DE102022200022A1 - Method and device for providing an aging state model for determining a current or predicted aging state for an electrical energy store using neural differential equations - Google Patents

Abstract

The invention relates to a computer-implemented method for providing a state of health model for determining a modeled state of health (SOH) of an electrical energy store (41) with at least one electrochemical unit, in particular a battery cell, in a technical device, with the following steps: - providing (S1) training data sets , which in each case, based on a specific point in time, assigns an empirically determined aging state to an operating variable profile (F(t)) of one or more operating variables (F);- providing a data-based aging model, which is designed with a neural differential equation system (10) to internal electrochemical states of the energy store (41), wherein the internal states (x(t)) can be mapped to an aging state, wherein at least one of the differential equations of the differential equation system (10) is a sum of a deterministic model term and a data-based correction term formed by a data-based correction model is, has, - training (S2, S3) of the data-based aging model based on the training data sets, so that model parameters of the data-based correction model are determined.

Description

technical field

The invention relates to electrical devices operated independently of the mains with electrical energy stores, in particular electrically driven motor vehicles, in particular electric vehicles or hybrid vehicles, and also to measures for determining a current or predicted state of health (SOH: State of Health) of the electrical energy store. Furthermore, the invention relates not only to mobile but also to stationary electrical energy stores.

Technical background

The power supply of mains-independent electrical devices and machines such. B. electrically driven motor vehicles, using electrical energy storage devices, usually device batteries or vehicle batteries. These provide electrical energy to operate the devices. However, energy converters, such as e.g. B. fuel cell systems, including hydrogen tank, into consideration.

Electrical energy stores or energy converters degrade over their service life and depending on their load or use. This so-called aging leads to a continuously decreasing maximum performance and storage capacity. The state of aging corresponds to a measure of the aging of energy storage devices. By convention, a new energy storage device has an aging state of 100% of its available capacity, which progressively decreases over its lifetime. A measure of the aging of the energy store (change in the state of aging over time) depends on an individual load on the energy store, i. H. in vehicle batteries of motor vehicles on the usage behavior of a driver, external environmental conditions and the vehicle battery type.

Although the current aging state of the energy storage device can be determined using a physical aging state model based on historical performance characteristics, this model is imprecise in certain situations. This inaccuracy of the conventional aging model makes it difficult to precisely determine the state and to predict the course of the aging state. However, the prediction of the progression of the state of aging of the energy store is an important technical variable, since it enables the remaining service life to be determined and a residual value of the energy store to be economically evaluated. Furthermore, the prediction of the aging condition is of added value in order to plan and carry out predictive maintenance intervals.

Disclosure of Invention

According to the invention, a method for providing an aging status model according to claim 1, a method for providing an aging status and corresponding devices according to the independent claims are provided.

Further developments are specified in the dependent claims.

• - Bereitstellen von Trainingsdatensätzen, die jeweils bezogen auf einen bestimmten Zeitpunkt einen empirisch bestimmten Alterungszustand einem Betriebsgrößenverlauf einer oder mehrerer Betriebsgrößen zuordnet;
• - Bereitstellen eines datenbasierten Alterungsmodells, das mit einem neuronalen Differentialgleichungssystem ausgebildet ist, um interne elektrochemische Zustände des Energiespeichers zu modellieren, wobei die internen Zustände auf einen Alterungszustand abbildbar sind, wobei mindestens eine der Differentialgleichungen des Differentialgleichungssystems eine Summe eines deterministischen Modellterms und eines datenbasierten Korrekturterms, der durch ein datenbasiertes Korrekturmodell gebildet wird, aufweist,
• - Trainieren des datenbasierten Alterungsmodells basierend auf den Trainingsdatensätzen, so dass Modellparameter des datenbasierten Korrekturmodells bestimmt werden.

According to a first aspect, a computer-implemented method for providing a aging state model for determining a modeled aging state of an electrical energy store with at least one electrochemical unit, in particular a battery cell, in a technical device is provided, with the following steps:

• - Provision of training data sets, each related to a specific point in time, assigns an empirically determined aging state to a performance variable curve of one or more performance variables;
• - Providing a data-based aging model, which is designed with a neural differential equation system to model internal electrochemical states of the energy store, wherein the internal states can be mapped to an aging state, wherein at least one of the differential equations of the differential equation system is a sum of a deterministic model term and a data-based correction term, which is formed by a data-based correction model,
• - Training the data-based aging model based on the training data sets, so that model parameters of the data-based correction model are determined.

