Method and device for the initial provision of an aging state model for energy storage systems based on active learning algorithms and time-discounted information evaluation

Abstract

The invention relates to a method for initially providing an at least partially data-based aging state model (4) for an electrical energy storage device (3), comprising the following steps: - Providing (S1) a number of energy storage devices (3) on a test bench (1) for measurement depending on a respective load profile, wherein the load profiles are different and characterize a time course of at least one load-bearing operating parameter for the energy storage device (3); - Operating (S2) the number of energy storage devices (3) with their respective assigned load profiles and recording operating parameter profiles over time; - At each given evaluation time, determine (S4) an aging state of a subset of the energy storage devices (3) as a label and generate a training data set with the operating parameter profiles and the determined label for each energy storage device (3) of the subset of energy storage devices (3); - Selecting (S6) the subset of energy storage devices (3) with their respective assigned load profiles, depending on an information measure from the respective energy storage devices (Info J ) of the subset of energy storage (3), which has a predictive covariance (Σ(x) t |x1, ... , x (t-1))) of the data-based aging state model is determined at at least one future point in time, wherein the information measure is determined depending on a standard deviation, a variance or a Shannon entropy of a given load pattern model, which is used to determine a future time course of the at least one operational variable in order to determine the predictive covariance at the future point in time.

Description

Technical area

[0001] The invention relates to methods and devices for initially providing an at least partially data-based aging state model for electrical energy storage devices of the same type. Technical background

[0002] The energy supply for the operation of mains-independent electrical devices and machines, such as electrically powered motor vehicles, is usually provided by electrical energy storage devices, such as device batteries or vehicle batteries.

[0003] Electrical energy storage devices degrade over their lifetime and depending on their load and use. This so-called aging leads to a continuously decreasing maximum power or storage capacity. The aging state corresponds to a measure of the aging of energy storage devices. For portable batteries used as electrical energy storage devices, according to convention, a new portable battery can have an aging state (in terms of its capacity, SOH-C) of 100%, which decreases noticeably over the course of its service life.

[0004] A measure of the aging of an electrical energy storage device (change in the aging state over time) depends on an individual load on the energy storage device, i.e. in the case of vehicle batteries of motor vehicles, on the usage behavior of a driver, external environmental conditions and the vehicle battery type.

[0005] In order to monitor the aging conditions of electrical energy storage devices in a large number of devices, operating variable data can be continuously recorded and transmitted in blocks as operating variable curves to a central unit external to the device.

[0006] For the model-based determination of the aging state of an electrical energy storage device based on operating parameter data, it is necessary to provide an initial aging state model. For this purpose, an initial measurement is planned for a specific number of energy storage devices, e.g., in a laboratory or test bench, in order to generate training data for an aging state model to be provided. For this purpose, the energy storage devices are operated in different ways, whereby, depending on the type of energy storage device, energy must be added or removed in order to simulate the operating cycles of the energy storage devices. The energy required for this is considerable and scales with the number of energy storage devices to be initially measured. The initial measurement can also be considerable, especially if sufficient training data still needs to be collected for aged energy storage devices. Disclosure of the invention

[0007] According to the invention, a method for initially providing an at least partially data-based aging state model for an energy storage type according to claim 1 and a corresponding device according to the independent claim are provided.

[0008] Further embodiments are specified in the dependent claims.

[0009] According to a first aspect, a method is provided for initially providing an at least partially data-based aging state model for an electrical energy storage device, comprising the following steps: - Providing a number of energy storage devices on a test bench for measurement depending on a respective assigned load profile, wherein the load profiles are different and characterise a temporal course of at least one loading operating variable for the energy storage device; - Operating the number of energy storage units with the respective assigned load profile and recording temporal operating variables; - At a given evaluation time, determining an aging state of a subset of the energy storage devices as a label and generating a training data set with the operating variable curves and the determined label for each energy storage device of the subset of energy storage devices; - Selecting the subset of energy storage devices with the respectively assigned load profile depending on an information measure for the subset of energy storage devices, which is determined with a predictive covariance of the data-based aging state model at at least one future point in time, wherein the information measure is determined depending on a standard deviation, a variance or a Shannon entropy of a predetermined load pattern model, which is used to determine a future temporal course of the at least one operating variable in order to determine the predictive covariance at the future point in time.

[0010] In particular, the subset of energy storage devices that yields the maximum information measure can be selected based on the associated information measure. The information measure indicates the information gain that can be obtained during further measurement of the subset of energy storage devices for the aging state model trained or parameterized with the resulting training data sets.

[0011] The state of health (SOH) is the key parameter for specifying the remaining battery capacity or remaining battery charge in portable batteries used as energy storage. The state of health represents a measure of the aging of the portable battery. In the case of a portable battery, battery module, or battery cell, the state of health can be specified as the capacity retention rate (SOH-C). The capacity retention rate SOH-C, i.e. the capacity-related state of health, is specified as the ratio of the measured instantaneous capacity to the initial capacity of the fully charged battery and decreases with increasing aging. Alternatively, the state of health can be specified as an increase in the internal resistance (SOH-R) or an impedance with respect to an internal resistance or an impedance at the beginning of the service life of the portable battery.The relative change in internal resistance SOH-R increases with increasing battery age.

