Method for determining a module state of an energy storage module, for generating training data sets and for energy management of a vehicle

Abstract

The present invention relates to a method for determining a module state of an energy storage module of a multilevel converter system, which comprises a plurality of energy storage modules and switching devices. The multilevel converter system is designed to connect and / or bridge each energy storage module to the respectively adjacent energy storage module in parallel and / or in series. Each energy storage module has at least one energy storage cell. In the method, charge carriers are exchanged with the energy storage modules in a pulsed manner in a charge carrier exchange process, specifically according to at least one pulse feature. During the charge carrier exchange process, at least one pulse response voltage and / or a pulse response temperature and / or a pulse response current intensity of at least one energy storage module is measured. A voltage parameter and / or a temperature parameter and / or a current intensity parameter are then provided as input parameters to an analysis model, and the analysis model outputs the module state of the energy storage module as an output parameter. The invention further relates to a method for generating training data sets for supervised learning, and to a method for energy management of a vehicle which comprises a multilevel converter system having a plurality of energy storage modules and switching devices.

Description

[0001] December 11, 2025

[0002] Pulsetrain GmbH M / PUIN-022-PC TR / TP / ra

[0003] Methods for determining the module state of an energy storage module, for generating training data sets, and for energy management of a vehicle.

[0004] Description

[0005] The present invention relates to a method for determining the module state of an energy storage module of a multilevel converter system comprising a plurality of energy storage modules and switching devices. The multilevel converter system is configured to connect each energy storage module in parallel and / or in series and / or by bridging it to the adjacent energy storage module. Each energy storage module has at least one energy storage cell. Furthermore, the invention relates to a method for generating training datasets for supervised learning, in particular for training an artificial neural network for use in the method for determining the module state of the energy storage module. Finally, the invention relates to a method for energy management of a vehicle comprising a multilevel converter system with a plurality of energy storage modules and switching devices.

[0006] The energy storage cell of the energy storage module includes at least one battery, for example, an accumulator. The energy storage module includes, in particular, lithium-ion cells, such as those used in electric vehicles. The module state of such an energy storage module includes, in particular, the state of charge (SoC), the state of health (SoH), and the state of power (SoP). These module states of the energy storage module cannot be measured directly; rather, they must be estimated using various methods. These methods include data-driven and model-based methods. M / PUIN-022-PC

[0007] 2

[0008] A disadvantage of known methods for determining such a module state of an energy storage module is that either the accuracy decreases over time due to cumulative errors or the methods are not suitable for continuous real-time applications.

[0009] From DE 10 2022 110 426 A1, a method for characterizing and / or optimizing at least one energy storage module is known. In this method, the energy storage modules are connected in such a way that at least one energy storage module is characterized and / or optimized based on its frequency response.

[0010] The present invention aims to provide a method for determining the state of an energy storage module of the type mentioned above, enabling the module state to be determined accurately and in real time during the use of the energy storage module in an application intended for the energy storage module. Furthermore, a method for generating training datasets for supervised learning is to be provided, suitable for training an artificial neural network for use in the method for determining the module state of the energy storage module. Finally, a method for energy management of a vehicle is to be provided, which efficiently manages the available stored energy.

[0011] According to the invention, this problem is solved by a method with the features of claim 1, a method with the features of claim 8, and a method with the features of claim 15. Advantageous embodiments and further developments are described in the dependent claims.

[0012] In the inventive method for determining a module state of an energy storage module of a multilevel converter system, which comprises a plurality of energy storage modules and switching devices, wherein the multilevel converter system is configured to connect and / or bridge each energy storage module in parallel and / or in series with the respective adjacent energy storage module, and wherein each energy storage module has at least one energy storage cell, in a charge carrier exchange process M / PUIN-022-PC

[0013] 3. Charge carriers are exchanged in pulsed fashion with the energy storage modules, according to at least one pulse characteristic. During the charge carrier exchange process, at least one pulse response voltage and / or pulse response temperature and / or pulse response current of at least one energy storage module is measured. A voltage parameter, which is the measured pulse response voltage or is derived from the measured pulse response voltage, and / or a temperature parameter, which is the measured pulse response temperature or is derived from the measured pulse response temperature, and / or a current parameter, which is the measured pulse response current or is derived from the measured pulse response current, are provided as input parameters to an analysis model. The analysis model then outputs the module state of the energy storage module as an output parameter.

[0014] In the method according to the invention, an exchange of charge carriers with the energy storage modules is carried out in a charge carrier exchange process. This exchange occurs with at least one pulse characteristic. During this charge carrier exchange process, the pulse response voltage and / or the pulse response temperature and / or the pulse response current of at least one energy storage module, and in particular of all energy storage modules, is measured. Surprisingly, it has been found that the measured pulse response voltage and / or the measured pulse response temperature and / or the measured pulse response current is particularly suitable as an input parameter for an analysis model that outputs the module state of the energy storage module as an output parameter. Advantageously, this allows the module state of the energy storage module to be determined very accurately, since the analysis model can represent the energy storage module very precisely.Furthermore, the procedure can be executed in real time during the operation of the energy storage module in its intended use to determine the module's state, as the analysis model can be generated before the energy storage module is deployed. The duration and amount of data required to generate the analysis model are not critical, since the analysis model is created before the actual operation of the energy storage module. By the time the energy storage module is commissioned, the analysis model has already been generated, so it is M / PUIN-022-PC.

[0015] 4 can then be used in real time in the method according to the invention to determine the module state of the energy storage module very quickly.