Aging state models for predicting an aging state can typically be provided as electrochemical models or data-based models. While physical aging models require a high level of expert knowledge in order to parameterize the differential equations contained therein, data-based aging models usually do not require expert knowledge, but allow no or only limited extrapolation. Therefore, an attempt is usually made to combine physical aging models with data-based models in order to obtain a good model prediction for a current len state of health as well as a good prediction accuracy for the state of health.

Such a combination has the disadvantage that the data-based correction model only compensates for modeling errors in the physical aging model, but does not allow any influence on the calculation of the aging state based on the physical aging model. The above method therefore envisages providing an aging state model which is based on neural differential equations. In this regard, the time-variant differential equations of the physical aging model, which are based on models of an internal system state, are provided as neural differential equations. Neural differential equations correspond to differential equations in which the physical term is supplemented by a data-based correction term on the right-hand side.

The data-based term is usually represented by a data-based regression model, in particular in the form of a Gaussian process model or a neural network. Neural differential equations have the advantage that the physical expert knowledge, which is available in the form of parameterized differential equations in the case of the physical aging model, can be supplemented by a data-based model component that represents a correction function with which deviations of the physical aging model from the real behavior of the energy storage devices are compensated can become.

annehmen, wobei der Modellterm f (α, x, F, t) das physikalische Alterungsmodell bezüglich eines internen Systemzustands x und zum Zeitpinkt t geltenden Werten der Betriebsgrößen F verkörpert und basierend auf Expertenwissen mit dem Parametervektor α parametriert ist. Hierbei kann der Zustandsvektor x z. B. Gleichgewichts- oder kinetische Parameter, und im Speziellen folgende elektrochemische Zustandsgrößen als interne Zustände oder Parameter umfassen:

• - Menge an zyklisierbarem Lithium,
• - Volumenanteil Anode,
• - Volumenanteil Kathode,
• - Reaktionsrate oder Reaktionskoeffizient bzgl. Anode,
• - Diffusionskoeffizient der Anode,
• - Schichtdicke bzgl. Anode,
• - Reaktionsrate oder Reaktionskoeffizient bzgl. Kathode,
• - Diffusionskoeffizient der Kathode,
• - Porositäten in Anode,
• - Porositäten in Kathode,
• - Bruggemann-Koeffizienten in Anode,
• - Bruggemann-Koeffizienten in Kathode,
• - Elektrolytkonzentration,
• - Kontaktwiderstände,
• - mechanische Partikelbelastung.

The neural differential equation can take the form $$\dot{x}\left(t\right)=ƒ\left(a,x,f,t\right)+G\left(\beta ,x,f,t\right)$$

assume, where the model term f (α, x, F, t) embodies the physical aging model with regard to an internal system state x and the values of the operating variables F valid at time t and is parameterized with the parameter vector α based on expert knowledge. Here, the state vector x z. B. Equilibrium or kinetic parameters, and in particular include the following electrochemical state variables as internal states or parameters:

• - Amount of cyclable lithium,
• - volume fraction anode,
• - volume fraction cathode,
• - reaction rate or reaction coefficient with respect to the anode,
• - diffusion coefficient of the anode,
• - layer thickness in relation to the anode,
• - reaction rate or reaction coefficient with respect to the cathode,
• - diffusion coefficient of the cathode,
• - porosities in anode,
• - porosities in cathode,
• - Bruggemann coefficients in anode,
• - Bruggemann coefficients in cathode,
• - electrolyte concentration,
• - contact resistances,
• - mechanical particle pollution.