[0012] The aging state of an electrical energy storage device is not typically measured directly. This would require a series of sensors inside the energy storage device, which would make the production of such an energy storage device costly and complex, as well as increasing the space required. Furthermore, practical measurement methods for directly and accurately determining the aging state of energy storage devices are not yet available on the market.

[0013] Precise methods for the indirect or model-based determination of the aging state of energy storage devices are computationally intensive. Therefore, for capacity reasons, the monitoring of energy storage devices in a large number of devices can be carried out in a central unit external to the device. For this purpose, the devices can transmit operating parameter profiles of the energy storage devices to the central unit, whereby a current electrochemical state and / or aging state is determined in the central unit. Depending on the model used, time series of operating parameters are recorded continuously or in time intervals as operating parameter profiles, such as battery current, battery temperature, state of charge and / or battery voltage for a device battery as an energy storage device. These are transmitted to the central unit in blocks and, if necessary, in compressed form.There, the operating variables are evaluated so that, based on one or more aging state models, a device-specific aging state and, if necessary, other variables can be calculated / determined. Furthermore, the operating variables from the numerous energy storage devices can be evaluated using statistical methods to improve the applied aging state models, thus significantly improving the determination and prediction of the aging state of the energy storage devices.

[0014] Due to electrochemical effects during the operation of an energy storage device that are difficult to describe physically, the use of a data-based model as or in combination with a physically based aging model has proven to be a useful aging state model.

[0015] A possible aging state model can be provided in the form of a hybrid aging state model, which corresponds to a combination of a physical aging model based on the evaluation of differential equations with a data-based correction model. In a hybrid model, a physical aging state can be determined using a physical or electrochemical aging model, and this can be subjected to a correction value resulting from the data-based correction model, in particular by addition or multiplication. The physical aging model can be based on electrochemical model equations that characterize the electrochemical states of a nonlinear differential equation system with regard to aging reactions, continuously calculate them according to a time integration method, and map them to the physical aging state for output, as SOH-C and / or SOH-R.The calculations can typically be performed in the central unit (cloud) at intervals of specified evaluation periods, e.g. once a week.

[0016] Furthermore, the correction model of the hybrid data-based aging state model can be designed with a probabilistic or artificial intelligence-based probabilistic regression model, in particular a Gaussian process model, and can be trained to correct the aging state obtained by the physical aging model. For this purpose, a data-based correction model can be provided to correct the capacitance-related aging state and, if necessary, another data-based correction model can be provided to correct the resistance-change-related aging state. Possible alternatives to the Gaussian process are other supervised learning methods, such as those based on a random forest model, an AdaBoost model, a support vector machine, or a Bayesian neural network.

[0017] When commissioning a new type of energy storage system, it is necessary to provide an aging state model to monitor the aging process. Since precise knowledge of the aging behavior of energy storage systems is generally not available at commissioning, it is necessary to initially specify an aging state model that can at least approximately indicate an aging state based on the provided operating parameter profiles. The initial specification of the aging state model therefore requires initial training of the data-based model.

[0018] Training data for such initial training has so far usually been determined in the laboratory or on a test bench and comprises a predetermined number of, in particular, randomly selected energy storage devices, each of which is operated with differing load profiles. These load profiles comprise cyclic current flows with individually predetermined durations and profiles, temperature profiles and the like, or are converted into these. Particularly in the case of device batteries as energy storage devices, these load profiles can comprise different charging and discharging current profiles at different temperatures and can be specified and characterized in a compressed manner, in particular using histogram data. These load profiles are converted into realistic time series of current supplies and current discharges, if necessary.also battery temperatures, which realize the load profile and curves of other corresponding operating variables, such as a battery voltage, a battery temperature and a state of charge, are recorded and stored together as operating variable curves.

[0019] At specified times, aging state measurements are taken using suitable additional models or measurement methods to determine a label for the operating parameter trends. This data is used to create training data sets that can be used to parameterize the aging state model and / or to train the data-based model in the case of a hybrid or purely data-based aging state model. Various aging state models and methods can be used to determine the label.

[0020] A possible model or method for determining the aging state of a portable battery used as an energy storage device could be a basic model according to which an SOH-C measurement is carried out by Coulomb counting or by forming a temporal current integral during the charging process, which is divided by the change in the state of charge between the beginning and the end of the respective charging and / or discharging phase. Advantageously, calibration is carried out using the open-circuit voltage characteristic during rest phases in order to include the state of charge curve in the calculation in the central unit. A sufficiently reliable indication of the aging state can be obtained, for example, if the portable battery is brought from a defined relaxed state from a fully discharged state of charge to a fully charged state during a charging process under reproducible load and ambient conditions.The maximum charge thus recorded can be related to the initial maximum charge capacity of the device's battery. Resistance-related aging states (SOH-R values) can also be calculated using voltage changes relative to a current change. These are usually related to a defined time interval, defined ambient conditions, and the energy flow direction of the system.

[0021] According to the above method, the aging states of a subset of the energy storage devices can now be determined as labels at predetermined times, and a training data set containing the operating variables and the determined label can be generated for each energy storage device in the subset of energy storage devices. The selection of the subset of energy storage devices with the associated load profile is carried out depending on the information metrics for all energy storage devices.

