Method and device for controlling an internal combustion engine having a plurality of burners
By detecting combustion measurement data and training machine learning models, optimizing the fuel and air mixing of each burner in the internal combustion engine, the temperature fluctuations and pollution problems caused by individual burners are solved, and the efficiency and power of the internal combustion engine are improved.
Patent Information
- Application Number
- CN202080029388.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2019-04-16
- Filing Date
- 2020-03-19
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2040-03-19
AI Technical Summary
The individual differences between multiple burners in existing internal combustion engines are difficult to effectively optimize, resulting in temperature fluctuations, pollutant emissions and efficiency problems. The optimization goals are opposite to each other and difficult to achieve approximately.
By detecting combustion measurement data, assigning burner identifiers, machine learning models are trained to generate burner-specific control data to optimize fuel-air mixing ratio, fuel supply flow, and injection pressure, and coordinate differences between burners. Use multiple valves to divide fuel levels and perform intensive learning training.
A specific optimization for each burner is achieved, reducing temperature and pressure limits, reducing pollutant emissions and wear, improving efficiency and power, enabling accurate modeling of complex interdependencies to adapt to different operating states.
Smart Images

Figure CN113646586B_ABST
Abstract
Description
Background Art
[0001] Modern internal combustion engines such as gas turbines, jet engines or internal combustion engines usually have multiple burners, which generate high temperatures and volume flows in separate or common combustion chambers, which are converted into kinetic energy. During the combustion process, pressure oscillations are usually excited and pollutants such as nitrogen oxides, carbon monoxide or unburned hydrocarbons are generated or emitted.
[0002] In many internal combustion engines, the combustion process in the burners is structured by adjusting different fuel levels. Here, fuel flows with different flow rates, inlet pressures and / or fuel-air mixture ratios are delivered in parallel to the corresponding burners in a manner differentiated into different fuel levels. For each fuel level, a separate annular pipeline is usually provided, which leads to all burners. By appropriately adjusting the individual fuel levels, the operation of the internal combustion engine can be optimized. Here, it is usually sought to reduce pressure fluctuations or temperature peaks in the combustion chamber, clogging of the fuel nozzles and / or pollutant emissions, as well as to increase the power or efficiency of the internal combustion engine. However, the optimization goals pursued are usually mutually contradictory and can only be achieved approximately. Summary of the invention
[0003] The object of the present invention is to provide a method and a device for controlling an internal combustion engine having a plurality of burners, which method and the device allow an improved optimization of the operation of the internal combustion engine.
[0004] This object is achieved by a method having the features of the invention, by a device having the features of the invention, by a computer program product having the features of the invention and by a computer-readable storage medium having the features of the invention.
[0005] According to the invention, in order to control an internal combustion engine having a plurality of burners, combustion measurement data are detected for the respective burners in a burner-specific manner and assigned to a burner identifier that identifies the respective burner. The internal combustion engine can in particular be a gas turbine, a jet engine or an internal combustion engine. Here, the respective burner can also have a plurality of combustion positions and / or be designed as a burner group. In addition, performance measurement data of the internal combustion engine are detected, and a performance value is determined based on the performance measurement data. A machine learning model is trained with the aid of the combustion measurement data, the assigned burner identifier and the performance measurement data to generate burner-specific control data, which optimizes the performance value when the burner is controlled in a burner-specific manner by the control data. Here, optimization should also be understood in particular as an approximation of an optimal value determined by a predetermined criterion. Such a criterion can be evaluated for the performance value in a so-called reward function, a loss function or a cost function. The control data generated by the trained machine learning model are output for controlling the burner in a burner-specific manner.
[0006] In particular, pressure fluctuations in the combustion chamber, pollutant emissions, gas temperature, spatial temperature distribution, fuel feed flow, mixture ratio or other parameters of the combustion process can be detected as combustion measurement data in a burner-specific manner. Performance values can relate in particular to the power or efficiency generated by the internal combustion engine, pressure fluctuations in the combustion chamber, temperature distribution, gas inlet temperature, local temperature peaks, pollutant emissions, fuel nozzle blockage, fuel consumption or wear of the internal combustion engine. By means of the control data, in particular, the mixture ratio between fuel and air, the fuel feed flow or the injection pressure can be controlled in a burner-specific manner.