[0016] The use of the voltage parameter in the method according to the invention is advantageous because the voltage level is characteristic of the current capacity of the energy storage module. In this way, conclusions can be drawn about how much energy is available, i.e., the state of charge, and about the general condition of the energy storage module, i.e., in particular its aging state. By measuring the pulse response voltage during a charge carrier exchange process with a pulse characteristic, especially a known current pulse, the internal resistance of the energy storage cell, particularly a battery cell, can also be advantageously determined. This internal resistance is an indicator of various processes, such as charge transfer, diffusion, and DC resistance, each of which provides information about the state of charge, the performance state, and the aging state, respectively.provide the lifespan of the energy storage module.

[0017] The alternative or additional use of the temperature parameter is advantageous because the temperature has a significant influence on the internal resistance of the energy storage cell and the overall efficiency.

[0018] The alternative or additional use of the current parameter is advantageous because the current provided by the multilevel converter system has a significant influence on the state of the multilevel converter system.

[0019] In the intended use of a multilevel converter system with an energy storage module, charge carrier exchange typically occurs in pulses. This property is advantageously exploited in the method according to the invention to determine the module state, since the analysis model determines the module state of the energy storage module based on a measured pulse response voltage and / or measured pulse response temperature and / or measured pulse response current. The pulse characteristics of the charge carrier exchange process are thus present during the operation of the multilevel converter system, so that it is not necessary to M / PUIN-022-PC

[0020] Five additional pulse characteristics for determining the module state of the energy storage module can be generated separately. This allows the method according to the invention to be advantageously carried out more efficiently when using a multilevel converter system as intended.

[0021] According to one embodiment of the procedure, the analysis model includes an artificial neural network, in particular a feed-forward neural network (FNN), a recurrent neural network (RNN), a support vector machine (SVM) or a long-short-term memory.

[0022] Feed-forward neural networks (FNNs) advantageously possess a simple architecture and high computational efficiency. They are therefore particularly well-suited for real-time applications such as monitoring the state of charge or the aging state of an energy storage module. They provide very fast predictions of the module state of an energy storage module in a multilevel converter system, which is beneficial for real-time battery management. Furthermore, feed-forward neural networks are less computationally intensive and faster to train compared to more complex networks such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs). They are especially effective at processing static data that does not require temporal sequencing.Feed-forward neural networks are therefore particularly well suited for predicting the state of charge and the state of aging, as they can learn nonlinear relationships between pulse response data of voltage and temperature on the one hand and module states, especially battery states, on the other, and are easy to implement and interpret.

[0023] A long-short-term memory (LSTM), on the other hand, is more complex, but it can better capture time-dependent patterns in the data. It is particularly useful for determining the module state of an energy storage module, especially battery performance, over many charge / discharge cycles, as it can provide long-term memory that helps model the gradual aging of the energy storage module over time. Unlike feed-forward neural networks, a long-short-term memory can process data sequences and detect temporal dependencies in the performance of the M / PUIN-022-PC.

[0024] 6

[0025] The energy storage module can be better taken into account. Therefore, the use of long-short-term memory is advantageous when historical data plays a crucial role in prediction.

[0026] According to a further embodiment of the method, the analysis model has a calculation rule which maps an arbitrary pulse response voltage and / or pulse response temperature and / or pulse response current to a module state.

[0027] The computational formula for the analysis model can be used as an alternative to, or in combination with, the artificial neural network. Advantageously, the artificial neural network is used as a machine learning model in combination with a conventional model estimation to determine the state of charge and aging of the energy storage module. Combining a computational formula with an artificial neural network can advantageously improve the accuracy and robustness of the determination of the energy storage module's state of charge.

[0028] According to a further embodiment of the method, the at least one pulse characteristic comprises one or more of the following characteristics of the charge carrier exchange process: the pulse duration, in particular a pulse duration of 200ms, 2s or 20s; the pulse frequency; the pulse amplitude; duty cycle.

[0029] The duty cycle (or duty level) is the ratio of pulse duration to period. The idle time after pulses is also a parameter that can be considered in the pulse response.

[0030] In the pulsed charge carrier exchange process with the energy storage modules, a current pulse is applied to an energy storage module, and the resulting pulse response voltage, temperature, and / or current intensity are measured. The pulse can be unipolar or bipolar. Furthermore, a pulse can be composed of multiple individual pulses. M / PUIN-022-PC

[0031] 7

[0032] If the pulse, also referred to as the main pulse in this case, is composed of a multitude of individual pulses, it is a defined sequence of at least two current flow phases (unipolar or bipolar pulses) and at least one, and in particular several, rest phases (pulse-off), which is maintained for a specific total duration. During a main pulse, the parameters of the individual pulses (e.g., amplitude, pulse width, duration, duty cycle) do not have to remain constant and can vary. A main pulse can be composed of several individual pulses if the selected main pulse length is longer than the longest pulse duration generated in the multi-level converter system. The data from the main pulse are used to determine the cell state. Individual pulses are not considered separately. The boundary between two main pulses is defined by a minimum rest period during which no current flows.

[0033] A single pulse is the smallest unit within a main pulse, consisting of a pulse-on phase and a pulse-off phase. The parameters of a single pulse (e.g., duration, pulse width, amplitude, duty cycle) can be variable.

[0034] If the main pulse is composed of a plurality of individual pulses, the pulse is defined as the current characteristic over a time interval that begins at the first current flow phase of an individual pulse and ends at the end of an nth current flow phase, i.e., at the end of a subsequent individual pulse, with a minimum rest period following the nth current flow phase, which exceeds a defined value. This value is, for example, in the range of 50 ms to 150 ms, particularly in the range of 80 ms to 120 ms, and preferably the minimum rest period is 100 ms.

[0035] These pulse characteristics are also commonly used in the charge carrier exchange process during the intended use of energy storage modules. In the inventive method, they are used in the same way to obtain the voltage parameters and / or the temperature parameters and / or the current parameters from which the analysis model then determines the module state of the energy storage module. Advantageously, the method can thus determine the module state in real time during operation of the energy storage module without requiring M / PUIN-022-PC

[0036] 8 is required to generate separate pulse characteristics for the analysis of the module state.