[0022] A method is thus proposed in which the initially provided aging state model is further trained at specified times and at the specified times only those energy storage devices (a specified number J of selected energy storage devices) are measured for label generation that have a high contribution to the further development or refinement of the aging state model. For this purpose, using a probabilistic regression model as a data-based model, the predictive covariance is determined at each evaluation time t. ∑(SOHj(t1),SOHj(t2),...,SOHj(tn)) of the data-based model, which results from the probabilistic regression model (Gaussian process model) provided in the aging state model. SOH j corresponds to the aging states of a specific energy storage device j of the selected energy storage devices at evaluation times, which results from future evaluation times t1...tn The predictive covariance Σ in Gaussian process models as probabilistic regression models does not depend on previously unseen aging conditions, but only on the operating characteristics x(t) = [m(x(t)), z(x(t)), Phys[x(t)]) as input variables of the Gaussian process model and can therefore be described as ∑(x¯j(t1),x¯j(t2),...,x¯j(tn)) where x(t) represents one or more operating variables, m(x(t)) histogram data as a possible indication of a load profile, z(x(t)) multidimensional electrochemical states of the physical aging model, such as the SEI thickness, the amount of cyclable lithium, the amount of active material, electrochemical concentrations and the like, and Phys[x(t)] corresponds to the physical (modeled) aging state at a specific evaluation time t. m(x(t)) as histogram data and z(x(t)) as electrochemical states represent operating characteristics for the evaluation of the Gaussian process model at a specific evaluation time t.

[0023] The operating characteristics x(t) can also include only a part of the quantities m(x(t)), z(x(t)), Phys[x(t)] and also additionally one or more values ​​of one or more of the original operating quantities x(t).

[0024] It can be provided that the information measure is formed using a sum of predictive covariances at future, particularly consecutive, evaluation times at which training data sets for training the aging state model have been determined. In particular, the information measure can be determined from the input vectors of the subset of several of the energy storage devices for the evaluation times of an entire measurement period.

[0025] The predictive covariances at future time points are obtained by evaluating the prediction model in prediction mode with load predictions.

[0026] For this purpose, the covariance function or kernel function of the Gaussian process can be used to derive the uncertainty or confidence of the Gaussian process as a function of the input variables. These predictive covariances represent a relevant information measure used in the procedure.

[0027] The total information measure for all selected energy storage units at a given evaluation point in time is: h(x)=∑t=t1Th(x¯1,...,x¯(t−1)) where x t ={x 1.t , ..., x J,t}, where J is the number of selected energy storage devices. The predictive covariance Σ is a matrix to which information measures, such as the determinant or the maximum eigenvalue, can be applied.

[0028] The respective entropy h(x t |x1, ..., x (t-1) ) can be calculated as predictive covariance h(x¯1,...,x¯(t−1))=12ln(2πe)Jdet(∑(x¯t|x¯1,...,x¯(t−1))) so that the information measure for the entire energy storage h(x¯)=∑t=t1T12ln(2πe)Jdet(∑(x¯t|x¯1,..,x¯(t−1))) amounts.

[0029] The conditional predictive covariances Σ can be calculated without knowledge of the model output of the aging state, since they depend only on the characteristics x(t) from predicted future operating variables. The predicted future operating variables are determined using a suitable load prediction model.

[0030] The operating variables are always known and deterministic in the laboratory. In reality, they can be predicted using load models, such as hidden Markov models, especially current and temperature as time series signals. The hidden Markov model characterizes a type of energy storage usage by defining operating states and the probabilities of transitions between the operating states.

[0031] The load-dependent prediction of current dynamics, such as charging and discharging (charging and discharging current profiles), can be performed probabilistically, preferably using NARX Gaussian processes or Deep Bayesian networks, or alternatively, using model-based approaches, particularly deep learning-based prediction models such as LSTMs, transformers, or alternative neural networks. The load-free (parking) and load-dependent (charging / discharging) time series signals are then reassembled using the load model and combined into time series signal forecasts (predicted operating variables). The uncertainties in the load prediction can be propagated in the aging prediction. This is done, for example, as follows: - using Monte Carlo methods, where a sufficiently large number of possible load predictions are sampled from the uncertainty in the load prediction, - Moment matching, used for propagation. The moment matching method is an alternative to sampling and is described, for example, in A. Girard, "Approximate Methods for Propagation of Uncertainty with Gaussian Process Models," Doctoral thesis, University of Glasgow, October 2004. It assumes that the probabilistic load prediction model outputs a normal distribution with mean and variance. This results in a distribution with an expected value. This can be calculated using the moment matching method. - or by integrating over the sampled distribution of the load prediction.

[0032] The load prediction model can be probabilistic, which is why the uncertainties σ L (x t) of the load prediction model results directly from the load sampling from the Hidden Markov Model, whereby the multiple load trajectories are evaluated with regard to their distribution function or their expected value and standard deviation.

[0033] The company size trends can be specified deterministically or can be modeled probabilistically.

[0034] Uncertainties in the prediction of farm size trends, from which the predictive farm characteristics m(x(t)) and z(x(t)) are determined at an assumed future evaluation time, can be taken into account for each evaluation time t by h˜_t=∫h(x¯t)pL(x¯t)dx where for the considered evaluation time t a probability distribution p L (x t ) of a load profile with which the associated energy storage is operated. This probability distribution, i.e. the probability of a feature xt The observed behavior can be derived from a probabilistic dynamic model and transferred to the information gain. Such a dynamic model generates artificial operating variables based on load profiles. The probability distribution indicates the probability that the generated operating variable corresponds to the actual operating variable.

