Method and apparatus for operating a fuel cell system by means of machine learning

Abstract

This invention relates to a method for determining system state variables (x) in a fuel cell system (1) at successive analysis moments. (i+1) A computer-implemented method for the fuel cell system (1) wherein the following steps are performed at each current analysis time (i+1): - Determine (S2) one or more operating state variables (y) of the fuel cell system (1) at the current analysis time (i+1). (i+1) ); - Using a trained recurrent neural network, based on the internal state vector (h) of the recurrent neural network at the previous analysis time (i) (i) And based on the running state variable (y) determined at the current analysis time (i+1) (i+1) To determine (S3) the current system state variable (x) (i+1) The internal state vector describes the internal state of the fuel cell system (1); and - based on the currently determined system state variable (x) (i+1) (S3) to run the fuel cell system (1).

Description

Technical Field

[0001] This invention relates to fuel cell systems, and more particularly to SOFC (Solid Oxide Fuel Cell) systems for operating fuel cell systems using natural gas. Specifically, this invention relates to determining the system state of a reforming unit for preparing hydrogen-rich gas from hydrocarbons contained in natural gas. Background Technology

[0002] Conventional fuel cell systems operating on natural gas have a reforming unit that converts the long-chain hydrocarbons in the natural gas into hydrogen-rich gas. This hydrogen-rich gas is then oxidized with oxygen in what is known as a fuel cell stack to generate electricity.

[0003] In addition to natural gas, exhaust gas from the cathode of the fuel cell stack is also recycled to a reforming unit used to convert hydrocarbons into natural gas. The exhaust gas recirculation rate can be variably adjusted here.

[0004] Through the exhaust gas recirculation, the fuel cell system exhibits strongly nonlinear behavior, determined by its internal system state. This internal system state is highly dependent on the operational changes of the fuel cell system. The internal system state of the fuel cell system is described by at least one system state variable, which may include the oxygen-to-carbon ratio at a specific point in the fuel cell system, particularly on the input side of the reforming unit, and / or fuel utilization rate. In particular, these system state variables form the basis of the operating type, and especially the basis for controlling or regulating the fuel cell system.

[0005] However, these system state variables cannot be measured online; instead, they must be determined through extensive measurements of gas composition. These measurements are time-consuming and cannot be used to estimate the internal system state during operation. Physical modeling of the oxygen-to-carbon ratio and fuel utilization rate is possible, but due to its complexity and long computation time, it is currently unsuitable for application during operation, and particularly unsuitable for regulation. Similarly, accurate real-time measurements are not possible.

[0006] In current fuel cell systems, the system state variables are determined in the form of oxygen-to-carbon ratio and fuel utilization rate to operate the fuel cell system. These system state variables are determined using an inaccurate physical model. Therefore, they can only inaccurately describe the current system state of the reforming unit. Summary of the Invention

[0007] According to the present invention, a method for operating a fuel cell system having a reforming device according to claim 1, and an apparatus and fuel cell system for operating a fuel cell system according to the parallel independent claims are provided.

[0008] Other designs are described in the dependent claims.

[0009] According to the first aspect, a method is provided for determining system state variables in a fuel cell system at successive analysis moments, wherein the following steps are performed at each current analysis moment:

[0010] - Determine one or more operating state variables of the fuel cell system at the current analysis time;

[0011] - Using a trained recurrent neural network, the current system state variables are determined based on the internal state vector of the recurrent neural network at a previous analysis time and based on the operating state variables determined at the current analysis time, wherein the internal state vector describes the internal state of the fuel cell system; and

[0012] - Operate the fuel cell system based on the currently determined system state variables.

[0013] To control a fuel cell system with exhaust gas recirculation, it is necessary to understand the internal system state in the form of system state variables, such as the oxygen-to-carbon ratio and / or fuel utilization rate. However, these variables can only be determined based on measurements of the gas composition delivered by the fuel cell. Such measurements are costly and cannot be performed during fuel cell operation. Physical modeling of the system state variables, namely the oxygen-to-carbon ratio and fuel utilization rate, is possible, but such modeling is too inaccurate for reliable regulation of the fuel cell system. Precise measurements are possible, but expensive. In particular, understanding these system state variables is essential for controlling or regulating the fuel cell system.

[0014] One approach to the above method involves using a recurrent neural network to determine the system state variables, namely the oxygen-to-carbon ratio or fuel utilization rate. For this purpose, the operating state variables should be measured variables present in the fuel cell system, such as temperature, volumetric mass flow, the gas composition of the delivered natural gas, and valve regulation.