[0037] According to a further embodiment of the method, the difference between the measured pulse response voltage and an average of pulse response voltages is provided to the analysis model as a voltage parameter. Alternatively or additionally, the difference between the measured pulse response temperature and an average of pulse response temperatures is provided to the analysis model as a temperature parameter. Alternatively or additionally, the difference between the measured pulse response current and an average of pulse response currents is provided to the analysis model as a current parameter.

[0038] Choosing such a voltage parameter, temperature parameter, and current parameter is advantageous if a corresponding voltage parameter, temperature parameter, and / or current parameter was also used when generating the analysis model. By selecting such a voltage parameter, temperature parameter, and / or current parameter when generating the analysis model, fluctuations in these parameters can be detected more effectively.

[0039] A multilevel converter system, preferably modular, describes a type of arrangement or circuit of several energy storage modules and switching devices, in particular transistors. The energy storage module can be a storage device for an electrical source, preferably frequency-dependent, such as a battery cell or accumulator. Each energy storage module can have at least or exactly one battery cell or accumulator. The transistors serve, for example, as switches by which current and / or voltage paths can be selected. The energy storage modules can thus be integrated into or excluded from a desired configuration.

[0040] The energy storage cell is in particular a battery cell, and the module state is in particular a cell state of the battery storage cell, e.g. the M / PUIN-022-PC.

[0041] 9

[0042] State of charge (SoC) and / or state of health (SoH) and / or state of performance (SoP) of the battery cell. The module state can also refer to the state of an individual battery storage cell.

[0043] The state of charge is defined as the energy stored in the energy storage modules or battery storage cells; the state of performance is defined as the peak power that the energy storage modules or battery storage cells can deliver or receive within a specific time interval.

[0044] The load carrier exchange process is, in particular, a loading and / or unloading process.

[0045] Each energy storage module can be connected in parallel and / or in series and / or bridged with its adjacent energy storage module. Preferably, each energy storage module can be connected in series with its adjacent energy storage module. The possibility of parallel connection is advantageous, but not necessary.

[0046] Preferably, the adjacent energy storage modules are interconnected via two current and / or voltage paths. Each path can be assigned a switching device, in particular a transistor.

[0047] For example, three transistors are provided between two adjacent energy storage modules. This allows the energy storage modules to be connected in parallel or in series, for example. Furthermore, switching devices, in particular transistors, are preferably provided for bridging energy storage modules.

[0048] The procedure for generating training datasets for supervised learning comprises the following steps: a) performing a pulsed charge carrier exchange at a training energy storage module of a training multilevel converter system, wherein the charge carrier exchange is performed with at least one training pulse feature; b) measuring a training pulse response voltage and / or a training pulse response temperature and / or a training pulse response current; M / PUIN-022-PC

[0049] 10 c) Determining a training module state of the training energy storage module after performing the charge carrier exchange; d) Generating a training data set comprising the following:

[0050] - a training voltage parameter, which is the training pulse response voltage or is derived from the training pulse response voltage, and / or a training temperature parameter, which is the training pulse response temperature or is derived from the training pulse response temperature, and / or a training current parameter, which is the training pulse response current or is derived from the training pulse response current; and

[0051] - the specific training module state.

[0052] This method for generating training datasets advantageously allows for the training of an artificial neural network, which can be used in the aforementioned method for determining the module state of an energy storage module. The training energy storage module used for this purpose corresponds in particular to the energy storage module described above, the training multilevel converter system used corresponds in particular to the multilevel converter system described above, and the training pulse feature used in the charge carrier exchange corresponds in particular to the pulse feature described above, possibly with variations thereof.

[0053] Advantageously, the data of the training data set establishes a relationship between the training voltage parameter and / or the training temperature parameter and / or the training current parameter on the one hand and the training module state of the training energy storage module on the other.

[0054] The training pulse response voltage and / or temperature and / or current intensity is a response to both the time interval during the pulse (on-time) and the time interval outside the pulse (off-time). The duration of these two time intervals is determined by the duty cycle. M / PUIN-022-PC

[0055] 11

[0056] When generating the training datasets, data is produced that can serve as input for training a neural network. In this case, the relationship between pulse response voltage and / or pulse response temperature and / or pulse response current intensity to a pulsed charge carrier exchange, especially pulsed currents, is the primary input for training the neural network.

[0057] According to one embodiment of the procedure, in step c) the training module state of the training energy storage module is determined by means of charge carrier counting}, open circuit voltage method}, model-based methods, capacitance testing and / or measurement of the internal resistance.

[0058] To generate the training dataset, the training module state is determined very accurately using conventional methods. Here, the estimated module states are compared with the actual states determined using conventional methods. Determining the training module state in real time during the operation of an energy storage module would not be practical with the method according to the invention, as the computation time for real-time determination would be too long and a very high computing capacity would also be required. However, these methods can be used to generate the training dataset and obtain a target module state for training an artificial neural network.

[0059] Charge carrier counting measures the charge entering and exiting the energy storage module to determine its state of charge. The open-circuit voltage method utilizes the relationship between open-circuit voltage and the energy storage module's state of charge. Model-based methods, particularly electrochemical or physical equivalent circuit models, can be generated. To account for the aging of the energy storage module, parameters can be determined using electrochemical impedance spectroscopy (EIS) measurements or pulse injection techniques. In electrochemical impedance spectroscopy, the impedance of the energy storage module, especially the battery cell, is analyzed at various frequencies. M / PUIN-022-PC

[0060] 12

[0061] Furthermore, Kalman filtering can be used, combining model predictions with actual measurements using a recursive algorithm. This produces a robust estimate that is particularly susceptible to measurement noise. While this method requires significant computational resources, these can easily be allocated during the generation of the training dataset.