[0035] It can be provided that, in the case of unknown operating variables of one or more of the energy storage devices, an artificial operating variable is generated and, before the summation, the expected value is formed via the predictive covariances with respect to a probability distribution that describes the probability that the artificial operating variable corresponds to the actual operating variable.

[0036] According to one embodiment, the predictive covariances can each be weighted with a weighting factor / discounting factor before summing, which weights earlier predictive covariances higher than later ones.

[0037] In addition to the currently planned load profiles, Active Learning can be used to generate additional load profiles, especially those that increase the information content. This allows Active Learning to actively intervene in load profiles to achieve the best possible operating points for the energy storage systems. Based on the probabilistic dynamic model, the load profile leads to operating variables with a corresponding probability distribution.

[0038] In particular, the weighting factor can be calculated using an index of a time step for consecutive evaluation points as a function a(σ L (x t )) of the standard deviation σ L (x t) or the Shannon entropy of the load prediction model for a discount factor can be determined. InfoJ=∑t=t1Ta(σL(x¯t))(h(x¯t|x¯1,...,x¯t−1))

[0039] This results in a discounting with the weighting factor a(σ L (x t )) is taken into account, with deterministically specified company size curves (x t |x1, ... , x t-1 ) or in a probabilistic model for generating artificial farm size trends InfoJ=∑t=t1Ta(σL(x¯t))∫12ln(2πe)ddet(∑(x¯t|(x¯t|x¯1,...,x¯t−1)))pL(x¯t)dx where a corresponds to a monotonically decreasing function which σ L (x t ) for small values ​​to a value close to “1” and for large values ​​to a value close to 0, so that there is strong discounting at high uncertainty and only little discounting at low uncertainty.

[0040] Possible functions are a) a(σ L (x t)) = 1- erf(c * σ L (x t )) with the Gaussian error function erf(x) b) a(σ L (x t )) = 1-(pi / 2)^-1*arctan(c*σ L (x t )) c) a(σ L (x t )) = 1-tanh(c*σ L (x t ))where the specified factor c describes how strong the discounting effect should be.

[0041] For this purpose, the company sizes must be predicted over a given prediction horizon of, for example, 3 months.

[0042] One reason for discounting is, in particular, that the load prediction has an uncertainty, e.g. if the load profiles and / or the operating size curves are still being adjusted due to active learning or because the aging state model has not yet been sufficiently trained.

[0043] In the sense of a variant of an active learning method, a number J (subset) of energy storage devices B can now be selected for measurement from the total number of energy storage devices B all at the current evaluation time, for which the respective information level is maximum. B∗=argmax{B⊂Ball,|B|=J}InfoJ(B)

[0044] The calculation can be performed using a greedy method. First, the most informative battery (with the highest information measure) is determined. Then, the battery that most complements this battery (with the highest information measure of the remaining batteries) is determined. Then, the battery that most complements the selected batteries, and so on.

[0045] The number J of energy storage devices to be selected can be determined, for example, by comparing the reduction in predictive variance on a validation dataset. Then, one can compare the reduction in predictive variance on the validation dataset and the number of energy storage devices, and, using a predefined threshold, decide which degree of reduction is desired for further measurement of an energy storage device.

[0046] According to the procedure, specific measurements of the aging state are now taken for the selected energy storage devices and not for the non-selected energy storage devices. As soon as sufficient new information is available, the data-based aging state model is retrained. This, together with the associated operating parameter profiles, produces a training data set for further developing the initial aging state model.

[0047] The training can be supplemented by automated hyperparameter tuning, e.g., using gradient-based methods or black-box methods such as Bayesian optimization.

[0048] The procedure described above can be terminated if a certain accuracy requirement is met, such as < 1.5% SOHC on a relevant, previously provided validation dataset, and other robustness requirements, such as evaluated via cross-validation, are met.

[0049] It can be provided that at each evaluation time, only the energy storage units of the subset of energy storage units are measured to determine an aging state as a label and the remaining energy storage units continue to be operated according to the load profile.

[0050] Furthermore, the data-based aging state model can be designed with a probabilistic data-based model, in particular a Gaussian process model, wherein for one of the energy storage devices, an input vector of the data-based model, which comprises at least one operating variable profile and / or at least one operating characteristic from the at least one operating variable profile, an internal state of the energy storage device and / or a physically modeled aging state, is mapped to the aging state of the energy storage device in question to be modeled or a correction variable for correcting a physically modeled aging state of the energy storage device in question, wherein the information measure for the subset of the energy storage devices is determined with a determinant of the predictive covariance to the respective energy storage device.

[0051] Furthermore, at each evaluation time, only the energy storage units of the subset of energy storage units can be measured as a label to determine an aging state, and the remaining energy storage units can continue to operate according to the assigned load profile.

[0052] Alternatively, at each evaluation time, only the energy storage devices of the subset of energy storage devices can be measured as a label to determine an aging state and continue to operate according to the load profile, while the remaining energy storage devices are removed from the test bench.

[0053] It can be provided that the aging state model is trained with the determined training data sets.