[0007] In order to carry out the method according to the invention, a device for controlling an internal combustion engine, a computer program product and a computer-readable storage medium are provided.
[0008] The method according to the invention and the device according to the invention can be executed or implemented, for example, by means of one or more processors, computers, application specific integrated circuits (ASICs), digital signal processors (DSPs) and / or so-called field programmable gate arrays (FPGAs).
[0009] One advantage of the invention is in particular that, by means of burner-specific measurement data detection and performance optimization, individual differences between the individual burners, which may be caused by manufacturing tolerances or different line routings, can be compensated in a coordinated manner. In this way, temperature limits, pressure or pollutant limits, material fatigue, wear or pollutant emissions can be reduced in many cases, and / or the efficiency or the generated power can be increased. By using machine learning models, complex interdependencies or correlations between burner-specific combustion measurement data, performance measurement data and control data optimized in a burner-specific manner can also be modeled relatively accurately. This also applies in particular to different operating states of the internal combustion engine.
[0010] Advantageous embodiments and developments of the invention are described below.
[0011] In an advantageous embodiment of the invention, the fuel supply flow to the respective burner can be divided into a plurality of fuel levels in a burner-specific manner according to the control data. By means of the different fuel levels, a plurality of fuel flows can be directed to each burner, which differ in their fuel / air mixture ratio, their introduction pressure or other introduction parameters. In this way, the combustion process can be structured spatially, temporally, physically and / or chemically in each burner and thus optimized in a burner-specific manner, in particular with regard to performance values.
[0012] Preferably, the fuel supply flow can be divided into the fuel stages by means of a burner-specific and / or fuel-stage-specific multi-way valve. By using a multi-way valve, the line expenditure for the internal combustion engine can usually be significantly reduced.
[0013] Advantageously, in order to briefly interrupt or reduce the fuel supply flow, the valve across the burners is actuated. In this way, a plurality of burners can be quickly adjusted downwards together, independently of their burner-specific actuation.
[0014] According to an advantageous development of the invention, the combustion measurement data can be detected in a fuel-grade-specific manner and respectively assigned to a fuel-grade identifier, which is additionally used for training the machine learning model. In this way, fuel-grade-specific interdependencies between the combustion measurement data and the performance measurement data or control data can also be modeled by the machine learning model. These interdependencies can advantageously be taken into account when optimizing the performance values.
[0015] According to a particularly advantageous embodiment of the present invention, the machine learning model can be trained by a reinforcement learning method. This reinforcement learning method is also generally referred to as reinforcement-learning. As a reinforcement learning method, for example, the so-called PGNRR method (PGNRR: Policy Gradient Neural Rewards Regression), the VOP-PGNRR method (VOP-PGNRR: Variable Objective Policy PGNRR, variable target policy PGNRR) or a model-based reinforcement learning method, such as the RCNN method (RCNN: Recurrent Control Neural Network, recursive control neural network) can be used. In particular, one or more so-called closed control strategies can be generated by the training. Such a closed control strategy can usually be executed very quickly and therefore allows a high temporal resolution when controlling the internal combustion engine.
[0016] The machine learning model and / or control strategy may preferably include an artificial neural network, a recursive neural network, a convolutional neural network, a Bayesian neural network, an autoencoder, a deep learning architecture, a support vector machine, a data-driven trainable regression model, a k-nearest neighbor classifier, a physical model and / or a decision tree.
[0017] According to another advantageous embodiment of the invention, the machine learning model can be further trained during the ongoing operation of the internal combustion engine based on currently detected combustion measurement data and performance measurement data. This allows continuous operation optimization and allows adaptation to changes in the internal combustion engine caused by operating time or environment.