[0062] The aforementioned methods are not suitable for continuous real-time monitoring. However, these methods are advantageous when generating the training dataset, as the module state can be determined with particular accuracy.

[0063] Another highly effective method for determining the training module state for the training dataset is capacity testing and internal resistance measurement. The latter method is non-invasive and can be performed at regular intervals. Temperature is specifically taken into account, as it affects resistance changes.

[0064] Steps a) to d) are performed multiple times, in particular with a variety of different training pulse characteristics. For example, at least one training pulse characteristic includes one or more of the following characteristics of the charge carrier exchange process:

[0065] The training pulse characteristic can be pulse duration. For example, training data sets can be generated where the pulse duration was 200 ms, 2 s, and / or 20 s. Alternatively or additionally, a training pulse characteristic can be pulse rate and / or pulse amplitude.

[0066] The training data sets are preferably collected at different pulse durations to ensure that the analysis model captures the behavior of the energy storage module, particularly the battery cell, under varying conditions. The same applies to pulse rate, pulse amplitude, and temperature. Alternatively or additionally, the state of health (SoH) and state of charge (SoC) levels can be varied. Therefore, it is possible that the training data sets will be different for various M / PUIN-022-PCs.

[0067] 13

[0068] State of Health (SoH) and State of Calorific Rate (SoC) levels can be recorded. Alternatively or additionally, rest periods and / or the type of pulses can be varied.

[0069] According to a further embodiment of the method, the multitude of different training pulse characteristics are variations of a target training pulse characteristic, which are within a range of ±10%, in particular ±5%, of the target training pulse characteristic. If the training pulse characteristic is the pulse duration, the training pulse characteristics can include variations of a target training pulse characteristic of ±5 ms.

[0070] The variation in pulse characteristics can affect, for example, the frequency, amplitude, or duty cycle (dual duty cycle). With alternative configurations, these variations can exceed the ranges of ±5% or ±10% and depend on the specific technology and application. For instance, training across a wider parameter range may be necessary to cover realistic scenarios.

[0071] These variations advantageously allow for the determination of how the energy storage module responds to variations in pulse characteristics. In particular, it can reveal what changes occur in the pulse response voltage, temperature, and / or current in response to variations in the pulse characteristics. This advantageously enables the generation of training datasets that facilitate improved training of an artificial neural network.

[0072] According to a further embodiment of the method, the generated training data set includes, as training voltage parameters, the difference between the training pulse response voltage and an average of pulse response voltages, and / or as training temperature parameters, the difference between the training pulse response temperature and an average of pulse response temperatures, and / or as training current parameters, the difference between the training pulse response current and an average of pulse response currents. For each charge carrier exchange with a training pulse characteristic, which is carried out at the training energy storage module, i.e., for example, at a battery cell, a multitude of training pulse response voltages and / or training pulse response temperatures are obtained. M / PUIN-022-PC

[0073] 14 and / or training pulse response current intensity. These can either be used unchanged, or the average is calculated for each value, and the average is subtracted from each value.

[0074] By selecting the appropriate training voltage, temperature, or current parameter, finer differences between the module states of the energy storage module can be detected. This normalization enables the trained neural network to more easily recognize patterns and fluctuations.

[0075] After the training datasets have been generated, the artificial neural network can be trained as follows: a) As described above, a large number of training datasets are generated, each with a different training pulse characteristic for one or more training energy storage modules; b) The training voltage parameter and / or the training temperature parameter and / or the training current parameter of a training dataset are provided to the artificial neural network as input parameters; c) The artificial neural network outputs a comparison module state of a training energy storage module as a training output parameter; d) The output comparison module state is compared to the specified training module state of the training dataset.e) At least one parameter and / or hyperparameter of the artificial neural network is adjusted depending on the result of this comparison. f) Steps b) to e) are performed for all training data sets generated in step a).

[0076] The parameters and / or hyperparameters of the artificial neural network are thus adjusted by comparing the comparison module state output by the artificial neural network with the training module state of the training dataset. The parameters and / or hyperparameters are then adjusted so that the comparison M / PUIN-022-PC

[0077] 15

[0078] The module state approaches the training module state of the training dataset, and the difference between these two module states becomes minimal.

[0079] The parameters of the artificial neural network, such as weights and distortions, can be adjusted during training using a technique called backpropagation. This method aims to minimize the error between the predicted module state (i.e., the module state generated by the artificial neural network) and the actual module state (i.e., the training module state). Various well-known optimization algorithms can be used for this purpose, updating the parameters based on gradients calculated from a loss function of the artificial neural network.

[0080] For example, ADAM or RMSprop can be used as optimization algorithms.

[0081] During training, the artificial neural network iteratively adjusts the parameters, particularly the weights, to minimize the loss function, specifically the mean absolute error or the mean squared error. The artificial neural network is thus trained using a training dataset that specifies the voltage and temperature responses to pulsed charging processes, especially pulsed currents, enabling the artificial neural network to establish a relationship between the voltage parameter and / or the temperature parameter and / or the current parameter on the one hand, and the module state, specifically the state of charge and the aging state of the energy storage module, on the other.

[0082] The energy management procedure for a vehicle that has a multilevel converter system with a variety of energy storage modules and switching devices comprises the following steps:

[0083] Determining a module state of an energy storage module according to one of the methods described above,

[0084] Perform at least one of the following actions based on the specified module state: M / PUIN-022-PC

[0085] 16

[0086] Controlling at least one switching device that disconnects or establishes a connection between the energy storage module and an electrical consumer of the vehicle;

[0087] Controlling at least one electrical consumer of the vehicle that is connected to the energy storage module;

[0088] Starting or stopping a charging or discharging process of the multilevel converter system.