[0054] According to a further aspect, an apparatus for carrying out one of the above methods is provided. Brief description of the drawings

[0055] Embodiments are explained in more detail below with reference to the attached drawings. They show: Fig. 1 a schematic representation of a test bench for measuring a plurality of vehicle batteries to create an initial aging state model; Fig. 2 a schematic representation of a hybrid aging state model; Fig. 3 a flowchart illustrating a method for the optimized creation of an initial aging state model; Fig. 4 a representation of an aging process based on the decrease in battery capacities of different vehicle batteries under different load patterns. Description of embodiments

[0056] Fig. Figure 1 shows a schematic arrangement of a test bench with a test bench unit 2 connected to a plurality of vehicle batteries 3 as electrical energy storage devices. The test bench unit 2 controls the vehicle batteries 3 according to a predetermined load pattern that characterizes various current and / or temperature profiles. The current and / or temperature profiles cause different loads and thus different cyclical aging of the vehicle batteries 3, for example, by specifying an ampere-hour throughput, specifications for charging profiles, in particular maximum charging currents depending on the state of charge during charging, maximum charging currents during discharging, maximum and average charging and discharging currents, and corresponding temperature conditions. The load patterns are specified for each of the vehicle batteries 3 and correspond to more or less stressful operating modes of the vehicle batteries 3.Temporal operating variables of the vehicle batteries 3 are continuously recorded and temporarily stored. For vehicle batteries, the operating variables include a battery voltage, a battery current, a battery temperature, and a state of charge.

[0057] The test bench 1 is used to collect data on the aging of the vehicle batteries 3 in order to provide a corresponding initial aging state model 4, which makes it possible to determine the current aging state of the vehicle battery in question with a specified minimum accuracy based on operating parameter profiles recorded during real operation. The aging state model 4 to be created for this purpose can be at least partially data-based and can have a data-based probabilistic regression model.

[0058] The measurement of the multitude of portable batteries 3 on test bench 1 involves determining the aging state of selected vehicle batteries 3 at regular intervals using a suitable method. This, in conjunction with the corresponding operating parameter profiles, which are specified by the load pattern assigned to the respective vehicle battery 3, results in training data sets that can be used to train the data-based / hybrid aging state model 4. The use of test bench 1 and the measurement of the multitude of vehicle batteries 3 incurs costs, which are caused on the one hand by the energy consumption and on the other hand by the occupancy time of the test bench and the other material consumption. These costs are to be reduced when measuring the multitude of vehicle batteries 3 without compromising the quality of the initially trained aging state model 4.

[0059] Fig. Figure 2 schematically shows, by way of example, the functional structure of an embodiment of a data-based aging state model 9 designed in a hybrid manner. The aging state model 9 comprises a physical aging model 5 and a data-based correction model 6.

[0060] Physical aging model 5 is a mathematical model based on nonlinear differential equations. Evaluating physical aging model 5 of aging state model 9 with operating parameter profiles, particularly since the beginning of the vehicle battery's service life, results in an internal state of the system of physical differential equations that corresponds to the physical internal state of the vehicle battery. Since the physical aging model is based on physical and electrochemical principles, the model parameters of the physical aging model are quantities that specify physical properties.

[0061] The time series of the operating variables x(t) of the vehicle battery to be evaluated are therefore directly incorporated into the physical aging state model 5, which is preferably designed as an electrochemical model and models corresponding internal electrochemical states z(t), such as layer thicknesses (e.g. SEI thickness), changes in cyclable lithium due to anode / cathode side reactions, rapid consumption of electrolytes, slow consumption of electrolytes, loss of active material in the anode, loss of active material in the cathode, etc., using nonlinear differential equations in a multidimensional state vector.

[0062] The physical aging model 5 thus corresponds to an electrochemical model of the respective vehicle battery 3. This model determines internal physical battery states z(t) depending on the operating variable curves x(t) in order to provide a physically based aging state SOHph = Phys[x(t)] of dimension of at least one, depending on the above-mentioned electrochemical states z(t), which are mapped linearly or non-linearly to a capacity maintenance rate (SOH-C) and / or an internal resistance increase rate (SOH-R) in order to provide these as aging states (SOH-C and SOH-R).

[0063] However, the model values ​​for the physical aging state SOHph provided by the electrochemical model are inaccurate in certain situations, and therefore, they are intended to be corrected with a correction variable k. The correction variable k is provided by the data-based correction model 6, which is trained using training data sets obtained by test bench 1.

[0064] Preferably, the data-based correction model corresponds to a Gaussian process model GP = GP[m(x(t)), z(x(t)), GP=GP[m(x(t)),z(x(t)),Phys[x(t)]))∼N(μ,Σ), where µ is an mx 1-dimensional vector that specifies the predictive mean at an evaluation time t, and Σ is the mx m-dimensional predictive covariance matrix of the Gaussian process model. The formulas for the mean and variance of the Gaussian process are as follows: μ(x¯)=kTCN−1y ∑(x¯)=c−kTCN−1k where N corresponds to the number of labels, x is m × d dimensional and specifies a set of m new points in the input space, where x(t) = [m(x(t)), z(x(t)), Phys[x(t)]]] and y each correspond to an N × 1 -dimensional vector, and where y each specifies a measured or labeled aging state. k corresponds to an N × m dimensional matrix of kernel scores that specify correlations encoded in the kernel of the Gaussian process between the m new points and the N measured points. C corresponds to an N × N dimensional matrix of kernel scores between the N measured points and c to an m × m matrix with kernel scores between the m new points. T denotes the transpose. See also Bishop, “Pattern Recognition and Machine Learning,” 2006.

