (31) of the same set being staged vertically one above the other, a minimum critical heat flow ratio being calculated for each set of in-core neutron flux detectors (31) using the deep neural network (37) with the inputs determined using at least the neutron flux measurements made by the neutron flux detectors (31 ) in-core of said set.

16.- The method of claim 14 or 15, wherein

- At the deviation calculation step, a set of deviations is calculated between each calculated critical heat flow ratio and the plurality of predetermined reference values;

- at the step of developing a control signal, producing a provisional control signal for each critical heat flux ratio calculated using the corresponding set of deviations, the provisional control signal being chosen from the list with at least the following values:

* automatic reactor shutdown or alarm;

* do nothing;

the control signal being constructed using at least more of the provisional control signals.

17. A method according to any preceding claim, wherein the critical heat flux ratio is calculated by the deep neural network (37) using the following formula:

where Y is the calculated critical heat flow ratio

X 0 is the vector of the inputs;

nH is the number of hidden layers of the deep neural network;

A, is a predetermined matrix of dimensions N, * N M , N, being the number of neurons of layer i and N M being the number of neurons of layer i-1;

s, is a predetermined nonlinear operator.

18.- The method of claim 17, wherein A is a matrix acting as an affine transformation on the components of the vector X M , the vector X M being the vector whose components are determined by the layer of neurons i-1.

19.- The method of claim 17 or 18, wherein s acts as a sigmoid function on all the components of the vector A, .C mi , the vector X M being the vector whose components are determined by the layer of neurons i -1.

20. A method according to any one of the preceding claims, wherein the method comprises a step of designing the deep neural network (37), the step of designing comprising the following sub-steps:

- constitution of a database comprising at least 100,000 reactor core states, each state being defined by a set of values ​​of said quantities characterizing the operation of the nuclear reactor (1) and by a minimum critical heat flux ratio value calculated by a reference 3D thermo-hydraulic calculation code using the set of values, said states being chosen to cover a plurality of predetermined category 2 accident situations, the control system (23) of the reactor being provided to deal with said situations accidental;

- learning of the deep neural network (37), using at least part of the database;

- validation of the deep neural network using a part of the database independent of that used for the learning phase, and testing of the neural network obtained by coupling to accidental neutronics and thermal-hydraulics codes to simulate accidental transients requesting the intervention of the control system (23).

21 .- Nuclear reactor (1) comprising:

- a heart (3);

- a control system (23);

- an information processing unit (35), configured to implement the operating method of any one of the preceding claims.