[0018] Furthermore, the machine learning model can be pre-trained during the calibration phase with the aid of a training internal combustion engine and / or with the aid of a simulation model of the internal combustion engine. Such a simulation model is often also referred to as a design model or digital twin. As a training internal combustion engine, a model of an internal combustion engine with the same or similar structure can preferably be used. Through such pre-training, the training of the internal combustion engine can usually be significantly shortened.
[0019] The machine learning model can advantageously include a plurality of sub-models specific to respective burners, each sub-model generating control data specific to respective burners. In this way, the training task can be decomposed into smaller sub-tasks, the interdependencies of which are generally easier to model and which can generally be executed faster.
[0020] In particular, submodels can be executed in parallel to generate burner-specific control data. In this way, execution times can often be significantly reduced.
[0021] According to an advantageous extension of the invention, a further machine learning model can be trained to generate further control data, by means of which the fuel flow distribution between the burners is set, in particular the total fuel flow distribution and / or the combustion temperature is set and the performance value and / or further performance value is optimized. Combustion measurement data, assigned burner identifiers and performance measurement data can be used as training data. In addition, simulation data from a simulation of an internal combustion engine or a part thereof can be used to train the further machine learning model. The use of simulation data is advantageous because, in general, relatively little suitable training data is available for such a cross-burner control process. The setting of the fuel flow distribution between the burners is particularly advantageous for annular combustion chambers, in which the burners are arranged in a common annular combustion chamber. Since each burner usually affects the entire annular combustion chamber, fluctuations across the burners can be compensated by controlling the fuel distribution across the burners. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] The embodiments of the present invention will be explained in more detail below based on the accompanying drawings. In the schematic diagrams:
[0023] Figure 1 A gas turbine with multiple burners is shown,
[0024] Figure 2 A diagram is shown for illustrating the fuel supply in a gas turbine according to the prior art.
[0025] Figure 3 A diagram is shown for illustrating the fuel supply in the case of a gas turbine according to the invention, and
[0026] Figure 4A device for controlling a gas turbine according to the invention is shown. DETAILED DESCRIPTION
[0027] Figure 1 The schematic diagram shows a gas turbine GT as an internal combustion engine. The gas turbine GT has a compressor V for compressing the inflowing air, a plurality of burners BN1 and BN2 for burning the supplied fuel, and a turbine T for converting the heat energy and kinetic energy generated by the combustion into rotational energy. The rotational energy is transmitted to the compressor V in particular via a drive shaft AW to drive the compressor V.
[0028] The application of the invention to a gas turbine is to be understood as an example only. Alternatively or additionally, the invention can also be used to control other internal combustion engines, such as jet engines or internal combustion engines.
[0029] Figure 2 A diagram is shown for explaining the fuel supply in a gas turbine according to the prior art. The gas turbine has a plurality of burners, which for the sake of clarity are shown in FIG. Figure 2 Only two of the burners BN1 and BN2 are shown in FIG. The gas turbine also has multiple fuel stages, of which, for the sake of clarity, Figure 2 Likewise, only two of the fuel levels FSA and FSB are explicitly indicated in the figure. In the different fuel levels FSA and FSB, a plurality of different fuel flows are introduced into the corresponding burner BN1 or BN2 via separate fuel nozzles. The fuel flows can differ from one another, for example, in their flow rate, fuel-air mixture ratio or introduction pressure. In this way, in each burner BN1 or BN2, the combustion process can be structured spatially, temporally, physically and / or chemically and thus in particular optimized.
[0030] According to the prior art, separate, stage-specific annular lines FRA and FRB are provided for the combustion stages FSA and FSB, respectively, which lead to all burners BN1 and BN2, respectively. The annular lines FRA and FRB are connected to the main supply line MSL of the gas turbine via controllable, fuel stage-specific valves VA and VB, respectively. The valves VA and VB divide the total fuel flow F to the individual fuel stages FSA and FSB.