[0015] The system state variables can be modeled using a known physical model. However, due to the tolerance of the operating state variables and the inaccuracy of the physical model, the estimation of the system state variables will be biased. These biases accumulate and may lead to significant inaccuracies in estimating the internal system state of the fuel cell system. Therefore, based on the above method, a recurrent neural network is proposed to make the physical model more accurate.

[0016] The combination with the physical model allows for the training of a recurrent neural network with a relatively small training dataset. This is particularly advantageous because determining the system state of a fuel cell system is very costly, and only a small number of reference measurements are available for training the recurrent neural network.

[0017] The use of the physical model allows knowledge of known physical relationships to be applied to determine the system state of the fuel cell system. The system state variables determined in this way can then be made more accurate through correction interventions using a trained recurrent neural network. The hybrid model formed in this way enables good predictions of the system state of the fuel cell system even when the recurrent neural network is trained using a small training dataset. In summary, the above method allows for the provision of system state based on system state variables, which would otherwise only result in a very inaccurate physical model of the system state, thus enabling improved control or regulation of the fuel cell system based on these system state variables.

[0018] Furthermore, the current system state variables can still be determined based on the changes in the running state variables at multiple previous analysis times.

[0019] It can be specified that the operating state variables include or depend on at least one of the following variables: the generated current, the temperature of the fuel cell unit of the fuel cell system, one or more volumetric flows and the fuel composition at the input of the fuel cell system, and one or more regulation amounts, particularly valve regulation amounts for controlling flow.

[0020] According to one implementation, a trained recurrent neural network can be configured to determine a system state variable correction based on the internal state vector of the recurrent neural network at a previous analysis time and the running state variable at the current analysis time, and to apply the system state variable correction to the system state variable modeled by means of a physical model using the running state variable at the current analysis time to determine the current system state variable.

[0021] In particular, the trained recurrent neural network can take into account the system state variables modeled by the physical model at the current analysis moment to determine the correction (K) of the system state variables.

[0022] Furthermore, trained recurrent neural networks can have GRU units or LSTM units.

[0023] According to one embodiment, the fuel cell system can operate according to the current system state variables, wherein in particular, one or more of the following variables, as adjustment quantities, are variably adjusted according to the current system state variables: recirculation mass flow, fuel mass flow, and oxygen mass flow.

[0024] According to another aspect, a device for operating a fuel cell system, particularly a control unit, is provided, wherein current system state variables in the fuel cell system are determined, and the device is configured to periodically perform the following steps:

[0025] - Determine one or more operating state variables of the fuel cell system at the current analysis time;

[0026] - Using a trained recurrent neural network, the current system state variables are determined based on the internal state vector of the recurrent neural network and the operating state variables at the current analysis time, wherein the internal state vector describes the internal state of the fuel cell system at a previous analysis time; and

[0027] - Operate the fuel cell system based on the currently determined system state variables.

[0028] According to another aspect, a fuel cell system is provided, including a fuel cell; a reforming unit; an exhaust gas recirculation unit; a sensing device for measuring a measurement variable as an operating state variable; and the aforementioned equipment. Attached Figure Description

[0029] The embodiments are explained in more detail below with reference to the accompanying drawings. Wherein:

[0030] Figure 1 A schematic diagram of a fuel cell system is shown;

[0031] Figure 2 A schematic functional diagram is shown for determining the system state of a fuel cell system;

[0032] Figure 3 An exemplary illustration of a GRU unit as an RNN unit is shown; and

[0033] Figure 4 A flowchart illustrating a method for determining system state variables of a fuel cell system is shown. Detailed Implementation

[0034] Figure 1A schematic diagram of fuel cell system 1 is shown. Fuel cell system 1 has fuel cell unit 2. Fuel cell unit 2 has fuel cell 3 for providing electrical energy. Fuel cell 3 can correspond to a solid oxide fuel cell (SOFC). Other battery technologies can also be used accordingly, such as PLFC (polymer electrolyte fuel cell), MCFC (molten carbonate fuel cell), and PAFC (phosphoric acid fuel cell).

[0035] The fuel cell 3 has an anode 5 and a cathode 6. An electrode 7 is arranged between the anode 5 and the cathode 6, wherein the anode 5 is separated from the cathode 6 by an electrolyte.

[0036] An oxidant, particularly air or oxygen, can be supplied to the cathode 6 of the fuel cell unit 2 via supply line 8. The oxidant is introduced via valve 9 and compressed by means of oxidation compressor 10. Oxidation compressor 10 can be variably adjusted to a predetermined oxygen mass flow. Oxidation compressor 10 can be configured as a continuously operating compressor with a variable-adjustment throttle valve to variably adjust the oxygen mass flow as one of the adjustment quantities.

