Method for determining a state of charge of a rechargeable battery cell

Abstract

The invention relates to a method for determining a state of charge (SoC j ) of a rechargeable battery cell (10), wherein for a first time (j) a value (U j ) a voltage (U), a value (I j ) of a current (I) and a value (Tj) of a temperature (T) of the battery cell (10) are measured, wherein a previous state of charge (SoC j-n) is provided, which was determined at a second time (jn) prior to the first time (j), wherein an artificial neural network (20) is provided, which has a time-sensitive module (24) with at least one attention layer and / or self-attention layer (35) on an input side (22), and a backbone module (28) downstream of the time-sensitive module (24) in the signal flow direction (26), wherein the backbone module (28) comprises an activatable memory and / or forgetting function (C, f) and / or a transformer module, wherein the values ​​(U j , I j , T j ) the voltage (U), the current (I) and the temperature (T) as well as the previous state of charge (SOC j-n ) an input vector (X j ) is generated as an input for the artificial neural network (DNN), and the state of charge (SoC j) at the first time (j) is determined as an output by means of the artificial neural network (20) and is output on an output page (44) when the input vector (X j ) is entered on the input page (22).

Description

The invention relates to a method for determining a state of charge of a rechargeable battery cell, wherein a value of a voltage, a current and a temperature of the battery cell are measured in each case for a first point in time.For electrically driven motor vehicles (also referred to below as electric vehicles), precise estimation of a current state of charge (SoC) of the battery cells of the electric vehicle is required, in particular for efficient travel planning. The estimation is usually carried out on the basis of voltage, current and temperature values which are measured in a battery cell. An SoC can be determined from the measured values mentioned on the one hand by means of optionally tabulated cell parameterization or functional dependence determined beforehand in corresponding tests. However, such a functional dependence is very strongly dependent in particular on the specifically used cell chemistry of the battery cell, and also on the battery and often also on an ageing state, so that usually no closed expression (not even approximately) can be specified here, but cell parameterization is very inefficient as a result of the many variables.A corresponding model of the battery cell can also be applied to the measured values, such as Kalman filters or sequential Monte Carlo methods (or models derived therefrom). Such filters are capable of providing sufficiently precise estimation results for battery cells based on lithium nickel manganese cobalt oxides ("NMC"). For battery cells with more complex cell chemistry, or with non-trivial electrochemical properties such as high hysteresis or a flat open circuit voltage (OCV) curve.Increasingly, for estimating battery cell states of charge, use is also being made of machine learning-based approaches which use, for example, an artificial neural network ("deep neural network", DNN). However, such approaches can be used in a truly targeted manner only for the specific cell chemistry for which the respective DNN has been designed and trained.WO 2020 / 234743 A1 discloses an estimation of an SoC of a battery on the basis of measurement values for voltage, current and temperature at a given point in time by means of an autoregressive DNN. Moreover, via a battery model that uses amp-hour integration, a previous SoC is provided as another input value to the DNN.In US 2022 / 0 057 451 A1 a model for the behaviour of a lithium (Li) battery is mentioned, which uses a "Neural Ordinal Differential Equation" (neural ODE) in which a derivative of a hidden state (DNN) is parameterized and solved. Here, a "performance" submodel describes the actual behavior of the battery, while a "degradation" submodel describes the temporal changes in the performance submodel, such as aging processes. For training the DNN, the gradient of the resulting loss function is analyzed.CN 110 232 43 B mentions a DNN for ascertaining an SoC of a battery cell, wherein preceding estimated values are stored and used further for subsequent estimates. In particular, the structure (e.g. number of input and output neurons) of the DNN is also determined only from the historical SoC values.The invention is based on the object of specifying a method for determining an SoC of a rechargeable battery cell which, on the one hand, determines values of the SoC that are as precise as possible, on the other hand is as versatile as possible, that is to say can be used in particular for different cell chemistries, and is also to be as resource-efficient as possible.The object mentioned is achieved according to the invention by a method for determining an SoC of a rechargeable battery cell, wherein a value of a voltage, a current and a temperature of the battery cell are measured in each case for a first point in time, wherein a preceding SoC is provided, which was determined