[0031] Due to the delivery to the corresponding fuel stages FSA or FSB via the annular lines FRA or FRB, the division of the fuel supply flow acts indiscriminately on all burners BN1 and BN2. However, due to geometric differences in the line guidance and changes in the burners BN1 and BN2 due to manufacturing tolerances or due to operating time, the fuel supply flow to the individual burners BN1 and BN2 and the individual combustion processes are at least slightly different. These differences usually also lead to different pollutant emissions, temperature peaks or pressure fluctuations in the individual burners BN1 and BN2, and cannot be compensated by the valves VA and VB located upstream of the annular lines FRA and FRB. Therefore, the controller of the internal combustion engine according to the prior art is usually configured based on the burner with the worst combustion value. However, as a result, the other burners do not operate in an optimal manner.
[0032] Figure 3 A diagram is shown for illustrating the fuel supply in the case of a gas turbine according to the invention. The gas turbine according to the invention has a plurality of burners, of which only two burners BN1 and BN2 are explicitly shown for the sake of clarity. The respective burner BN1 or BN2 can also have a plurality of firing positions and / or be designed as a burner group. Furthermore, the gas turbine according to the invention has a plurality of fuel stages, which are shown in FIG. Figure 3 Only two of the fuel levels, FSA and FSB, are explicitly indicated in FIG. As described above, the fuel levels FSA and FSB are used to divide the fuel supply flow to the respective burners BN1 and BN2. However, in combination with Figure 2 In contrast to the described gas turbine according to the prior art, in the gas turbine according to the invention this division takes place individually for each burner BN1 and BN2 or each burner group.
[0033] In order to divide the fuel supply flow to the individual fuel stages FSA and FSB, controllable burner-specific multi-way valves MV1 and MV2 are provided. Thus, a multi-way valve MV1 and MV2 is specifically assigned to each burner BN1 and BN2. The multi-way valves MV1 and MV2 are connected to the main supply line MSL of the gas turbine via a cross-stage annular line FR, via which the total fuel flow F is directed to the multi-way valves MV1 and MV2.
[0034] Preferably, a valve (not shown) across the burners is connected upstream of the multiway valves MV1 and MV2 in order to temporarily interrupt or reduce the fuel supply flow. Such interruption of the fuel supply flow is also commonly called tripping, and such reduction of the fuel supply flow is commonly called unloading.
[0035] By means of the burner-specific division of the fuel flow into the individual fuel stages FSA and FSB, burner-specific differences in the combustion process can be compensated and, moreover, can be optimized burner-specifically. In particular, pollutant emissions, local temperature peaks, pressure fluctuations, power, efficiency, fuel consumption, service life, wear, fatigue and / or clogging of the fuel nozzles can be optimized individually for each burner. By means of the stage-specific annular lines which may be omitted, the piping expenditure can also be reduced in many cases.
[0036] In addition, the individual burners BN1 or BN2 can be operated in a targeted manner at different power points, different combustion temperatures and / or different fuel quantities. As a result, the formation of carbon monoxide, especially at lower loads, and / or the formation of nitrogen oxides can generally be effectively reduced. As a result, in many cases the gas turbine can be operated at very low power and remain on standby while complying with emission regulations.
[0037] Figure 4 A schematic diagram shows a device for controlling a gas turbine GT according to the invention. Figure 4 Used in Figure 3 For the same reference numerals as in Figure 3 These entities can be Figure 3 Designed as described in .
[0038] The gas turbine GT has a turbine controller CTL which implements the control method according to the invention. Figure 4 In the embodiment, the turbine controller CTL is shown outside the gas turbine GT. Alternatively, the turbine controller CTL can also be implemented completely or partially in the gas turbine GT. The turbine controller has one or more computers or processors PROC for executing the method steps of the turbine controller CTL, and one or more memories MEN for storing data to be processed by the turbine controller CTL. In particular, a non-volatile memory is provided in the turbine controller CTL in order to control a computer program implementing the method according to the invention.