[0037] In order to provide electricity, synthetic gas containing fuel, especially synthetic gas containing hydrogen, is delivered to the anode 5 of the fuel cell unit 2 via fuel supply line 11.

[0038] To generate fuel-containing synthesis gas, a reforming unit 12 is connected upstream of fuel cell unit 2. The reforming unit 12 is connected to fuel cell unit 2, or particularly to the anode 5 of fuel cell 3, via a fuel supply line.

[0039] To produce syngas containing fuel, a mixture of natural gas or other reactants, particularly oxygen and / or water vapor, is fed to reforming unit 12, where it is reformed into syngas containing fuel. Natural gas is compressed by fuel compressor 13 and delivered to reforming unit 12 via fuel valve 14. Fuel compressor 13 can be variably operated to pre-determine the fuel mass flow rate. The fuel compressor 13 can be configured as a continuously operating compressor with a variable-adjustment throttle valve to variably adjust the fuel mass flow as one of the adjustment quantities.

[0040] Downstream of the fuel compressor 13, additional reactant water, particularly in the form of steam, is added to the natural gas via exhaust gas recirculation 15.

[0041] The exhaust gas recirculation 15 of the fuel cell system 1 allows at least a portion of the exhaust gas from the fuel cell 3 (anode exhaust gas in the illustrated case) to be recirculated to the input side of the reformer 12 for mixing with natural gas and the additional reactants. For this purpose, the exhaust gas recirculation 15 may have a recirculation compressor 16 configured to recirculate the water-containing anode exhaust gas to the reformer 12. The recirculation compressor 16 can be variably operated to pre-determine the recirculation mass flow rate. The recirculation compressor 16 can be configured as a continuously operating compressor with a variable-adjustment throttle valve to variably adjust the recirculation mass flow as one of the adjustment quantities.

[0042] During reforming, long-chain alkanes are completely reformed, while methane is partially reformed in reforming unit 12 and partially reformed in fuel cell 3. Furthermore, unused syngas can be recycled back to the fuel cell via exhaust gas recirculation 15 through reforming unit 12, which improves the fuel efficiency of fuel cell system 1. Various alkanes from natural gas are introduced into reforming unit 12 along with water vapor and reformed there. Depending on the composition of the natural gas, different amounts of water vapor are required to optimize reforming. Similarly, the concentration in the syngas depends on the composition of the natural gas used.

[0043] The internal state of fuel cell system 1 is related to the operation of fuel cell 3. The state of fuel cell system 1 is reflected in system state variables, particularly the oxygen-to-carbon ratio or fuel utilization rate of fuel cell system 1. Since the oxygen-to-carbon ratio or fuel utilization rate cannot be directly measured or can only be measured through considerable expenditure, understanding the internal state of fuel cell system 1 can help the fuel cell system operate more efficiently. Thus, for example, system state variables can be determined based on the internal state of fuel cell system 1 and used to adjust the operation of fuel cell system 1 in a known manner.

[0044] In order to determine the system state variables that serve as characteristic variables of the system state of fuel cell system 1, it is necessary to periodically detect the measured variables that serve as operating state variables, which describe the current operating state of fuel cell system 1.

[0045] Thus, the operating state of the fuel cell system can be detected, for example, using temperature sensors, volumetric flow sensors, and gas composition sensors, as a description of temperature, one or more volumetric flow parameters, and fuel gas composition. These operating states allow the system state variables to be determined from the internal state. Furthermore, the operating state variables may include one or more regulation parameters provided for manipulating the fuel cell system, particularly valve regulation parameters for controlling flow rates, etc.

[0046] Although the system state determined by the system state variables can be determined by a physical model based on the operating state variables, the measured operating state variables and / or the physical model are erroneous, which may lead to the accumulation of errors through modeling. These errors make it more difficult to predict the system state of fuel cell system 1.

[0047] exist Figure 2 The diagram illustrates the methods used to determine the system state variable x. (i) In particular, the system state variable x in the form of oxygen-carbon ratio or fuel utilization rate (i) System state variable x in the form of (i) The functional circuit diagram of system state model 20 is shown. This functional circuit diagram illustrates a hybrid model that can be used to determine the internal system state of fuel cell system 1.

[0048] The system state model 20 has recursive characteristics and has model blocks 21 that are executed at successive analysis times i-1, i, i+1.

[0049] At each analysis time i, model block 21 uses the measured variables and internal state vector h detected at the current analysis time. (i) Obtain the current running state variable vector y (i+1) As input.