at a second point in time, which is in particular situated directly before the first point in time, wherein an artificial neural network is provided, which has a time-sensitive module having at least one attenuation layer and / or self-attenuation layer on an input side, and a backbone module downstream of the time-sensitive module in the signal flow direction, and wherein the backbone module comprises an activatable memory and / or forget function and / or a transformer module.Here, it is arranged that an input vector is generated from the values of the voltage, the current and the temperature of the battery cell measured at the first time and from the preceding SoC as an input for the artificial neural network, and the SoC is determined as an output at the first time by means of the artificial neural network and is output on an output side as a result when the input vector is input on the input side. Advantageous and partly per se inventive embodiments are the subject matter of the dependent claims and of the following description.In this case, an SoC of a battery cell of an electric vehicle is preferably determined, and in particular the determination takes place during the operation of the electric vehicle. An electric vehicle preferably includes any electrically driven motor vehicle (motor vehicle). The electric vehicle is in particular provided by a vehicle operated purely battery-electrically, that is to say without an internal combustion engine for hybrid operation. However, the electric vehicle can also be provided by a plug-in hybrid.A battery cell is understood here to mean, in particular, the smallest unit of a battery that is accessible from the outside, for which the measurements mentioned can be carried out. A battery cell in this case comprises in particular in each case an anode contact and a cathode contact for electrically contacting and connecting the battery cell to other battery cells to form a battery or a battery assembly, wherein the two said contacts are preferably in each case connected via current collectors to the associated electrochemically active layers in the interior of the battery cell (that is to say anode layers or cathode layers).The voltage and the current are now measured at the mentioned contacts for a specific point in time of a query, i.e. the first point in time. Moreover, at the first time, a temperature measurement of the battery cell is also carried out by means of a correspondingly configured measuring device (i.e. for example a thermometer with a temperature probe in the interior of the battery cell).Furthermore, an SoC is provided which was determined at a second point in time before the first point in time, in particular immediately before it. This preceding SoC may also have been determined, on the one hand, by the method described herein (then by providing an even more preceding SoC), or may have been determined in some other way, such as via amp-hour integration or the like.From the values of the voltage U j, the current I j and the temperature T j at the first time j and the SoC at the second time (SOC j-n) determined in this way, an input vector X j for a DNN still to be described is now generated, wherein in particular the individual components of the input vector are given by the mentioned variables, i.e.In particular, the input vector X can comprise j further components. The DNN, which is preferably implemented on suitable hardware of the electric vehicle, takes as input the input vector X j with the mentioned variables and generates the SoC from this as an output by corresponding propagation of the input vector X j by the DNN at the first point in time.For this purpose, the DNN is preferably suitably trained or provided with parameters ("weights") of the individual neurons, which are "learned" beforehand in a training phase on a structurally identical DNN (which, however, can be implemented physically on another hardware and in particular outside the electric vehicle), preferably on the basis of measurements of the respective SoC without the DNN and corresponding error comparisons (and backpropagation).The DNN has a time-sensitive module on an input side, at which the input vector is transferred to the DNN, and a backbone module behind the time-sensitive module in the signal flow direction. The time-sensitive module comprises at least one attenuation layer and / or self-attenuation layer, the function of which is described below. The backbone module comprises an activatable reminder and / or forget function and / or a transformer module, and is provided in particular with at least one long short-term memory (LSTM) layer. A layer is a layer of neurons in the DNN.In view of the pronounced temporal relationships in the practical behavior of battery cells, the inclusion of temporal information is of decisive importance for an accurate determination of the SOC of battery cells. This is taken into account in particular by the time-sensitive module, which enables the explicit modelling of dependencies between different inputs by an attenuation or self-attenuation mechanism within the architecture.Many existing approaches to determining SoC using a DNN each utilize recurrent neural networks (RNNs) with a recurrent entity (i.e., at least one neuron with a feedback loop) to model the complicated relationship between the SOC and the observable physical quantities such as voltage, current, and temperature of Li-ion battery cells. Compared to feedforward neural networks (RNNs with recurrent units have the ability to use past information, which provides them with a relative advantage for determining a SoC. This is due to the fact that the current SoC is closely interwoven with earlier and in particular immediately preceding SoC values.However, RNNs are often subject to an inherent limitation in long term dependence detection, primarily due to the phenomenon of so-called "gradient disappearance" ("vibrating gradient") during traditional backpropagation training. The present invention solves this problem by introducing a feedback mechanism through the use of the SOC j-n at the previous second time (in addition to the currently measured values U j, I j, T j). By integrating a previous SoC determination as an input and thus including a feedback loop, the model represented by the DNN is able to efficiently capture long-term dependencies and the temporal dynamics inherent in the battery system without a gradient disappearance within the recurrent architecture becoming an appreciable problem.The time-sensitive module thus enables the method to effectively record the temporal dependencies and patterns of the behavior of the battery cell that are present in the data. Conversely, the backbone module of the described type, as a result of its activatable memory and / or forget function, serves to model the sequential nature of the said measured values of the battery cell, which facilitates the detection of long-term dependencies.An activatable reminder and / or forget function includes in particular a number of neurons in one layer or a plurality of layers, which may temporarily store information if an associated activation function of the reminder of said function assumes a corresponding value for this purpose, wherein the stored information may be read out, possibly depending on a further activation function, until an activation function of the forget causes the information to be deleted. In particular, an LSTM layer or a so-called gated recurrent unit (GRU) layer has such an activatable memory and / or forget function.An attenuation or self-attenuation layer makes it possible to recognize relationships in an input vector X j( i.e. in the present case then the input vector, or a vector arising from the input vector by a corresponding embedding or encoding, which represents a battery cell model), and to prioritize individual components of the input vector for a predefined request. For this purpose, a matrix W Q trained specifically for the request is applied to the input vector in order to generate a request ("query") for the attention from each component x j of the input vector X j. The request Q j= W Q · X j is "applied" to a set of "key values" ("key"), which in turn are formed by a matrix W K specifically trained for the request, and represents the possible relationships in the entries of the input vector, which may be relevant for the request. For this purpose, the request Q j is multiplied by the matrix W TK.The result of the "request" against the key values is processed by an activation function, which can be provided, for example, by a softmax function or by a component-by-component logistic function (preferably with subsequent normalization), and multiplied by a value matrix W V trained specifically for the request, which represents the possible result values of the request, onto which the request is thus projected by the matrix multiplication.The self-attenuation in the self-attenuation layer described gives priority to various components of the input vector, and thus of the request vector, in different time steps. This improves the ability of the method to detect temporal dependencies between the features and physical observables, thereby improving the precision in determining the SoC. The method is thus able to determine the SoC sufficiently precisely for such different cell chemistries as LiNiMnCoO2(NMC) and LiFePO4(LFP).Preferably, in a training phase of the artificial neural network, a value of a voltage, a current, and a temperature of the battery cell are each measured at a plurality of training times, wherein for each of the training times, in particular on the basis of the respectively determined voltage value and / or current value and / or temperature value, an SoC is determined independently of the artificial neural network as a reference SoC for the respective training time, wherein for each of the training times the SoC is determined by the artificial neural network on the basis of the respectively determined voltage value, current value and temperature value and on the basis of a preceding SoC which is determined at a time lying before the respective training time, and wherein internal parameters for the artificial neural network are determined and adapted accordingly for each training time on the basis of the reference SoC and the SoC determined by the artificial neural network. In the training phase, therefore, the