[0039] A gas turbine GT has multiple burners, for clarity, Figure 4 Only two of the burners BN1 and BN2 are shown in the figure. Here, the respective burner BN1 or BN2 can also have multiple firing positions and / or be designed as a burner group. Each burner BN1 and BN2 is fed in a burner-specific manner via an electric or hydraulic multi-way valve MV1 and MV2, respectively. The multi-way valves MV1 and MV2 divide the fuel supply flow to the burners BN1 and BN2 into individual fuel stages in a burner-specific or burner-group-specific manner, as already described in conjunction with Figure 3 described.
[0040] The multi-way valves MV1 and MV2 are each actuated by burner-specific control data (here, SD1 and SD2, respectively), by means of which the division of the fuel supply flow to the various combustion stages is set for each multi-way valve MV1 and MV2 in a burner-specific or burner group-specific manner. Furthermore, the mixture ratio between fuel and air, the injection pressure or other control parameters can also be set in a burner-specific or burner group-specific manner via the control data SD1 and SD2.
[0041] At the respective burners BN1 and BN2, burner-specific sensors AS1 and AS2 are arranged, respectively, which preferably measure burner-specific combustion measurement data VD1 and VD2 of the burners BN1 and BN2 concerned, respectively, in real time. In the present embodiment, in particular, the acoustic vibrations of the respective burners BN1 and BN2 generated by the pressure fluctuations in the respective combustion chambers are measured. Here, the intensity of the acoustic vibrations is detected separately for each burner BN1 and BN2 in a plurality of frequency ranges, preferably between 15 Hz and 10 kHz, and is output as combustion measurement data VD1 and VD2, respectively. In addition, burner-specific pollutant emissions (in particular pollutant emissions of nitrogen oxides, carbon monoxide or unburned hydrocarbons), chemical compositions of pollutants, combustion temperatures or combustion temperature distributions, inlet temperatures or exhaust gas temperatures, pressures, fuel-air mixture ratios, fuel supply flows or other combustion parameters can also be detected as combustion measurement data VD1 and VD2 by sensors AS1 and AS2.
[0042] The combustion measurement data VD1 and VD2 measured and output by the respective sensors AS1 and AS2 are respectively assigned to burner identifiers BK1 and BK2 identifying the burners BN1 and BN2 concerned, and transmitted to the turbine controller CTL.
[0043] Furthermore, the gas turbine GT has a performance sensor PS, which preferably measures performance measurement data PD of the gas turbine GT in real time. The performance sensor PS can include in whole or in part the sensors AS1 and AS2. The performance measurement data PD can be measured in a burner-specific manner or across burners and, if necessary, can be part of the combustion measurement data VD1 and VD2 measured by the sensors AS1 and AS2. The performance measurement data PD can include, in particular, measurement data relating to the power, efficiency, fuel consumption or wear of the gas turbine GT, to pressure fluctuations in the combustion chamber, to the emission of pollutants (in particular nitrogen oxides, carbon monoxide or unburned hydrocarbons), to the inlet temperature, to the exhaust gas temperature, to local temperature peaks or to the blockage of the fuel nozzles, as well as measurement data relating to other operating variables related to the performance of the gas turbine GT.
[0044] The combustion measurement data VD1 are transmitted to the machine learning model NN of the turbine controller CTL in a manner assigned to the burner identifier BK1 and the combustion measurement data VD2 are transmitted to the burner identifier BK2. The performance measurement data PD are likewise transmitted to the machine learning model NN by the performance sensor PS. The performance measurement data PD are used together with the combustion measurement data VD1 and VD2 assigned to the burner identifiers BK1 and BK2 as training data to train the machine learning model NN in a data-driven manner.