[0089] Improved module condition monitoring, particularly of the state of charge and aging of a battery cell, can extend the lifespan of the energy storage module. This improved module condition monitoring can be used, in particular, to activate energy storage modules at the ideal time, thereby extending their lifespan.

[0090] A further aspect of the invention relates to a device for determining the module state of an energy storage module of a multilevel converter system comprising a plurality of energy storage modules and switching devices, wherein the multilevel converter system is configured to connect and / or bridge each energy storage module in parallel and / or in series with the respective adjacent energy storage module, and wherein each energy storage module has at least one energy storage cell. The device comprises a control unit configured to control a charging unit such that charge carriers are pulsedly exchanged with the energy storage modules in a charge carrier exchange process, according to at least one pulse characteristic.The device further comprises an analysis unit configured to measure at least one pulse response voltage and / or one pulse response temperature and / or one pulse response current of at least one energy storage module during the charge carrier exchange process, wherein a voltage parameter that is the measured pulse response voltage or is derived from the measured pulse response voltage, and / or a temperature parameter that is the measured pulse response temperature or is derived from the measured pulse response temperature, and / or a current parameter that is the measured pulse response current or is derived from M / PUIN-022-PC.

[0091] The measured pulse response current, derived from parameter 17, is provided as an input parameter to an analysis model of the analysis unit. An output unit coupled to the analysis unit then outputs the module state of the energy storage module as an output parameter.

[0092] This device is particularly designed to carry out methods according to the invention for determining the module state of an energy storage module of a multilevel converter system.

[0093] A further aspect of the invention relates to an energy management device for managing the energy of a vehicle that has a multilevel converter system with a plurality of energy storage modules and switching devices. The energy management device includes the device described above for determining the module state of an energy storage module of the multilevel converter system. The energy management device is configured to determine the module state of an energy storage module using the device described above and to perform at least one of the following actions based on the determined module state:

[0094] Controlling at least one switching device that disconnects or establishes a connection between the energy storage module and an electrical consumer of the vehicle;

[0095] Controlling at least one electrical consumer of the vehicle that is connected to the energy storage module;

[0096] Starting or stopping a charging or discharging process of the multilevel converter system.

[0097] This energy management device is particularly designed to carry out the inventive methods for energy management of a vehicle.

[0098] The invention will now be explained with reference to an exemplary embodiment and the attached drawings. M / PUIN-022-PC

[0099] 18

[0100] Figure 1 shows an example of the structure of a multilevel converter system,

[0101] Figure 2 shows an embodiment of a device for generating training data,

[0102] Figure 3 shows an example of an artificial neural network which is used in embodiments of the methods according to the invention and

[0103] Figure 4 shows an embodiment of a device for determining a module state.

[0104] Figure 1 shows an example of a multilevel converter system 1. It comprises a plurality of adjacent energy storage modules 10, 12, 14, 16. The energy storage modules 10, 12, 14, 16 each have at least one energy storage cell. In the present embodiment, the energy storage cell is a battery cell. The energy storage modules 10, 12, 14, 16 are each interconnected via several paths, with a switching device configured as a transistor 18 being arranged in each path. The adjacent energy storage modules 10, 12, 14, 16 can be connected in series or parallel by means of the transistors 18, or one of the energy storage modules 10, 12, 14, 16 can be bypassed. By bypassing an energy storage module 10, 12, 14, 16, it is excluded from a configuration if required.

[0105] The following describes a device and a method for generating training datasets for supervised learning with reference to Figure 2:

[0106] The training datasets can then be used to train an artificial neural network. The trained artificial neural network is then used as an analysis model to determine the module state of one of the energy storage modules 10, 12, 14, 16. In the embodiments described below, the module state is determined to be the state of charge and the aging state of the battery cell of the respective energy storage module 10, 12, 14, 16. M / PUIN-022-PC

[0107] 19

[0108] A training multilevel converter system 1' is used to generate the training data. The training multilevel converter system 1' corresponds to the multilevel converter system 1 as described with reference to Figure 1 and as used in the inventive method for determining the module state of the energy storage module of the multilevel converter system 1.

[0109] Alternatively, the training data can also be obtained using individual battery cells with a cell cycler. These individual battery cells are of the same type (chemistry, format, etc.) as the battery cells to be used in the multilevel converter system. In the cell cycler, the battery cell can be subjected to artificial aging by repeatedly charging and discharging it to investigate how the cell properties change over many cycles.

[0110] First, a capacity test is performed on a battery cell to determine its state of aging. This involves a complete discharge and a complete charge using a method that combines constant current and constant voltage charging (CC / CV method).

[0111] In the actual process for generating training data sets, a pulsed charge carrier exchange is performed in a training energy storage module of the training multilevel converter system 1', whereby the charge carrier exchange occurs with at least one training pulse characteristic. For this purpose, the training multilevel converter system 1' is connected to a charging unit 2 and a control unit 3. The training multilevel converter system 1', or rather the individual training energy storage modules of this system 1', can be charged and discharged by means of the charging unit 2, with this process being controlled by the control unit 3. In this embodiment, pulsed currents are used for the charge carrier exchange. Bipolar pulses are used, whose pulse lengths, pulse durations, and pulse frequencies are varied.These pulse characteristics correspond to typical pulse characteristics as they are typically used in the operation of a multilevel converter system 1 M / PUIN-022-PC.

[0112] 20

[0113] The pulse durations used to generate the training data sets are, for example, 200 ms, 2 s and 20 s.