Other parameters for electrochemical or mechanical or physical modeling can also be included.

g(β, x, F, t) corresponds to a data-based component, where β corresponds to a model parameter vector of the data-based correction term. The model parameter vector includes model parameters and hyperparameters of the underlying data-based model. The model term f(α, x, F, t) includes electrochemical action chains that are modeled from non-linear state equations. The differential terms on the left-hand side of the differential equations define internal states of the energy storage device, with an aging state SOH=k(γ,x) being able to be determined from a functional combination of the internal states and combination parameters γ, in particular from a linear combination of the internal states x.

In order to model and predict an aging state with the neural differential equation system, the course of operating variables is evaluated, in particular by solving the differential equation system using a numerical time integration method. For this purpose, in the case of a device battery as an energy store, the operating variables such as battery current, battery temperature, battery voltage, state of charge and/or the like must be available as high-resolution time series.

The calculation of such a physical aging model is complex and cannot be carried out in a control device of the technical device due to a lack of computing capacity. In this regard, the evaluation can be carried out in a central unit that is in communication with the technical device. To do this, however, it is necessary to transmit the high-resolution time series of the operating variables to the central unit, which leads to a high volume of data.

The extension to a neuronal differential equation is carried out by supplementing each of the individual differential equations with the data-based correction term β, which depends on the model parameter vector and the internal state variables and advantageously also on operating variables.

In contrast to conventional physical aging models, which are made more precise or corrected by a suitable correction model learned from the residuals, the correction term in the above approach for a neural differential equation system acts deeper in the model and is included in the integration of the differential equation, which is the case with conventional data-based or hybrid aging state models is not possible.

The data-based correction model can include a deep neural network, where the training is performed using a gradient-based method by minimizing a predetermined loss function.

Training data sets are used to train the neural differential equation. B. when it is put into operation, each assigned to an aging state reached after a period of operation. These training data sets can be identical to the training data on which a purely physical aging model was previously parameterized. The training can be carried out in accordance with known, in particular gradient-based, methods using a loss function in accordance with a least squares method. The model parameters of the model parameter vector β for the data-based correction term are adjusted in order to minimize the loss.

The data-based correction model can include a probabilistic regression model, in particular a Gaussian process model, with the training being carried out using a gradient-based method by maximizing a log-likelihood.

Thus, in the case of a Gaussian process, the maximization of a log-likelihood can also be used instead of the above-mentioned minimization of the loss function, and thus a probabilistic modeling can be taken into account.

The parameter vector α with the physical parameters α and the model parameters of the model parameter vector β can be parameterized together in the training process with the aid of the training data sets.

• - Parameter des Modellterms und Modellparameter des datenbasierten Korrekturmodells gleichzeitig mithilfe eines gradientenbasierten Optimierungsverfahren optimiert werden, oder
• - wobei in einem ersten Schritt Parameter des Modellterms angepasst und in einem zweiten Schritt Modellparameter des datenbasierten Korrekturmodells der mindestens einen Differentialgleichung bei fixierten Parametern des Modellterms trainiert werden.

The training of the data-based state of health model can be performed by

• - Parameters of the model term and model parameters of the data-based correction model are optimized simultaneously using a gradient-based optimization method, or
• - wherein in a first step parameters of the model term are adjusted and in a second step model parameters of the data-based correction model of the at least one differential equation are trained with fixed parameters of the model term.

Thus, the training method can provide for simultaneous training of the parameters α and β. Alternatively, the training method can also be carried out in two steps, after which the parameter vector α is first parameterized with the physical parameters α based on a set of training data sets and then the physical parameters α are fixed and only the model parameter vector β is trained with the same or a further training set of training data .

Provision can be made for the at least one differential equation to have a stochastic term.

Thus, the model equation of the neural differential equation can be extended with a stochastic term to set up a stochastic differential equation system. This allows for greater robustness since the correction terms can also account for errors due to noise.

Furthermore, the method can be carried out in a device-external central unit, with the operating variable profile being determined in the technical device and being transmitted to the device-external central unit.

• - Bereitstellen eines Betriebsgrößenverlaufs einer oder mehrerer Betriebsgrößen;
• - Bereitstellen eines trainierten datenbasierten Alterungszustandsmodells, das auf einem neuronalen Differentialgleichungssystem basiert und trainiert ist, um einen Betriebsgrößenverlauf einem Alterungszustand zuzuordnen;
• - Bestimmen des modellierten Alterungszustands durch Auswerten des Alterungszustandsmodells mit dem Betriebsgrößenverlauf.