SoC generated by the DNN in each time step is in particular matched to a SoC measured elsewhere (i.e. without the DNN). The SoC can be measured elsewhere here in particular by means of an open-circuit voltage and / or an amp-hour integration. The weights of the individual connections are preferably trained as internal parameters of the DNN. Depending on the specific implementation, the weights of the backbone module and of the self-attention layer and, if appropriate, of yet further layers can be trained successively.Advantageously, at the first time, an ageing state and / or a residual capacity and / or an internal resistance of the battery cell is additionally measured, wherein the ageing state measured at the first time or the measured residual capacity or the measured internal resistance enters the request vector as a component in each case. This additional information can be determined with reasonable effort and is often used for the battery cells anyway in an electric vehicle. By including the above-mentioned variables in the determination of the SoC, the precision can be further increased.Advantageously, the artificial neural network is provided with a backbone module, which has at least one LSTM layer. An LSTM layer comprises the required reminder and / or forget function, which can temporarily store information depending on a reminder activation function until a deletion of the information is caused by a forget activation function. As a result, those temporal relationships in the measured variables which were determined as relevant by the self-attenuation layer of the time-sensitive module can be stored until they lose relevance by successive subsequent information regarding the progressive development of the SoC and the physical variables in the battery cell (which is determined by the forget-activation function).However, a number of GRU layers or a transformer architecture can also be used as backbone module. However, it has been found that LSTM is already very effective in detecting sequential dependencies over longer periods of time. In contrast, a transformer architecture provides a very universal possibility of modelling temporal and contextual dependencies. However, a transformer architecture typically requires very large amounts of data for this purpose and therefore also has higher computing and storage requirements. Thus, a backbone module based on a number of LSTM layers includes an advantageous compromise between accurate detection of long term dependencies and computational efficiency, which is particularly suitable for use in an electric vehicle.It has proven to be further advantageous if the artificial neural network is provided with a time-sensitive module which comprises a bypass function which additionally passes an input carried out into the self-attenuation layer past the self-attenuation layer, and links it to the output resulting from the self-attenuation layer by a linear arithmetic operation, preferably a component-by-component multiplication. The self-attenuation layer thus not only represents a relationship between the individual components of the input vector (or a vector formed from the input vector by encoding or embedding), but also applies a prioritization of the components resulting therefrom directly to these.Expediently, the artificial neural network is provided with a time-sensitive module, which has at least one embedding of time information behind the at least one self-attention layer. As a result, an input vector X j can be input into the time-sensitive module, and an intermediate result resulting from the self-attention layer, provided with the time information, can be transferred to the backbone module and stored there over a plurality of time steps by the reminder function and used accordingly, while already subsequent vectors X j+1, X j+2 etc. are transferred to the time-sensitive module for the determination of the self-attention.In an advantageous embodiment, the artificial neural network is provided with a time-sensitive module in which a softmax function is used as an activation function in the at least one self-attenuation layer. The softmax function for a vector-valued argument v is the exponential function of the respective vector component v j, normalized by the sum of all the components exposed, i.e.The softmax function (which can also be formed by an exponential function on a basis other than e, e.g. 10) thus amplifies existing differences between the components v j of the vector v and thus more clearly emphasize the relationships resulting from the input vector between the values I j, U j and T j of the current I, the voltage U and the temperature measured at the first point in time and the preceding state of charge SoC j-n.The artificial neural network is preferably provided downstream of the backbone module with at least one fully connected layer in the signal flow direction. A fully connected layer is a layer in which each neuron of the layer is connected to each neuron of the preceding layer, and preferably to each neuron of the following layer. As a result, the logical and temporal relationship of the input data (and of the preceding SoC) which is detected by the time-sensitive