[0050] Alternatively, multiple internal state vectors h determined at n previous analysis times can be considered. (i-n...i-1) .

[0051] Based on these inputs, the function of model block 21 is executed at each analysis time i in order to determine the current system state variable x for the subsequent analysis time i+1. (i+1) And update the internal state vector h (i+1) .

[0052] In detail, system state model 20 includes a building module 22 that applies a physical model to determine the physically modeled system state variables x'. The physically modeled system state variables x' are determined using a physical model that maps physical and / or chemical dependencies. This physical model is based on the currently running variable vector y. (i+1) Determine the system state variable x' for physical modeling. The current operating variable vector is, for example, the current temperature, volumetric mass flow, gas composition, and at least one adjustment amount and the average volumetric mass flow entering the reformer 12 from the last analysis time to the current analysis time.

[0053] The system state variable x', determined by the physical modeling module, is fed to the summing block 23. The current system state variable x(i) corresponds to the absolute description of the system state of fuel cell system 1.

[0054] In particular, when x' is the oxygen-to-carbon ratio φ, the oxygen-to-carbon ratio φ can be determined according to the following physical model known from the prior art:

[0055]

[0056] Where K1 and K2 represent the constants to be parameterized. Explain the recirculated mass flow, and Explain the fuel mass flow.

[0057] Alternatively, the fuel utilization rate FU can be determined based on the following physical models known from the prior art:

[0058]

[0059] K3 and K4 correspond to the constants to be parameterized.

[0060] The system state variable x' in physical modeling is like the current running state variable vector y. (i+1) and the final determined internal state vector h (i) That way, all of them are fed into RNN unit 24 (RNN: recurrent neural network), where the current internal state vector h is determined. (i+1) RNN unit 24 corresponds to a type of neural network that is particularly well-suited for processing time signal sequences and for predicting time-varying processes. RNN unit 24 can be constructed, for example, as a GRU unit (Gated Recurrent Unit) or an LSTM unit (Long-Short-term Memory) or a variant thereof.

[0061] The current internal state vector determined by RNN unit 24 The system is processed in optional correction block 25 to provide a system state variable correction K, and in summation block 23, this system state variable correction is added to the system state variable x' determined at the previous analysis time i-1 in the physical model. For the physical model described above, the system state variable correction K corresponds to the correction amount for the oxygen-carbon ratio φ or the fuel utilization rate FU.

[0062] The correction block 25 can be a predefined function that modifies the internal state vector h. (i) The correction K is assigned to the system state variable. The correction block 25 can be specifically constructed as a neural network, preferably as a separate feedforward neural network, which will take the current internal state vector h... (i+1) The correction K is assigned to the system state variable. In particular, when the LSTM is used as RNN unit 24, the correction block 25 can correspond to the "output gate" of the LSTM.

[0063] Alternatively, RNN unit 24 can also be constructed as, in addition to the current internal state vector h (i+1) In addition, it provides system state variable correction K, which allows the setting of correction block 25 to be omitted.

[0064] After the calculation in function block 21, the current system state variable x is provided at the output of summation block 23. (i+1) Furthermore, an internal state vector h is provided at the output of RNN unit 24. (i+1) For analysis at time i+1. In the calculation of the next analysis time i+2, they are used in conjunction with the measured variable y at the next analysis time i+2. (i+2) They are assumed to be running state variables together.

[0065] The GRU unit, used as an example of RNN unit 24, is trained using known training methods, such as backpropagation. For this purpose, the system state variable x is measured at k pre-given analysis times Tk to T and at analysis time T. (T) At analysis time Tk, the system state variable x is assumed or measured. (T-k) The running state variable vector y is set at analysis time Tk...T. (T-k...T) As training variables, and execute the corresponding recursive training method in a known manner to train the trainable parameters of the RNN.

[0066] Figure 3 An exemplary illustration of a GRU unit, exemplified as RNN unit 24, is shown. The following functions are performed in the GRU unit:

[0067]

[0068]

[0069] Where h t Corresponding to the internal state vector, y t Corresponding to the running state variable vector, W, U, and b correspond to the parameter matrix or parameter vector of the GRU unit, σ g This corresponds to the sigmoid function.

[0070] For the system state variable x determined at the current analysis time i+1 (i+1)The analysis is used to operate the fuel cell system. In particular, the limits of the adjustment amount are defined based on the system state variables, as described, for example, in M. Carré et al., “Feed-forward control of a solid oxide fuel cell system with anode offgas recycle”, Journal of PowerSources 282 (2015), pp. 498-510.

[0071] exist Figure 4 The document describes a flowchart illustrating the method performed using control device 10.