According to a further aspect, a computer-implemented method for determining a modeled aging state of an electrical energy store with at least one electrochemical unit, in particular a battery cell, in a technical device at a specific point in time is provided, with the following steps:

• - providing a performance of one or more performance variables;
• - Providing a trained data-based aging model based on a neural differential equation system and is trained to assign an operating variable profile to an aging state;
• - Determination of the modeled aging status by evaluating the aging status model with the course of operating variables.

The neural differential equation system can be designed to obtain internal state variables, with the modeled aging state resulting at each time step as a functional combination of the internal state variables, in particular from a linear combination of the internal state variables.

In particular, the modeled state of health can be corrected using a further data-based correction model that is trained on a residue of the state of health based on training data sets.

The energy store can be used to operate a device such as a motor vehicle, a pedelec, an aircraft, in particular a drone, a machine tool, a consumer electronics device such as a mobile phone, an autonomous robot and/or a household appliance.

According to a further aspect, a device for carrying out the above method is provided.

character list

• 1 eine schematische Darstellung eines Systems zur Bereitstellung von fahrer- und fahrzeugindividuellen Betriebsgrößen zur Bestimmung eines Alterungszustands einer Fahrzeugbatterie in einer Zentraleinheit;
• 2 eine schematische Darstellung eines funktionalen Aufbaus eines datenbasierten Alterungszustandsmodells; und
• 3 ein Flussdiagramm zur Veranschaulichung eines Verfahrens zum Erstellen des datenbasierten Alterungszustandsmodells.

Embodiments are explained in more detail below with reference to the accompanying drawings. Show it:

• 1 a schematic representation of a system for providing driver and vehicle-specific operating variables for determining an aging condition of a vehicle battery in a central unit;
• 2 a schematic representation of a functional structure of a data-based aging model; and
• 3 a flowchart to illustrate a method for creating the data-based aging model.

Description of Embodiments

The method according to the invention is described below using vehicle batteries as electrical energy stores in a large number of motor vehicles as devices of the same type. A data-based state of health model for the respective vehicle battery can be implemented in a control unit in motor vehicles. As described below, the aging state model can be continuously updated or retrained in a vehicle-external central unit based on operating variables of the vehicle batteries from the vehicle fleet. The aging state model is operated in the central unit and used for aging calculation and aging prediction. Furthermore, the overall aging model, but specifically the data-based aging model, can be used in the vehicle for efficient data processing, so that the data transfer to the cloud is carried out in an optimized manner.

The above example is representative of a large number of stationary or mobile devices with a network-independent energy supply, such as vehicles (electric vehicles, pedelecs, etc.), systems, machine tools, household appliances, IOT devices and the like, which have a corresponding communication connection (e.g. LAN, Internet) with a device-external central unit (cloud).

1 shows a system 1 for collecting fleet data in a central unit 2 for the creation and operation as well as for the evaluation of an aging model. The aging state model is used to determine an aging state of an electrical energy store such. B. a vehicle battery or a fuel cell in a motor vehicle. 1 shows a vehicle fleet 3 with several motor vehicles 4.

One of the motor vehicles 4 is in 1 shown in more detail. The motor vehicles 4 each have a vehicle battery 41 as a rechargeable electrical energy store, an electric drive motor 42 and a control unit 43 . The control unit 43 is connected to a communication module 44 which is suitable for transmitting data between the respective motor vehicle 4 and a central unit 2 (a so-called cloud).

The motor vehicles 4 send the operating variables F to the central unit 2, which indicate at least variables which influence the aging state of the vehicle battery 41. In the case of a vehicle battery, the operating variables F can include time series of a battery current, a battery voltage, a battery temperature and a state of charge (SOC: State of Charge), both at pack, module and/or cell level. The operating variables F are recorded in a fast time frame from 1 Hz to 100 Hz and can be transmitted regularly to the central unit 2 in uncompressed and/or compressed form.

Furthermore, the time series using compression algorithms for the purpose Minimizing the data traffic to the central unit 2 are transmitted in blocks to the central unit 2 at intervals of several hours up to several days.