[0045] In the present embodiment, the machine learning model NN is implemented by an artificial neural network. The artificial neural network is preferably trained by means of a reinforcement learning method (Reinforcement-Learning). Here, the performance measurement data PD are evaluated in a reward function, which determines a performance value of the gas turbine GT based on the performance measurement data PD, which performance value will be optimized by the training. Such a reward function is also generally referred to as a Reward-Function. A large number of known reinforcement learning methods can be used to perform the training. Preferably, the so-called PGNRR method (PGNRR: Policy Gradient Neural Reward Regression), the so-called VOP-PGNRR method (VOP-PGNRR: Variable Target Policy PGNRR) or a model-based reinforcement learning method such as the so-called RCNN method (RCNN: Recursive Control Neural Network) can be used for the training.
[0046] In this embodiment, for each burner (here BN1 and BN2) of the gas turbine GT, the machine learning model NN includes burner-specific sub-models ML1 and ML2, respectively, which are trained with the aid of training data PD and VD1, BK1 and VD2, BK2 to generate specific control data, here SD1 and SD2, respectively, through the corresponding burners BN1 and BN2. The sub-models ML1 and ML2 are trained in parallel, and the trained sub-models ML1 and ML2 are executed in parallel in the production operation of the gas turbine GT. By executing the sub-models ML1 and ML2 in parallel, the application of the trained machine learning model NN can be greatly accelerated, so that high time resolution can be achieved when generating the control data SD1 and SD2.
[0047] In this context - following the traditional language usage in this field, training is to be understood as a mapping of the input data of the machine learning model (here NN) to one or more target variables, which is optimized during the training phase according to predetermined criteria. Here, a training structure of the machine learning model NN is formed that is optimized for predetermined criteria. This training structure can, for example, include the networking structure of the neurons of the neural network and / or the weights of the connections between neurons, which are constructed by training so that the predetermined criteria are met as well as possible. The predetermined criteria can be mapped in the reward function to be optimized.
[0048] In the present embodiment, the machine learning model NN receives as input data from the gas turbine GT burner-specific combustion measurement data VD1 and VD2, which are respectively assigned to the burner identifiers BK1 and BK2 in question. As further input data, the machine learning model NN receives performance measurement data PD. The output data of the machine learning model NN are fed back to the gas turbine GT as burner-specific control data SD1 and SD2. The multi-way valve MV1 is actuated as described above with the aid of the control data SD1, and the multi-way valve MV2 is actuated as described above with the aid of the control data SD2.
[0049] The machine learning model NN is trained so that its output data (here, the burner-specific control data SD1 and SD2) optimizes the performance value when the multi-way valves MV1 and MV2 are actuated in a burner-specific manner by the control data SD1 and SD2, thereby optimizing the performance measurement data PD to a certain extent. The performance measurement data PD are used as an optimization criterion for the machine learning model NN. Figure 4 The training is thus reduced to an optimization problem, for which a large number of optimization methods are available.
[0050] By optimizing the performance values, the machine learning model NN is trained to generate advantageous, i.e. performance-enhancing, control data SD1 and SD2 for the operation of the gas turbine GT from the combustion-specific combustion measurement data VD1 and VD2. With the aid of the control data SD1 and SD2 optimized in a burner-specific manner, the efficiency of the gas turbine GT can generally be increased and / or pollutant emissions or wear of the gas turbine GT can be reduced. In addition, local temperature peaks can be avoided or clogging of the fuel nozzles can be reduced.
[0051] The machine learning model NN is preferably further trained during ongoing production operation of the gas turbine GT based on currently detected combustion and performance measurement data. In particular, a closed control loop for continuous optimization can be provided in order to continuously adapt the gas turbine GT to changes caused by the environment or operating time.
[0052] According to an embodiment variant of the invention, the combustion measurement data VD1 and VD2 can also be detected in a fuel-grade-specific manner by means of the sensors AS1 and AS2, and a fuel-grade identifier identifying the respective fuel grade can be assigned to the combustion measurement data. The burner-specific and fuel-grade-specific combustion measurement data VD1 and VD2 can then be transmitted as input data to the machine learning model NN in a manner assigned to the respective fuel-grade identifier in order to train the machine learning model. The additional differentiation according to the fuel grade makes it possible to model the fuel-grade-specific interdependencies as well and to take them into account when optimizing the performance values.