[0072] In step S1, the operating state variables are measured at a short sampling rate, for example, <1s, preferably less than 0.5s, and provided as an operating state variable vector together with the adjustment amount if necessary.

[0073] In step S2, the current system state variable at the current analysis time i+1 is determined by means of the hybrid model from the running state variable and the finally determined internal state vector.

[0074] In step S3, based on the determined current system state variable x (i) To operate the fuel cell system. For example, M. Carré et al. describe a possible operating mode in “Feed-forward control of a solid oxide fuel cell system with anode offgas recycle”, Journal of Power Sources 282 (2015), pp. 498-510. Here, the oxygen-to-carbon ratio ϕ or the fuel utilization rate FU, determined by the model, is used as the limit value that must not be exceeded or fallen below when selecting the operating point of the fuel cell system.

Claims

1. Computer-implemented method for determining a system state variable (x (i+1) ) in a fuel cell system (1) at successive analysis instants, wherein the following steps are performed at each current analysis instant (i+1): - determining (S2) one or more operating state variables (y (i+1) ) of the fuel cell system (1) at a current analysis instant (i+1); - Using a trained recurrent neural network, based on the internal state vector (h) of the recurrent neural network at the previous analysis time (i) (i) And based on the running state variable (y) determined at the current analysis time (i+1) (i+1) To determine (S3) the current system state variable (x) (i+1) The internal state vector describes the internal state of the fuel cell system (1). The trained recurrent neural network is constructed based on the internal state vector (h) at the previous analysis time (i). (i) ) and the running state variable (y) at the current analysis time (i+1). (i+1) ) to determine the system state variable correction (K), and to use the operating state variable (y) (i+1) At the current analysis time (i+1), the system state variable correction is applied to the system state variable (x') modeled using the physical model to determine the current system state variable (x). (i+1) );as well as -Based on the currently determined system state variables (x) (i+1) (S3) to operate the fuel cell system (1).

2. The method according to claim 1, wherein, The current system state variable (x) is also determined (S3) based on the changes of the operating state variable at multiple previous analysis times. (i+1) ).

3. The method according to claim 1 or 2, wherein, The operating state variables include or depend on at least one of the following variables: the generated output current, the temperature of the fuel cell unit (2) of the fuel cell system (1), one or more volumetric flows and the fuel composition at the input of the fuel cell system (1), and one or more regulation amounts, wherein the regulation amounts include valve regulation amounts for controlling the flow rate.

4. The method according to claim 1 or 2, wherein, The trained recurrent neural network considers the system state variables (x') modeled by the physical model at the current analysis time (i+1) to determine the correction (K) of the system state variables.

5. The method according to claim 1 or 2, wherein, The trained recurrent neural network has GRU units (24) or LSTM units.

6. The method according to claim 1 or 2, wherein, The fuel cell system (1) operates according to the system state variables by limiting the adjustment amount to the current system state variable (x). (i+1) The value is determined.

7. A device for operating a fuel cell system (1), wherein the current system state variable (x) in the fuel cell system (1) is determined. (i+1) The device is configured to periodically perform the following steps: - Determine (S2) one or more operating state variables (y) of the fuel cell system (1) at the current analysis time (i+1). (i+1) ); - Using a trained recurrent neural network, based on the internal state vector (h) of the recurrent neural network at the previous analysis time (i) (i) And based on the running state variable (y) determined at the current analysis time (i+1) (i+1) To determine (S3) the current system state variable (x) (i+1) The internal state vector describes the internal state of the fuel cell system (1), wherein the trained recurrent neural network is constructed based on the internal state vector (h) at the previous analysis time (i). (i) ) and the running state variable (y) at the current analysis time (i+1). (i+1) ) to determine the system state variable correction (K), and to use the operating state variable (y) (i+1) At the current analysis time (i+1), the system state variable correction is applied to the system state variable (x') modeled using the physical model to determine the current system state variable (x). (i+1) );as well as -Based on the currently determined system state variables (x) (i+1) (S3) to operate the fuel cell system (1).

8. The device according to claim 7, wherein, The device is a control unit.

9. A fuel cell system (1), comprising: - Fuel cell (3); - Reforming unit (12); - Exhaust gas recirculation (15); - A sensing device used to measure variables that are part of the operating state variables; as well as The device according to claim 7 or 8.

10. A computer program product comprising a computer program configured to perform all the steps of the method according to any one of claims 1 to 6.

11. An electronic storage medium having a computer program stored thereon, the computer program being configured to perform all the steps of the method according to any one of claims 1 to 6.

Images