The central unit 2 has a data processing unit 21, in which the method described below can be executed, and a database 22 for storing data points, model parameters, states and the like.

A data-based aging state model is implemented in the central unit 2, which is implemented as a neural differential equation system.

The state of health model can be regularly, i. H. e.g. g. after the respective evaluation period has expired, in order to determine the current aging status of the relevant vehicle battery 41 of the assigned vehicle fleet based on the temporal progression of the operating variables (in each case since the respective vehicle battery was commissioned or starting from a point in time with known internal system states). . In other words, it is possible to determine an aging state of the relevant vehicle battery 41 based on the progression of the operating variables F (operating variable progression) of one of the vehicle batteries 41 of the motor vehicles 4 of the associated vehicle fleet 3 .

The state of health (SOH: State of Health) is the key variable for indicating a remaining battery capacity or remaining battery charge. The aging state is a measure of the aging of the vehicle battery or a battery module or a battery cell and can be specified as a capacity retention rate (SOH-C) or as an increase in internal resistance (SOH-R). The capacity retention rate SOH-C is given as the ratio of the measured instantaneous capacity to an initial capacity of the fully charged battery and decreases as the battery ages. The relative change in internal resistance SOH-R increases as the battery ages.

As for the time series, compression algorithms can be used for the purpose of minimizing the data traffic to the central unit 2 for the transmission. Furthermore, an event-based transmission can take place, so that the data transfer is triggered and takes place when z. B. a stable or known WiFi network connection has been identified.

2 shows a schematic representation of the functional structure of the neural differential equation system 10, which is solved at each time step with the operating variables F(t) provided.

The differential equation system 10 implements a plurality of interdependent differential equations and includes a model block 11 for providing the right-hand side model terms for each of the differential equations. The model terms correspond to the physical model equations f1(α,x,F,t), f2(α,x,F,t), f3(α,x,F,t), ... which the physical aging model with respect to an internal embody the system state x and the values of the operating variables applicable at time t and are parameterized with the parameter vector α based on expert knowledge.

Furthermore, a correction block 12 is provided, which provides a data-based correction term g(β, x, F, t). In each time step the result of the model block and the correction block 12 is added in a summation block 13 and the result represents the state vector x(t) for the current time step. This is delayed by a delay block 14 in the model block 11 and the correction block 12 as x( t-1) provided.

The internal system states x can correspond to the following state variables for a portable battery: SEI layer thickness x _SEI , amount of cyclable lithium x _cLi , amount of cyclable solvent x _csol1,2 , amount of cathode material x _Li , loss of active material x _AML -, and amount to x _AME+

$${\dot{x}}_{cLi}=ƒ2\left(a,x,F,t\right)+g2\left(\beta ,x,F,t\right)$$

$${\dot{x}}_{csol\mathrm{1,2}}=ƒ3\left(a,x,F,t\right)+g3\left(\beta ,x,F,t\right)$$

$${\dot{x}}_{Li}=ƒ4\left(a,x,F,t\right)+g4\left(\beta ,x,F,t\right)$$

$${\dot{x}}_{AML-}=ƒ5\left(a,x,F,t\right)+g5\left(\beta ,x,F,t\right)$$

$${\dot{x}}_{AML+}=ƒ\text{6}\left(a,x,F,t\right)+g6\left(\beta ,x,F,t\right)$$

$$\text{mit }x={\left(x_{SEI}{x}_{cLi},x_{csol\mathrm{1,2}},x_{Li},x_{AML-},x_{AML+}\right)}^T$$