module and has been temporarily stored in the backbone module and processed with information from other points in time can be read out particularly efficiently and in the process mapped to an SoC value.The invention further specifies an electrically driven motor vehicle, comprising at least one rechargeable battery cell, means for measuring a voltage, a current and a temperature of the battery cell, means for providing an initial SoC, means for providing an artificial neural network, which has on an input side a time-sensitive module having at least one attention layer and / or self-attention layer, and, downstream of the input side in the signal flow direction, a backbone module, which comprises an activatable memory and / or forget function and / or a transformer module. Furthermore, the electric motor vehicle comprises means for determining an SoC of the battery cell on the basis of the artificial neural network as an output on an output side by means of an input vector of the artificial neural network, which in each case comprises a value of a voltage, a current and a temperature at a point in time of a query and an initial SoC at a point in time preceding the point in time of the query.The electric motor vehicle according to the invention shares the advantages of the method according to the invention for determining an SoC of a rechargeable battery cell. The advantages mentioned for the method and for its developments can be transferred analogously, in a mutatismutant manner, to the electric motor vehicle.An exemplary embodiment of the invention is explained in more detail below with reference to drawings. Here, in each case, diagrammatically show: FIG. 1 is a block diagram of an electric vehicle having a rechargeable battery made up of individual battery cells and a computer unit having a DNN for determining the SoC of the battery cells, FIG. 2 shows a layout for the DNN of FIG. 1 ; and FIG. 3 shows a block diagram of a method for determining an SoC for a battery cell of the electric vehicle according to FIG. 1 by means of the DNN according to FIG. 2.Mutually corresponding parts and sizes are each provided with the same reference numerals in all figures.FIG. 1 is a schematic block diagram of an electrically driven motor vehicle 1 (hereinafter also electric vehicle 1 for simplification purposes), which comprises a battery system 2 with a rechargeable battery 4 for supplying individual electric motors 6 of the respective wheels 8 with electrical energy via a corresponding cable system 9. The electric motors 6 can also be replaced here by one or two electric axle motors which drive one axle each instead of the direct drive of the individual wheels 8 shown here. The battery system 2 has a plurality of battery cells 10 which are embodied in the present case as Li-ion cells and which are interconnected in the battery 4 in a manner not shown in more detail, so that the battery system 2 discharges the individual battery cells as homogeneously as possible when power is drawn from the battery 4. The battery system 2 further comprises a battery management system, not shown in detail, which is configured, inter alia, to control a discharging process of the battery 4 during the power extraction and a charging process of the battery 4, i.e. in particular to monitor and correspondingly regulate the compliance with electrical characteristic variables such as maximum voltages and maximum currents.The battery system 2 has a number of voltage measuring devices 12, a number of current measuring devices 14 and a number of temperature measuring devices 16, which are each preferably assigned to individual battery cells 10, and by means of which a voltage U, a current I and a temperature T can be measured for the respective battery cell 10 concerned. In particular, individual voltages or currents can be electrically tapped at the respective battery cell 10 for the voltage and current measurement devices 14, wherein the actual measurement of the tapped current I or the tapped voltage U takes place outside the battery 4, for example in measurement circuits of the battery system 2 specifically configured for this purpose. The temperature measurement devices 16 can in particular comprise temperature probes, not shown in detail, which penetrate into the interior of the individual battery cells 10 or measure the temperature at the surface of the battery cells 10.The electric vehicle 1 also has a computer unit 18, which can be implemented in a central vehicle onboard computer, not shown in more detail, on which fundamental control functions and driving assistance systems are also implemented, among other things. However, the computer unit 18 can also be implemented on dedicated hardware which does not implement any other function beyond that described below for determining an SoC.In the computer unit 18, a DNN 20 is implemented in a manner yet to be described, by means of which an SoC of the individual battery cells 10 is determined, for which purpose the measured values of the voltage U, the current I and the temperature T of the respective battery cell 10 are used as input values to the DNN 20, which are calculated by means of the above-mentioned. The measuring