[0114] In this embodiment, the pulse is composed of a multitude of individual pulses. It is therefore a main pulse, a sequence of several individual pulses. Each individual pulse represents a current flow phase. Between two individual pulses, there is a rest phase within the main pulse, during which essentially no current flows. During this main pulse, the parameters of the individual pulses vary (e.g., amplitude, pulse width, duration, duty cycle).

[0115] If the resting phase after a single pulse exceeds a minimum resting period during which no current flows, this single pulse is the last of the single pulses that make up the main pulse. The minimum resting period is, for example, 100 ms. The main pulse then extends from the beginning of the first current flow phase, i.e., the first single pulse, to the end of the last current flow phase, i.e., the end of the last single pulse.

[0116] Furthermore, a variety of different training pulse feature variations around specific target training pulse features can be used to generate training datasets. These variations can, for example, be in a range of + / - 10%, and in particular + / - 5%.

[0117] Subsequently, the system's response to the pulsed charge carrier exchange is measured. For this purpose, the device includes a determination unit 4. Specifically, the training pulse response voltage and / or the training pulse response temperature and / or the training pulse response current is measured. Thus, after a charge carrier exchange induced by a current pulse, the voltage and temperature of the battery cell are measured. Appropriate sensors are provided for this purpose in the training multilevel converter system 1'.

[0118] The training module status of the training energy storage module is then determined. In this embodiment, the state of charge and the aging state of the battery cell are determined. The training M / PUIN-022-PC

[0119] 21

[0120] The module state is determined using conventional state estimation methods. For example, a charge carrier count can be performed. Alternatively or additionally, the state of charge and the state of aging can be determined using the open-circuit voltage method, model-based methods, a capacitance test, and / or a measurement of the internal resistance, as is generally known.

[0121] These procedure steps can be performed for a variety of battery cell aging states. For each aging state, training pulse response voltages to bipolar pulses for charge carrier exchange can be collected at different battery cell states of charge. The battery cell's state of charge is determined each time using conventional state estimation methods, such as charge carrier counting. The battery cell can then be discharged stepwise, measuring the training pulse response voltage, temperature, and / or current at each state of charge. These discharge steps can be repeated until the battery cell reaches a cutoff voltage.

[0122] Furthermore, the steps described above are performed for a variety of battery cell aging states. After multiple charge carrier exchange processes, the battery cell is aged through a large number of cycles until a further aging state is reached. A capacity test is then performed again to determine the new aging state. Additionally, the pulsed charge carrier exchange processes and the corresponding measurements of the training pulse response voltage, temperature, and current are recorded, and the state of charge and aging state are again determined using conventional state estimation methods.

[0123] In an alternative embodiment, instead of measuring the training pulse response voltage, temperature, and current, the voltage responses are used to obtain the parameters of an equivalent circuit model of the battery cell. M / PUIN-022-PC

[0124] 22

[0125] In another alternative embodiment, the parameters of the equivalent circuit model of the battery cell are determined as an alternative to or in addition to the training pulse response voltage, training pulse response temperature, and training pulse response current.

[0126] In a further embodiment, the process steps described above are carried out for a large number of battery cells with different cell chemistry and / or cell formats. The battery cells are, in particular, lithium-ion batteries.

[0127] After each pulsed charge carrier exchange process, including the measurement of the training pulse response voltage, temperature, and current, and the determination of the battery cell's training module state, a training data set is generated. This data set includes at least one training voltage parameter, which is either the training pulse response voltage or derived from it. In this embodiment, the training data set also includes a training temperature parameter, which is either the training pulse response temperature or derived from it. Furthermore, the training data set includes the determined training module state. Optionally, the training data set can also include the cell chemistry of the battery cell used and the parameters of the equivalent circuit model.

[0128] In the present embodiment, a parameter derived from the training pulse response voltage is used as the training voltage parameter instead of the training pulse response voltage itself. In this case, the training voltage parameter is the difference between the measured training pulse response voltage and an average of pulse response voltages.

[0129] Similarly, the difference between the measured training pulse response temperature and an average of pulse response temperatures is used as a training temperature parameter. M / PUIN-022-PC

[0130] 23

[0131] Similarly, the difference between the measured training pulse response current and an average of pulse response currents is used as the training pulse response current.

[0132] The training data sets generated in this way are transferred from the determination unit 4 to a memory 5, which stores the training data sets.

[0133] These training datasets are used to train an artificial neural network. Figure 3 shows the structure of such an artificial neural network. In this case, it is a so-called feed-forward neural network.

[0134] The artificial neural network includes an input layer 20. The first input parameter 20-1, the training temperature parameter, and the second input parameter 20-2, the training voltage parameter, are entered into the artificial neural network via this input layer 20.

[0135] In another alternative embodiment, data on the cell chemistry of the battery cell are provided to the artificial neural network as a third input parameter 20-3. Furthermore, data on the cell format and / or the nominal capacity can optionally be provided to the artificial neural network.

[0136] In a further alternative embodiment, the parameters of the equivalent circuit model are provided to the artificial neural network either alternatively or additionally.

[0137] The input parameters 20-1 to 20-4 are each contained in the training data sets stored in memory 5.

[0138] The artificial neural network further comprises hidden layers 21 and an output layer 22, as is known per se. The first output parameter 22-1 of output layer 22 is the aging state of the battery cell. The second output parameter 22-2 is the state of charge of the battery cell. M / PUIN-022-PC

[0139] 24

[0140] During the training of the artificial neural network, as depicted in Figure 3, the previously acquired input parameters 20-1 and / or 20-2, and optionally the input parameters 20-3 and 20-4, are provided to the artificial neural network. The artificial neural network outputs an associated aging state and charge state of an associated battery cell via the output layer. These output parameters 22-1 and 22-2 represent a so-called output comparison module state. This comparison module state is compared with the predefined training module state of the training dataset.