[0065] The correction model 6 receives operating characteristics x on the input side t , which are derived from the predicted or given courses of the operating variables x(t0...t n) for future evaluation points in time and may also include one or more of the internal electrochemical states of the differential equation system of the physical model 5. Furthermore, the correction model 6 can receive the physical aging state SOHph obtained from the physical aging model 5 on the input side. The operating characteristics x t of the current evaluation time t are generated in a feature extraction block 8 based on the operating variable curves x(t0...t). Furthermore, the internal states z(x(t)) from the state vector of the electrochemical physical aging model 5 and advantageously the physical aging state SOHph are input to the correction model 6 as operating characteristics x t supplied.

[0066] From the predicted or given operating variable curve x(t0...t n), operating characteristics m(x(t)) can be generated in the central unit 2 for each vehicle fleet 3 or, in other embodiments, already in the respective motor vehicles. These operating characteristics relate to an evaluation period. The evaluation period between two evaluation points in time can range from a few hours (e.g., 6 hours) to several weeks (e.g., one month) for determining the aging state. A typical value for the evaluation period is one week.

[0067] The operating characteristics m(x(t)) can, for example, include characteristics related to the evaluation period and / or accumulated characteristics and / or statistical quantities determined over the entire service life to date. In particular, the operating characteristics can include, for example: electrochemical states, such as SEI layer thickness, changes in cyclable lithium due to anode / cathode side reactions, rapid uptake of electrolyte solvents, slow uptake of electrolyte solvents, lithium deposition, loss of active anode material and loss of active cathode material, information on impedances orthe internal resistances, histogram features such as temperature versus state of charge, charging current versus temperature and discharging current versus temperature, in particular multi-dimensional histogram data relating to the battery temperature distribution versus state of charge, the charging current distribution versus temperature and / or the discharging current distribution versus temperature, the current throughput in ampere-hours, the accumulated total charge (Ah), an average capacity increase during a charging process (in particular for charging processes where the charge increase is above a threshold proportion [e.g. 20% ΔSOC] of the total battery capacity), the charging capacity as well as an extreme value (e.g. maximum) of the differential capacity during a measured charging process with a sufficiently large swing in the state of charge (smoothed curve of dQ / dU: charge change divided by change in battery voltage) or the accumulated mileage.These variables are preferably converted in such a way that they best characterize real-world usage behavior and are normalized in the feature space. The operating characteristics m(x(t)) can be used in whole or in part for the procedure described below.

[0068] To determine a corrected aging state SOH to be output, the outputs SOHph, the correction variable k of the physical aging model 5, and the data-based correction model 6, which is preferably implemented as a Gaussian process model, are combined. In particular, these can be added in a summing block 7 or otherwise multiplied (not shown) to obtain the modeled aging state SOH to be output for a current evaluation period. In the addition case, the confidence of the Gaussian process can also be used as the confidence of the corrected aging value SOH of the hybrid model to be output. The confidence or confidence value of the Gaussian process model thus characterizes the modeling uncertainty of the mapping of operating characteristic points to an aging state.

[0069] The initial training of the hybrid aging state model 9 takes place in test bench 1. For this purpose, training data sets are created that assign the operating parameters of a vehicle battery operating in accordance with a load profile to an empirically or model-based aging state as a label. The goal is for the hybrid aging state model to be able to accurately predict the non-linear, sudden decrease in capacity. This can be achieved through a combination of electrochemically modeled causal chains and machine learning.

[0070] For example, to determine an aging state, a base model can be provided as a label for training the hybrid or data-based aging state model. According to this model, an SOH-C measurement is performed by Coulomb counting or by forming a temporal current integral during the charging process under defined ambient conditions, which is divided by the change in the state of charge between the beginning and end of the respective charging and / or discharging phase. A sufficiently reliable indication of the aging state can be obtained, for example, if the vehicle battery is brought from a defined relaxed state to a fully charged state during a charging process under reproducible load and ambient conditions. The maximum charge thus recorded can be related to an initial maximum charge capacity of the vehicle battery.Resistance-related aging states (SOH-R values) can also be calculated using voltage changes relative to a current change. These are usually related to a defined time interval, defined environmental conditions, and the energy flow direction of the system.

[0071] The determination of an aging state as a label can be carried out in a conventional manner by evaluating the operating variables with an additional aging model under defined load and environmental conditions for label generation, such as constant temperature, constant current, and the like. Other models can be used to determine the aging state. The training of the data-based correction model can be carried out in the conventional way based on the training data sets and the residuals of the modeled aging state.

[0072] In Fig. Figure 3 shows a flowchart describing the process for measuring the plurality of vehicle batteries 3 on the test bench 1. The process is executed in the test bench control unit 2 and results in the provision of an initial data-based aging state model 4, which enables sufficient accuracy in determining the aging state based on operating parameter profiles of vehicle batteries 3 in real operation.

[0073] In step S1, the test bench 1 is equipped with a plurality of brand-new (or in a reference state) vehicle batteries 3 of the same type, and each of the vehicle batteries 3 is assigned a predefined load pattern. The load patterns are different and each represent a load ranging from low to high for the assigned vehicle battery 3. The load patterns provide information from which temporal profiles of a battery current in conjunction with temperature profiles can be derived, which stress the vehicle batteries 3 in different ways. This load pattern can also directly predetermine the operating variable profiles. Alternatively, the operating variable profiles can also be predefined using a predefined probabilistic model based on the load patterns. The load patterns then have a predictive character and represent a forecast for the future.Advantageously, they include a point estimator, especially for temperature and current, as well as a confidence.