[0053] According to another embodiment of the present invention, an additional machine learning model (not shown) can be set up, which is trained to generate additional control data, by which the fuel flow distribution between burners BN1 and BN2 is set and the performance value and / or other performance values are additionally optimized.
Claims
1. A method for controlling an internal combustion engine (GT) having a plurality of burners (BN1, BN2), wherein a) combustion measurement data (VD1, VD2) are detected in a burner-specific manner for the respective burner (BN1, BN2) and are assigned a burner identifier (BK1, BK2) which identifies the respective burner (BN1, BN2), b) detecting performance measurement data (PD) of said internal combustion engine (GT), determining a performance value based on said performance measurement data, c) training a machine learning model (NN) with the aid of the combustion measurement data (VD1, VD2), the assigned burner identifier (BK1, BK2) and the performance measurement data (PD) to generate burner-specific control data (SD1, SD2), which optimize the performance value when the burner (BN1, BN2) is operated in a burner-specific manner by means of the control data (SD1, SD2), wherein the machine learning model (NN) comprises a plurality of sub-models (ML1, ML2) specific to the respective burner (BN1, BN2), each sub-model generating control data (SD1, SD2) specific to the respective burner (BN1, BN2), and d) outputting control data (SD1, SD2) generated by a trained machine learning model (NN) for burner-specific control of the burner (BN1, BN2).
2. The method according to claim 1, characterized in that A fuel supply flow to the respective burner (BN1, BN2) is divided into a plurality of fuel stages (FSA, FSB) in a burner-specific manner according to the control data (SD1, SD2).
3. The method according to claim 2, characterized in that The fuel supply flow is divided into the fuel stages (FSA, FSB) by burner-specific multiway valves (MV1, MV2).
4. The method according to any one of claims 1 to 3, characterized in that To briefly interrupt or reduce the fuel supply flow, a valve across the burner is actuated.
5. The method according to any one of claims 1 to 3, characterized in that Combustion measurement data are detected in a fuel-grade-specific manner and respectively assigned to fuel-grade identifiers, and the assigned fuel-grade identifiers are used for training the machine learning model (NN).
6. The method according to any one of claims 1 to 3, characterized in that The machine learning model (NN) is trained by a reinforcement learning method.
7. The method according to any one of claims 1 to 3, characterized in that The machine learning model (NN) includes an artificial neural network, a recursive neural network, a convolutional neural network, a Bayesian neural network, an autoencoder, a deep learning architecture, a support vector machine, a data-driven trainable regression model, a k-nearest neighbor classifier, a physical model and / or a decision tree.
8. The method according to any one of claims 1 to 3, characterized in that The machine learning model (NN) is further trained during ongoing operation of the internal combustion engine (GT) based on currently detected combustion measurement data (VD1, VD2) and performance measurement data (PD).
9. The method according to any one of claims 1 to 3, characterized in that In a calibration phase, the machine learning model (NN) is pre-trained with the aid of a training internal combustion engine and / or with the aid of a simulation model of an internal combustion engine.
10. The method according to claim 1, characterized in that The sub-models (ML1, ML2) are executed in parallel to generate burner-specific control data (SD1, SD2).
11. The method according to any one of claims 1 to 3, characterized in that A further machine learning model is trained to generate further control data, by means of which the fuel flow distribution and / or the combustion temperature between the burners (BN1, BN2) is set and the performance value and / or further performance values are optimized.
12. A device for controlling an internal combustion engine (GT) having a plurality of burners (BN1, BN2), configured to carry out the method according to any one of claims 1 to 11.
13. A computer program product for controlling an internal combustion engine having a plurality of burners (BN1, BN2), configured to carry out the method as claimed in any one of claims 1 to 11.
14. A computer-readable storage medium having the computer program product according to claim 13.