[0141] The artificial neural network is then optimized so that the difference between the output comparison module state and the specified training module state is minimal. For this purpose, at least one parameter and / or a hyperparameter of the artificial neural network is adjusted depending on the result of this comparison.

[0142] The steps described above for training the artificial neural network are then performed for all training datasets. In this way, a trained artificial neural network is obtained which can output the output parameters 22-1 and 22-2 of output layer 22 in real time for the input parameters of input layer 20. The parameters and / or hyperparameters of the artificial neural network have been adjusted so that the output parameters 22-1 and 22-2 correspond very closely to the parameters of the training module state of the training dataset, which were obtained conventionally when generating the training dataset.

[0143] Before the artificial neural network is actually used in the operation of a multilevel converter system 1, for example in a vehicle, the analysis model provided by the artificial neural network can be validated. For example, a cross-validation process can be performed to find the ideal hyperparameters for a feed-forward neural network. Each hyperparameter set can be repeated several times, for example five times, and the average mean absolute error can be calculated. After the cross-validation process, all errors resulting from the different hyperparameters are aggregated, and the hyperparameter set with the lowest error is selected for the analysis model M / PUIN-022-PC.

[0144] 25 are selected. The analysis model is then tested on new battery cells. Only after such validation is the analysis model integrated into the device, for example in a vehicle.

[0145] In another embodiment, separate artificial neural networks are used to determine the aging state on the one hand and the charge state on the other. The first artificial neural network is then trained to minimize the deviation of the determined aging state from the aging state of the training data set, and the second artificial neural network is trained to minimize the deviation of the output charge state from the charge state of the training data set.

[0146] After training the artificial neural network to determine the state of charge and the aging state of a battery cell, or the two separate artificial neural networks to determine these states, they are integrated into a device for determining the module state of an energy storage module 10, 12, 14, 16 of a multilevel converter system 1. This device allows the module state to be determined in real time. The device is explained below with reference to Figure 4:

[0147] Multilevel converter system 1 corresponds to training multilevel converter system 1'. It is connected to both charging unit 2 and control unit 3. Charge carriers are exchanged with energy storage modules 10, 12, 14, and 16 via charging unit 2 and control unit 3 in a charge carrier exchange process. The energy storage modules 10, 12, 14, and 16 are charged and discharged in pulsed cycles. The charge carrier exchange occurs according to at least one pulse characteristic. This pulse characteristic relates to the pulse duration, pulse frequency, and / or pulse amplitude. The same or similar pulse characteristics were used when generating the training datasets so that the artificial neural network was adapted to the actual operation of multilevel converter system 1.

[0148] The load carrier exchange processes occur in this case during the intended operation of the multilevel converter system 1, for example, during vehicle operation. This includes driving, recuperation, or M / PUIN-022-PC operation.

[0149] 26

[0150] When charging at a charging station, pulsed charge carriers are exchanged using the multilevel converter system 1.

[0151] During such a charge carrier exchange process, the pulse response voltage and / or the pulse response temperature of at least one energy storage module, in particular all energy storage modules 10, 12, 14, 16, is measured. For this purpose, the analysis unit 6, which is connected to the control unit 3, is provided. A voltage parameter, which is the measured pulse response voltage or is derived from the measured pulse response voltage, is then obtained from the analysis unit 6. Alternatively or additionally, a temperature parameter, which is the measured pulse response temperature or is derived from the measured pulse response temperature, is obtained.

[0152] The trained artificial neural network(s) are stored in analysis unit 6. The voltage parameter and / or the temperature parameter are provided to the respective artificial neural network(s) as input parameters. Additionally, data on the cell chemistry of the battery cell and / or parameters of the equivalent circuit model can be provided to the artificial neural network(s) as further input parameters. The artificial neural network(s) output the aging state and the state of charge for one or all energy storage modules 10, 12, 14, 16 from output parameters 22-1 and 22-2 via an output unit 7 connected to analysis unit 6.In this way, the module state, namely the state of charge and / or the state of aging, of an energy storage module 10, 12, 14, 16, namely a battery cell, is determined in real time during the operation of the multilevel converter system 1.

[0153] The method for determining the module state of an energy storage module 10, 12, 14, 16 of the multilevel converter system 1 is then used in a method and an energy management device for the energy management of a vehicle that includes such a multilevel converter system 1. After the module state of an energy storage module 10, 12, 14, 16 has been determined, at least one switching device is activated, which is an M / PUIN-022-PC

[0154] 27

[0155] The connection of the energy storage module 10, 12, 14, 16 to an electrical consumer of the vehicle is disconnected or established. In this way, one or more energy storage modules 10, 12, 14, 16 can be switched on or off depending on their respective state of charge and / or aging. This allows the energy storage modules 10, 12, 14, 16 to be used particularly efficiently and increases the overall service life of the multilevel converter system 1.

[0156] Furthermore, the module state of an energy storage module 10, 12, 14, 16 determined by the method described above can be used to control at least one electrical consumer of the vehicle that is connected to the energy storage module 10, 12, 14, 16, or to start or stop a charging process of the multilevel converter system 1.

[0157] In the embodiment described above, voltage parameters and / or temperature parameters are used as input parameters for the analysis model. Alternatively or additionally, current parameters can also be used as input parameters for the analysis model, analogous to the voltage and / or temperature parameters.