[0074] In step S2, the plurality of vehicle batteries 3 are operated according to the predetermined load pattern.

[0075] In step S3, a check is made to determine whether an evaluation time t0 has been reached. The evaluation time can be scheduled at regular intervals, such as intervals between one week and two months.

[0076] If an evaluation time is reached (alternative: Yes), the process continues with step S4, otherwise it returns to step S2.

[0077] In step S4, based on the previous training state of the aging state model for a subset of the vehicle batteries 3, an information measure Info Jwhich indicates the information gain that can be obtained during the further measurement of the subset of the vehicle battery 3. This information measure can be specified using a predictive covariance Σ of the data-based aging state model. In particular, when using a Gaussian process model as a probabilistic regression model as part of the aging state model, the predictive covariance Σ of an energy storage device (index j) depends ∑(SOHj(t1),SOHj(t2),...,SOHj(tn)) does not depend on previously undetermined aging states, but only on input variables x(t0...t n ) for t0 the current time and t1....t n the future points in time or time steps.

[0078] An information measure Info J defined: InfoJ=∑t=t1Ta(σ(σL(x¯t))(h(x¯t|x¯1,...,x¯(t−1))) InfoJ=∑t=t1Ta(σL(x¯t))∫12 ln(2πe)ddet(∑(x¯t|x¯1,...,x¯(t−1)))pL(xt)dx where a discounting with the weighting factor a(σ L (x t )) is taken into account, with deterministically specified operating variables x1, ... , x (t-1) or in a probabilistic model for generating artificial farm size trends

[0079] a corresponds to a monotonically decreasing function, the σ L (x t ) for small values ​​to a value close to “1” and for large values ​​to a value close to 0, so that there is strong discounting at high uncertainty and only little discounting at low uncertainty.

[0080] Possible functions are a) a(σ L (x t )) = 1- erf(c * σ L (x t )) with the Gaussian error function erf(x) b) a(σ L (x t )) = 1-(pi / 2)^-1*arctan(c* σ L (x t )) c) a(σ L (x t )) = 1-tanh(c*σ L (x t))where the specified factor c describes how strong the discounting effect should be.

[0081] As described above, an information measure for a subset B of the total number of vehicle batteries B all as InfoJ=h(x¯)=∑t=t1Ta(σL(x¯t))12ln(2πe)ddet(∑(x¯t|x¯1,...,x¯(t(t−1)))) where x t ={x 1.t , ..., x J,t} for an evaluation time t for all vehicle batteries considered (index 1...J) the operating characteristics x derived from the operating variables t γ < 1 and corresponds to a discount factor that decreases over the time steps k (increasing k)

[0082] As a result, in step S4, the information measures Info Jabout the expected information gain from measuring a subset of vehicle batteries 3 from the total of vehicle batteries 3. The evaluated information measures refer to the future, but are discounted for temporal relevance and take into account already propagated uncertainties regarding load and aging prediction.

[0083] In a step S5, information measures are calculated for subsets B of vehicle batteries, each with a predetermined number J of vehicle batteries, and the subset B of vehicle batteries having a maximum information measure is selected. B∗=argmax{B⊂Ball|B|=J}InfoJ(B)

[0084] In step S6, the vehicle batteries 3 of the selected subset are now measured and in step S7 a label in the form of a measured aging state is determined at the current evaluation time.

[0085] This performs active learning. The goal of active learning is to reduce test bench duration and thus costs without compromising model accuracy. This can also be achieved by actively intervening in the operating parameter curve in a closed control loop using active learning, e.g., by adjusting the current or temperature on the test bench to create a relevant state in at least one battery.

[0086] Subsequently, in step S6, the selected vehicle batteries 3 are used to perform a highly accurate measurement under reproducible, defined conditions (temperature, charging / discharging direction, current, ambient conditions, ...) of the aging state as a label.

[0087] The aging state can be determined using a suitable aging state model or a suitable measurement method to determine the aging state.

[0088] A possible model or method for determining the state of aging is a measurement using Coulomb counting or by calculating a temporal current integral during the charging process. Here, the transferred charge is divided by the change in the state of charge between the beginning and the end of the respective charging and / or discharging phase. Calibration is advantageously carried out using the open-circuit voltage characteristic during rest phases in order to calculate the state of charge progression in the central unit. A sufficiently reliable indication of the state of aging can be obtained, for example, if the vehicle battery is brought from a defined relaxed state during a charging process from a fully discharged charge state to a fully charged state under reproducible load and ambient conditions.The maximum charge thus recorded can be related to an initial maximum charge capacity of the vehicle battery 3.

[0089] For the multitude of vehicle batteries, there are three aging states, as for example in Fig. 4. It is evident that a sudden death event S is occurring. The earlier such information can be measured and incorporated into the model through training, the faster the training can be. The goal here is for the aging model to be able to predict the nonlinear, sudden decline quite accurately.

[0090] In conjunction with the operating parameter curves resulting from the load pattern and the determined aging state as a label, training data sets are available with which the aging state model 4 can be trained. The newly obtained training data sets can be used to further train the data-based probabilistic regression model. In addition, automated hyperparameter tuning can be performed, e.g., using a gradient-based method or a black-box method such as Bayesian optimization.