devices for measuring the voltage of the measuring devices 12, the measuring devices 14 and the temperature measuring devices 16 were measured.FIG. 2 shows the layout of the DNN 20 according to FIG. 1. The DNN 20 has a time-sensitive module 24 on an input side 22. Downstream of the time-sensitive module 24 in the signal flow direction 26, the DNN 20 has a backbone module 28 with a number of LSTM layers 30. Downstream of the backbone module 28 in the signal flow direction 26, the DNN 20 has at least one fully connected layer 32.At the input side 22, an input vector X j is passed to an input layer 33 of the DNN. The input vector X j in this case contains as components values U j, I j and T j of the voltage U, of the current I and of the temperature T, each measured at a specific point in time j. Moreover, the input vector X j has, as a further component, a value SoC j-n of the state of charge SoC, which value is preferably immediately preceding.The input vector X j can still be embedded or encoded by the input layer 33 accordingly for the subsequent layers.The input vector X j or a modified input vector X j' resulting from said encoding is now transferred to the time-sensitive module 24. This has a self-attention layer 35 in which a request Q j= W Q · X j(') is generated from the input vector X j or from the modified input vector X j' by multiplication with a weighting matrix W Q trained specifically for the request. The request Q j is "applied" by matrix multiplication to a set of "key values", which in turn are formed by a weighting matrix W K specifically trained for the request. A first activation function F1 is applied to the product Q j · W Tk which is given in the present case by a softmax function. The vector-valued result of the activation function F 1 (Q j · W TK) is multiplied by a value matrix W V specifically trained for the request, which represents the possible result values of the request. The self-attention layer 35 thus works out a temporal relationship between the individual components of the (modified) input vector X j(').The time-sensitive module 24 further comprises a bypass function 40 which additionally passes the (modified) input vector X j(') around the self-attenuation layer 35 and multiplies it component by the vector Y j= F1 (Q j ·W TK) ·W V resulting from the self-attenuation layer 35. By means of the bypass function 40, the temporal relationship, worked out by the self-attenuation layer 35, between the individual components of the (modified) input vector X j(') and the resulting prioritization of the components is also applied directly to said components.The time-sensitive module 24 also includes embedding time information 42. As a result, the vector Y j resulting from the self-attention layer 35 can be transferred to the backbone module 28 after said combination with the (modified) input vector X j('), which has passed the bypass function 40, as an intermediate result Z j provided with the time information, and, in a manner to be described below, a plurality of time steps can be stored by a reminder function and used accordingly, while subsequent vectors X j+1, X j+2 etc. are already transferred to the time-sensitive module 24 for the determination of the self-attention in the self-attention module 35.Backbone module 28 has a number of LSTM layers 30, a possible implementation of which is shown schematically and simplified in FIG. 2 for the sake of completeness. The LSTM layer 30 has an activatable forget function f, the value fj of which depends on the one hand on the input (here the vector Z j) transferred to the LSTM layer 30 and on the other hand on an immediately preceding output h j-1 of the LSTM layer 30. Furthermore, the LSTM layer 30 has a cell C in which, depending on the input (here the vector Z j) and the immediately preceding output h j-1 an internal state C j-1 of the LSTM layer 30, which is used for the calculation of the respective output of the LSTM layer 30 (and thus also of the immediately preceding output h j-1) is continuously updated in each step depending on the input Z j and on the current value f j of the forget function f (that is to say C j-1 → C j). The cell C thus forms an activatable memory function, since the values of the input Z j can be used further in a modified form for later purposes by updating the internal state C j-1 → C j of the cell C. Finally, depending on the immediately preceding output h j-1, from the input Z j and from the updated internal state C j of the cell C, a current output h j of the LSTM layer 30 is generated, which on the one hand can be passed on to subsequent layers such as the at least one fully connected layer 32 and on the other hand is also used further for the internal calculations of the LSTM layer 30 in the next time step. It should be noted that the output h j of the LSTM layer 30 is vector valued.The individual operations of the activatable forget function f, the activatable reminder function formed by the cell C and the generation of the output h j are carried out here by means of corresponding, specifically trained weighting matrices W f, W i, W C, W o( and optionally further weighting matrices or bias vectors, in each case not shown), activation