In the exemplary case of these six internal state variables, the differential equation system has the following structure, with the functions g1 to g6 each corresponding to the data-based correction term, which can be formed by a Gaussian process model or a neural network: $${\dot{x}}_{SEI}=ƒ1\left(a,x,f,t\right)+G1\left(\beta ,x,f,t\right)$$

$${\dot{x}}_{cLi}=ƒ2\left(a,x,f,t\right)+G2\left(\beta ,x,f,t\right)$$

$${\dot{x}}_{csOl1.2}=ƒ3\left(a,x,f,t\right)+G3\left(\beta ,x,f,t\right)$$

$${\dot{x}}_{Li}=ƒ4\left(a,x,f,t\right)+G4\left(\beta ,x,f,t\right)$$

$${\dot{x}}_{AML-}=ƒ5\left(a,x,f,t\right)+G5\left(\beta ,x,f,t\right)$$

$${\dot{x}}_{AML+}=ƒ\text{6}\left(a,x,f,t\right)+G6\left(\beta ,x,f,t\right)$$

$$\text{with}x={\left(x_{SEI}{x}_{cLi},x_{csOl1.2},x_{Li},x_{AML-},x_{AML+}\right)}^T$$

The state vector x(t) is now evaluated in an aging state block 15 in order to obtain an aging state SOH.

The aging state SOH results for each time step as a functional combination of the internal state variables, in particular from a linear combination of the internal state variables: SOH=k(γ,x).

The neural differential equation system 10 is solved numerically, it being possible to use a difference in the values of the state variables at successive time steps as the differential term of the internal state variables.

In addition, the state of health block 15 can be supplied with a further correction variable, which results from the evaluation of a further data-based correction model in a further correction block 16 . The further correction model makes it possible to correct the residue of the aging state that results during or after the evaluation in the aging state block 15 . The further data-based correction model can be trained after the training of the neural differential equation system has been completed, based on the identical or further training data sets. The further data-based correction model can be designed as a probabilistic regression model, in particular as a Gaussian process model or as a Bayesian neural network.

3 shows a flowchart to illustrate a method for training the neural differential equation system 10. For this purpose, in step S1 first training data sets are provided, which each map a time series of an operating variable profile of one or more operating variables to an aging state determined empirically, for example. The aging status can be determined, for example, using a basic model based on Coulomb counting or an evaluation of an OCV characteristic.

So e.g. B. an SOH-C measurement by Coulomb counting or by forming a time current integral during the charging process, which is divided by the swing in the state of charge between the beginning and end of the relevant charging and / or discharging phase. In this case, the calibration on the open-circuit voltage characteristic (OCV characteristic) advantageously takes place in idle phases in order to also calculate the state of charge profile in the central unit 2 . A sufficiently reliable indication of the aging state for use as a label can be obtained, for example, if the vehicle battery is brought from a fully discharged state of charge to a fully charged state during a charging process from a defined relaxed state under reproducible load and environmental conditions. The maximum charge thus detected can be related to an initial maximum charge capacity (nominal battery capacity) of the vehicle battery.

wobei SOH_mess(t_n) den als Label ermittelten Alterungszuständen zum Zeitpunkt t_n entspricht.In step S2, an initial parameterization of the neural differential equation system 10 is subsequently carried out, with the physical parameters α of the model term and the model parameters β of the data-based correction term being able to be trained together. The training is carried out using a loss function that provides a loss as least square by minimizing the physical parameters of the parameter vector and the model parameters β. $$\left(a*,\beta *\right)=\underset{a,\beta }{\text{argmin}}{\displaystyle \sum _{i-1}^n{\left(SO{H}_{mess}\left(t_n\right)-SOH\left(t_n,a,\beta \right)\right)}^2}$$

where SOH _mess (t _n ) corresponds to the aging states determined as a label at time t _n .

If the correction term is formed by a Gaussian process model, a maximization of the log-likelihood can also be used instead of the above minimization, so that a probabilistic modeling can be taken into account.

If the physical parameters have converged to a sufficient degree, determined by a predetermined convergence criterion, then in step S3, subsequent training takes place with a further training set of training data sets in order to train the model parameter vector β for the correction terms g1, g2... of the differential equation system 10. The training is based on a gradient-based training method, which can be applied since the differential equations including the data-based correction terms are differentiable.

zwischen den Iterationen weniger als eine vom Anwender vorgegebene zweite Schwelle Δ2 verändert oder beides.Known convergence criteria for parameter estimation in ordinary differential equations can be used. So e.g. B. the convergence can be determined in particular if the parameters α and β between the iterations change less than a first threshold Δ1 specified by the user or if the residual ${{\displaystyle {\sum }_{i=1}^n\left(SO{H}_{meas}\left(t,n\right)-SOH\left(t,n,a,\beta \right)\right)}}^2$

changed less than a second threshold Δ2 specified by the user or both between the iterations.