[0158] M / PUIN-022-PC

[0159] 28

[0160] Reference symbol list:

[0161] 1 Multilevel converter system

[0162] 1' training multilevel converter system

[0163] 2 charging units

[0164] 3 Control unit

[0165] 4. Determination unit for training module status

[0166] 5 Memory slots for training data sets

[0167] 6 analysis units

[0168] 7 Output unit

[0169] 10, 12, 14, 16 Energy storage module

[0170] 18 transistors (switching devices)

[0171] 20 Input layer

[0172] 20-1 first input parameter

[0173] 20-2 second input parameter

[0174] 20-3 third input parameter

[0175] 20-4 fourth input parameter

[0176] 21 hidden layers

[0177] 22 Output layer

[0178] 22-1 first output parameter

[0179] 22-2 second output parameter

Claims

M / PUIN-022-PC 29 Claims 1. A method for determining the module state of an energy storage module of a multilevel converter system comprising a plurality of energy storage modules and switching devices, wherein the multilevel converter system is configured to connect and / or bridge each energy storage module in parallel and / or in series with the respective adjacent energy storage module, and wherein each energy storage module has at least one energy storage cell in which charge carriers are pulsedly exchanged with the energy storage modules in a charge carrier exchange process according to at least one pulse characteristic, during the charge carrier exchange process at least one pulse response voltage and / or one pulse response temperature and / or one pulse response current of at least one energy storage module is measured, a voltage parameter which is the measured pulse response voltage or is derived from the measured pulse response voltage, and / or a temperature parameter.which is the measured pulse response temperature or is derived from the measured pulse response temperature, and / or a current parameter which is the measured pulse response current or is derived from the measured pulse response current, are provided as input parameters to an analysis model, the analysis model outputs the module state of the energy storage module as an output parameter.

2. Method according to claim 1, characterized in that the analysis model comprises an artificial neural network, in particular a feed-forward neural network, a recurrent neural network, a support vector machine (SVM) or a long-short-term memory.

3. Method according to claim 1 or 2, characterized in that the analysis model has a calculation rule which maps an arbitrary pulse response voltage and / or pulse response temperature and / or pulse response current to a module state. M / PUIN-022-PC 30 4. Method according to one of the preceding claims, characterized in that the at least one pulse feature comprises one or more of the following features of the charge carrier exchange process: - the pulse duration, in particular a pulse duration of 200ms, 2s or 20s; - the pulse rate; - the pulse amplitude; - the duty cycle.

5. Method according to one of the preceding claims, characterized in that the analysis model is provided with the difference between the measured pulse response voltage and an average value of pulse response voltages as a voltage parameter and / or the analysis model is provided with the difference between the measured pulse response temperature and an average value of pulse response temperatures as a temperature parameter and / or the analysis model is provided with the difference between the measured pulse response current and an average value of pulse response currents as a current parameter.

6. Method according to one of the preceding claims, characterized in that the energy storage cell is a battery cell and the module state is a cell state of the battery storage cell.

7. Method according to one of the preceding claims, characterized in that the cell state is the state of charge and / or aging state of the battery cell and / or the performance state.

8. Procedure for generating training datasets for supervised learning, comprising the following steps: M / PUIN-022-PC 31 a) Performing a pulsed charge carrier exchange in a training energy storage module of a training multilevel converter system, wherein the charge carrier exchange is performed with at least one training pulse feature; b) Measuring a training pulse response voltage and / or a training pulse response temperature and / or measuring a training pulse response current; c) Determining a training module state of the training energy storage module after performing the charge carrier exchange; d) Generating a training data set comprising the following: - a training voltage parameter, which is the training pulse response voltage or is derived from the training pulse response voltage, and / or a training temperature parameter, which is the training pulse response temperature or is derived from the training pulse response temperature, and / or a training current parameter, which is the training pulse response current or is derived from the training pulse response current; and - the specific training module state.

9. Method according to claim 8, characterized in that in step c) the training module state of the training energy storage module is determined by means of charge carrier counting, open circuit voltage method, model-based methods, capacity testing and / or measurement of the internal resistance.

10. Method according to claim 8 or 9, characterized in that steps a) to d) are performed multiple times with a plurality of different training pulse characteristics.

11. Method according to one of claims 8 to 10, characterized in that the at least one training pulse feature comprises one or more of the following features of the charge carrier exchange process: - the pulse duration, in particular a pulse duration of 200 ms, 2 s or 20 s; - the pulse rate; M / PUIN-022-PC 32 - the pulse amplitude.

12. Method according to one of claims 8 to 11, characterized in that the plurality of different training pulse characteristics are variations of a target training pulse characteristic which are in a range of ±10%, in particular ±5%, of the target training pulse characteristic.

13. Method according to one of claims 8 to 12, characterized in that the generated training data set comprises as training voltage parameter the difference between the training pulse response voltage and an average of pulse response voltages and / or as training temperature parameter the difference between the training pulse response temperature and an average of pulse response temperatures and / or as training current parameter the difference between the training pulse response current and an average of pulse response currents.

14. Method according to claim 2, characterized in that the artificial neural network was trained by: a) generating a plurality of training datasets according to one of claims 8 to 13, with a plurality of different training pulse characteristics for one or more training energy storage modules; b) providing the training voltage parameter and / or the training temperature parameter of a training dataset as input parameters to the artificial neural network; c) the artificial neural network outputs a comparison module state of a training energy storage module as a training output parameter; d) comparing the output comparison module state with the predetermined training module state of the training dataset; e) adjusting at least one parameter and / or hyperparameter of the artificial neural network depending on the result of this comparison. M / PUIN-022-PC 33 f) where steps b) to e) are performed for all training data sets generated in step a).

15. Method for energy management of a vehicle having a multilevel converter system with a plurality of energy storage modules and switching devices, wherein the method comprises the following steps: Determining a module state of an energy storage module according to any one of claims 1 to 7; and Perform at least one of the following actions based on the specified module state: - Controlling at least one switching device that disconnects or establishes a connection between the energy storage module and an electrical consumer of the vehicle; - Controlling at least one electrical consumer of the vehicle that is connected to the energy storage module; - Starting or stopping a charging or discharging process of the multilevel converter system.

Images