functions (such as sigmoid functions) being used component by component, which determine a "degree of reminder or forget".The output h j of the LSTM layer 30 is now transferred to the at least one fully connected layer 32, where the current state of charge SoC j is generated and output at an output side 44 of the DNN 20.FIG. 3 schematically shows a block diagram of the sequence of a method according to which the SoC is determined in the electric vehicle 1 according to FIG. 1 with the aid of the DNN according to FIG. 2.In a training phase 50, values U a-m, I a-m, T a-m of the voltage U, of the current I and of the temperature T of a battery cell 10' identical in construction to the battery cell 10 are respectively measured at a plurality of training times 52 a- munder laboratory conditions. Moreover, for the training times 52 a- mduring which the battery cell 10' is connected to an electrical load and is thus successively discharged, a reference state of charge SoC-R a-m is determined in each case independently of the DNN 20, for example by means of amp-hour integration.For each of the training times 52 a- m, the respective state of charge SoC a-m is now determined by the DNN 20 on the basis of the values U a-m, I a-m, T a-m the voltage U, the current I and the temperature T of the battery cell 10'. For this purpose, the time of the state of charge SOC a-1,..., m-1 preceding the respective training time 52 a- mis also provided.On the basis of the respective deviation Δ a-m of the reference state of charge SoC-R a-m determined for the respective training time 52 a- mfrom the state of charge SoC a-m determined in each case by the DNN 20, the DNN 20 is now trained in that in particular the weighting matrices W Q, W K, W V of the self-attention layer 35 and the weighting matrices W f, W i, W C, W o of the LSTM layer 30 (and further ones, Weighting matrices, not shown, for example of the fully connected layer 32) can be adapted (preferably by error backpropagation) as a function of the deviations Δ a-m.During operation of the electric vehicle 1, the values U j, I j and T j of the voltage U, of the current I and of the temperature T of the battery cell 10 are now determined at a concrete first point in time j of a query. Moreover, for a second time j-n preceding the first time j, the state of charge is provided as a preceding state of charge SOC j-n. The same may in particular have been determined in an earlier step in an analogous manner by the DNN 20 or have been determined by amp-hour integration (in particular at initial values). The input vector X j is now formed from the voltage U j, measured at the first point in time j, the measured current I j, the measured temperature T j and the preceding state of charge SOC j-n. The input vector X j can additionally contain an internal resistance R int,j, an aging state SoH j( "state of health") or further measured variables of the battery cell at the first time j (dashed line), in this case the training phase 50 described above changes accordingly.From the input vector X j the DNN 20 then generates the state of charge SoC j at the first time j. The same is displayed to a driver of the electric vehicle 1 and / or used in the onboard computer of the electric vehicle 1 for charge planning and further stored for subsequent calculations of the state of charge SoC j+1 at a subsequent time j+1.Although the invention has been illustrated and described in more detail by the preferred exemplary embodiment, the invention is not restricted by the disclosed examples and other variations can be derived therefrom by the person skilled in the art without departing from the scope of protection of the invention.List of reference characters1 Electrically driven motor vehicle / electric vehicle 2 Battery system 4 Battery 6 Electric motor 8 Wheel 9 Cable system 10 (') Battery cell 12 Voltage measuring device 14 Current measuring device 16 Temperature measuring device 18 Computer unit 20 DNN 22 Input side 24 Time-sensitive module 26 Signal flow direction 28 Backbone module 30 LSTM layer 32 Fully-Associated layer 33 Input layer 35 Self-Attention layer 40 Bypass function 42 Embedding of time information 44 Output side 50 Training phase 52 a- m Trainings times C Cell / activatable memory function C j-1 / j Internal state of the cell f (j)( Value of the) activatable forget function(s) F1 First Activation function h j-1 / j output (of the LSTM layer) I (j)( value of) current(s) Q j request R int,j internal resistance SoC state of charge SOC j-n preceding value of state of charge SoH j state of aging T (j)( value of) temperature U (j)( value of) voltage W Q / K weighting matrix W f / i / C / o, Weighting matrix W V Value matrix X j(')( modified) Input vector Y j Resulting vector Z j Intermediate result (input into the LSTM layer) Δ a-m DeviationReferences included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedWO 2020 / 234743 A1

[0005] US 2022 / 0 057 451 A1

[0006] CN 110 232 43 B

[0007]

Claims