In a further embodiment, a stochastic term can also be applied to the model terms and correction terms of the differential equations.

In addition, the further data-based correction model 16 can be trained based on the training data sets provided. For this purpose, the residue of the aging state, which results from evaluating the state vector x(t), is trained as a correction variable for the training data points determined by the training data sets. The further correction variable is used in particular for the additive loading of the modeled aging state SOH.

Claims (14)

Computer-implemented method for providing a state of health model for determining a modeled state of health (SOH) of an electrical energy store (41) with at least one electrochemical unit, in particular a battery cell, in a technical device, with the following steps: - Providing (S1) of training data sets, each related to a specific point in time, assigns an empirically determined aging state to an operating variable profile (F(t)) of one or more operating variables (F); - Providing a data-based aging model, which is designed with a neural differential equation system (10) to model internal electrochemical states of the energy store (41), wherein the internal states (x(t)) can be mapped to an aging state, wherein at least one of the differential equations the differential equation system (10) has a sum of a deterministic model term and a data-based correction term, which is formed by a data-based correction model, - Training (S2, S3) of the data-based aging model based on the training data sets, so that model parameters of the data-based correction model are determined. procedure after claim 1 , wherein the training of the data-based aging model is carried out by - parameters (α) of the model term and model parameters (β) of the data-based correction model being optimized simultaneously using a gradient-based optimization method, or - wherein in a first step parameters (α) of the model term are adjusted and in in a second step, model parameters (β) of the data-based correction model of the at least one differential equation are trained with fixed parameters (α) of the model term. Procedure according to one of Claims 1 until 2 , wherein the data-based correction model of the neural differential equation system (10) comprises a deep neural network, the training being carried out using a gradient-based method by minimizing a predetermined loss function. Procedure according to one of Claims 1 until 3 , wherein the data-based correction model comprises a probabilistic regression model, in particular a Gaussian process model, the training being carried out using a gradient-based method by maximizing a log-likelihood. Procedure according to one of Claims 1 until 4 , wherein the at least one differential equation has a stochastic term. Procedure according to one of Claims 1 until 5 , The method being carried out in a device-external central unit (2), the operating variable profile being determined in the technical device and being transmitted to the device-external central unit (2). Computer-implemented method for determining a modeled state of health (SOH) of an electrical energy store (41) with at least one electrochemical unit, in particular a battery cell, in a technical device at a specific point in time, with the following steps: - Providing an operating variable curve (F(t)) of one or more operating variables; - Providing a trained data-based aging state model, which is based on a neural differential equation system and is trained to assign an operating variable profile (F(t)) to an aging state; - Determining or predicting the modeled aging status by evaluating the aging status model with the operating variable progression in time steps. procedure after claim 7 , wherein the neural differential equation system (10) is designed to obtain internal state variables (x(t)), the modeled aging state being at each time step as a functional combination of the internal state variables, in particular from a linear combination of the internal state variables (x( t)), results. procedure after claim 8 , wherein the modeled aging state is corrected using a further data-based correction model that is trained on a residual of the aging state based on training data sets. Procedure according to one of Claims 1 until 9 , wherein the energy store (41) for operating a device such as a motor vehicle, a pedelec, an aircraft, in particular a drone, a machine tool, a consumer electronics device such as a mobile phone, an autonomous robot and/or a household appliance. Procedure according to one of Claims 1 until 10 , wherein the aging state model is designed to indicate or predict a modeled aging state (SOH) as a capacity-related aging state, as an aging state related to internal resistance change or in the form of an electrochemical parameter or internal state. Device for carrying out one of the methods according to one of Claims 1 until 11 . Computer program product comprising instructions which, when the program is executed, cause at least one data processing device to carry out the steps of the method according to one of the Claims 1 until 11 to execute. Machine-readable storage medium, comprising instructions which, when executed by at least one data processing device, cause the latter to carry out the steps of the method according to one of Claims 1 until 11 to execute.

Images