[0091] In a subsequent step S8, it is checked whether the trained aging state model for a provided validation data set exceeds a sufficient accuracy of, for example, a maximum error of 1.5% SOHC. If this is the case (alternative: yes), the method continues with step S2; otherwise (alternative: no), the measurement of the vehicle batteries 3 is terminated. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] A. Girard, “Approximate Methods for Propagation of Uncertainty with Gaussian Process Models”, Doctor thesis, University of Glasgow, October 2004

[0031]

Claims

[1] Method for initially providing an at least partially data-based aging state model (4) for an electrical energy storage device (3), comprising the following steps: - Providing (S1) a number of energy storage devices (3) on a test bench (1) for measurement depending on a respective load profile, wherein the load profiles are different and characterize a temporal course of at least one loading operating variable for the energy storage device (3); - operating (S2) the number of energy storage devices (3) with the respective assigned load profile and recording temporal operating variable curves; - At a given evaluation time, determining (S4) an ageing state of a subset of the energy storage devices (3) as a label and generating a training data set with the operating variable profiles and the determined label for each energy storage device (3) of the subset of energy storage devices (3); - Selecting (S6) the subset of energy storage devices (3) with the respective assigned load profile depending on an information measure of the respective energy storage devices (Info J ) of the subset of energy storages (3), which is associated with a predictive covariance (Σ(x t |x1, ... , x (t-1) )) of the data-based aging state model is determined at at least one future point in time, wherein the information measure is determined as a function of a standard deviation, a variance or a Shannon entropy of a given load pattern model which is used to determine a future temporal course of the at least one operating variable in order to determine the predictive covariance at the future point in time. [2] Method according to claim 1, wherein the information measure (Info J ) with a sum of predictive covariances (Σ(x t |x1, ... , x (t-1))) is formed at future, in particular successive, evaluation times at which training data sets have been determined for training the aging state model. [3] Method according to claim 2, wherein the predictive covariances are each weighted with a weighting factor before the summation, which weights predictive covariances in the near future more highly than those further away. [4] Method according to claim 2 or 3, wherein the sum of the predictive covariances (Σ(x t |x1, ..., x (t-1) )) at future, especially consecutive, evaluation times of a sum of predictive covariances discounted with a weighting factor (Σ(x t |x1, ... , x (t-1) )), where the weighting factor is determined with a monotonically decreasing function (a) depending on the standard deviation, the variance or the Shannon entropy of the load pattern model. [5] Method according to one of claims 2 to 4, wherein, in the case of unknown operating variable profiles of one or more of the energy storage devices, an artificial operating variable profile is generated and, before the summation, the expected value is calculated using the predictive covariances (Σ(x t |x1, ... , x (t-1) )), in particular on the predictive covariances (Σ(x t |x1, ... , x (t-1) )) is formed, with respect to a probability distribution that describes the probability that the artificial farm size trend corresponds to the actual farm size trend. [6] Method according to one of claims 1 to 5, wherein the future course of the at least one operating variable is carried out with the aid of a probabilistic load prediction, wherein in particular uncertainties of the load prediction model result from a sampling of load trajectories according to a predetermined hidden Markov model which specifies a type of use of the electrical energy storage device, and an expected value and a standard deviation are determined from the load trajectories with regard to their distribution function. [7] Method according to claim 6, wherein the probabilistic load prediction is carried out using a method of moments, a Monte Carlo sampling or by integrating over the sampled distribution of the at least one operating variable. [8] Method according to one of claims 1 to 7, wherein the data-based aging state model (4) is designed with a probabilistic data-based model, in particular a Gaussian process model, wherein for one of the energy stores (3) an input vector of the data-based model, which comprises at least one operating variable profile (x(t)) and / or at least one operating characteristic (m(x(t))) from the at least one operating variable profile, an internal state of the energy store (3) and / or a physically modeled aging state, is mapped to the aging state of the respective energy store (3) to be modeled or a correction value for correcting a physically modeled aging state of the respective energy store (3), wherein the information measure (Info J ) for the subset of energy storage devices with a determinant of the predictive covariance to the respective energy storage device. [9] Method according to one of claims 1 to 8, wherein at each evaluation time only the energy storage devices (3) of the subset of energy storage devices (3) are measured to determine an ageing state as a label and the remaining energy storage devices (3) continue to be operated according to the assigned load profile. [10] Method according to one of claims 1 to 9, wherein at each evaluation time only the energy storage devices (3) of the subset of energy storage devices (3) are measured as a label to determine an ageing state and continue to be operated according to the load profile, while the remaining energy storage devices (3) are removed from the test bench. [11] Method according to one of claims 1 to 10, wherein the load profile and / or the resulting operating variable curve is adapted for the further measurement of one or more of the energy storage devices using active learning. [12] Method according to one of claims 1 to 11, wherein the subset of energy storage devices is determined depending on the associated information measure (Info J ) are selected for which a maximum amount of information (Info J ) results. [13] Apparatus for carrying out one of the methods according to one of claims 1 to 12. [14] A computer program product comprising instructions which, when the program is executed by at least one data processing device, cause the device to carry out the steps of the method according to any one of claims 1 to 12. [15] Machine-readable storage medium comprising instructions which, when executed by at least one data processing device, cause the device to carry out the steps of the method according to one of claims 1 to 12.

Images