Method for determining a state of charge (SoC j) of a rechargeable battery cell (10), wherein for a first point in time (j) a value (U j) of a voltage (U), a value (I j) of a current (I) and a value (Tj) of a temperature (T) of the battery cell (10) are respectively measured, wherein a preceding state of charge (SOC j-n) is provided, which was determined at a second point in time (j-n) situated before the first point in time (j), wherein an artificial neural network (20) is provided, which has - on an input side (22) a time-sensitive module (24) with at least one attention layer and / or self-attention layer (35), and - after the time-sensitive module (24) in the signal flow direction (26) a backbone module (28), wherein the backbone module (28) comprises an activatable reminder and / or forget function (C, f) and / or a transformer module, wherein from the values (U j, I j, T j) of the voltage (U) measured at the first time (j), The method according to the invention is to generate an input vector (X j) as an input for the artificial neural network (DNN) of the current (I) and the temperature (T) and from the previous state of charge (SoC j-n) and wherein the state of charge (SoC j) is determined at the first time (j) by means of the artificial neural network (20) as an output and is output on an output side (44) when the input vector (X j) is input on the input side (22).Method according to Claim 1, wherein, in a training phase (50) of the artificial neural network (20), a value (U a-m, I a-m, T a-m) of a voltage (U), of a current (I) and of a temperature (T) of the battery cell (10) are in each case measured at a plurality of training times (52a-m), wherein a reference state of charge (SoC-R a-m) is determined for the respective training time (52a-m) for each of the training times (52a-m) independently of the artificial neural network (20), wherein for each of the training times (52a-m), the state of charge (SoC a-m) is determined by the artificial neural network (20) on the basis of the respectively determined voltage value (U a-m), current value (I a-m) and temperature value (T a-m) and on the basis of a preceding state of charge (SoC a-1,...,m-1) which is determined at a time lying before the respective training time (52a-m), and wherein internal parameters for the artificial neural network (20) are determined and adapted accordingly for the respective training time (52a-m) on the basis of the reference state of charge (SoC-R a-m) and the state of charge (SoC a-m) determined by the artificial neural network (20).Method according to Claim 1 or Claim 2, wherein at the first time (j) an ageing state (SoH j) and / or a residual capacity and / or an internal resistance (R int,j) of the battery cell (10) is additionally measured, and wherein the ageing state (SoH j) measured at the first time (j) or the measured residual capacity or the measured internal resistance (R int,j) each enters the input vector (X j) as a component.Method according to one of the preceding claims, wherein the artificial neural network (20) is provided with a backbone module (28) which has at least one long short-term memory layer (30).Method according to one of the preceding claims, wherein the artificial neural network (20) is provided with a time-sensitive module (24) which comprises a bypass function (40) which additionally passes the self-attenuation layer (35) with an input made into the self-attenuation layer (35), and links it to the output (Y j) resulting from the self-attenuation layer (35) by means of a linear arithmetic operation.Method according to one of the preceding claims, wherein the artificial neural network (20) is provided with a time-sensitive module (24) which has at least one embedding of time information (42) after the at least one self-attention layer (35) in the signal flow direction (26).Method according to one of the preceding claims, wherein the artificial neural network (20) is provided with a time-sensitive module (24), in which a softmax function is used as a first activation function (F1) in the at least one self-attention layer (35).Method according to one of the preceding claims, wherein the artificial neural network (20) is provided after the backbone module (28) with at least one fully connected layer (32) in the signal flow direction (26).Electrically driven motor vehicle (1) comprising: - at least one rechargeable battery cell (10), - means (12, 14, 16) for measuring a voltage (U), a current (I) and a temperature (T) of the battery cell (10), - means for providing an initial state of charge (SoC j-n), - means (18) for providing an artificial neural network (20) which - - has a time-sensitive module (24) having at least one attention layer and / or self-attention layer (35) on an input side (22), and - has a backbone module (28) downstream of the input side (22) in the signal flow direction (26), said backbone module having an activatable memory and / or forget function (C, f) and / or a transformer module, and - means for determining a state of charge (SoC j) of the battery cell (10) on the basis of the artificial neural network (20) as an output on an output side (44) by means of an input vector (X j) of the artificial neural network (20), which respectively comprises a value (U j, I j, T j) a voltage (U), a current (I) and a temperature (T) at a point in time (j) and the initial state of charge (SoC j-n) preceding the point in time (j).

Images