METHOD FOR TRAINING A MACHINE MODEL TO DETERMINE A BATTERY SYSTEM STATE

Description

[0001] With the increasing use of battery systems as an energy source for transport vehicles, the analysis of the aging behavior of these battery systems is becoming increasingly important.

[0002] Simulating aging behavior and predicting a (future) battery system state often requires a battery cell-specific aging model, whose aging parameters must be laboriously determined in the laboratory. This is time-consuming and costly. Furthermore, the aging model is often only valid for battery systems that use this specific battery cell type. VATANI MOHSEN ET AL: "Cycling Lifetime Prediction Model for Lithium-ion Batteries Based on Artificial Neural Networks", 2018 IEEE PES INNOVATIVE SMART GRID TECHNOLOGIES CONFERENCE EUROPE (ISGT-EUROPE), IEEE, October 21, 2018, pages 1-6, discloses a method for training a machine model.

[0003] The object of the present invention is to improve the known methods. A method for training a machine model to describe a battery system is proposed, comprising: providing training data comprising constant battery description vectors and, for each of these battery description vectors, a sequence of temporally consecutive measured value vectors; processing the training data in the machine model, thereby obtaining a simulated battery system state; providing a cost function that describes a deviation of the simulated battery system state from a reference battery state; and training the machine model, wherein the cost function is minimized by adapting parameters of the machine model.

[0004] The battery description vectors describe properties of the battery system that do not change over time. The measured value vectors describe the battery system states of various variables that are measured over time.

[0005] The simulated battery system state can correspond to a current battery system state or a future battery system state. If it is a future battery system state, it can also be referred to as a prediction of the battery system state.

[0006] The claimed method can increase the amount of data required for proper training of the machine model. For example, predicting an end-of-life (EOL) typically requires measurement vectors from battery systems that have reached the end of their service life. In many cases, such measurement vectors are not available for all battery cell types. By jointly processing different battery systems, each described by battery description vectors, predictions can also be made for battery systems for which no complete sets of chronologically consecutive measurement vectors obtained in laboratory tests are available.

[0007] One embodiment of the method provides for the adjustment of machine model parameters under boundary conditions. Specifying boundary conditions can contribute to improved machine model quality.

[0008] For example, the boundary conditions can specify that a specific battery system state of the measured value vector may only change monotonically over time. It is also conceivable to specify that the battery system states must be subject to a certain statistical distribution. Furthermore, the boundary conditions can include known relationships between different physical quantities of the measured value vectors.

[0009] In embodiments of the method, it is provided that the parameters of the machine model comprise at least one parameter that describes a temporal distribution of at least one measured value of the measured value vector.

[0010] Further embodiments may provide that the boundary conditions include a minimum and / or maximum of a parameter value. For example, the boundary conditions may specify that the charge of the battery system must not exceed a maximum value.

[0011] The boundary value conditions can include known relationships between different physical quantities of the measured value vectors.

[0012] For example, the described method can take into account that a current flow causes the battery system to heat up. It can also be taken into account that at the end of the service life, the temperature increase when a current flows through the battery system can be higher than at the beginning of the service life (BOL). This can allow conclusions to be drawn about the impedance of the battery system as well as an increase in impedance as the battery system ages. It is also conceivable that the proposed method can take into account that at BOL the battery cells of the battery system relax more quickly when a current flowing through the battery system is switched off than at EOL. One reason for this may be that current and potential densities become more inhomogeneous with increasing service life of the battery system.In connection with temperature measurements, it can also be considered that the time of year and the current location of the battery system, which can be determined, for example, by NSS systems, especially GPS, are influenced, and not exclusively by the current flowing through the battery system. Furthermore, the method also allows for a similarity measure between different cell chemistries and the spatial proximity of battery cells within the battery system to be taken into account. By utilizing this diverse information, the training of the described machine model can be significantly improved.

[0013] The measured value vectors evaluated in connection with the described method can, in particular, include current values, voltage values, and / or temperature values. It is also conceivable for the measured value vectors to include a location within the battery system. Furthermore, a season or season can also be taken into account (e.g., winter, summer; dry season, rainy season).

[0014] The battery description vectors, i.e. properties of the battery system which remain unchanged over the lifetime of the battery system, can describe a cell chemistry of a battery cell of the battery system and / or a physical structure of a battery cell of the battery system and / or a number of battery cells of the battery system and / or a wiring architecture of batteries of the battery system and / or a nominal energy density of the battery system and / or a nominal power density of the battery system.

[0015] The cell chemistry may, for example, include information on the cathode material and / or the anode material and / or the geometry of the cathode and / or anode material and / or the electrolyte.

[0016] Information about the cathode material may include information about the active material. For example, the active material of the cathode material may comprise nickel and / or manganese and / or cobalt and / or iron and / or phosphorus. Furthermore, the information may specify a precise composition (e.g., Li(1-x)Ni 1 / 3 Mn 1 / 3 C 1 / 3 O 2 ). It may also state whether and, if so, which conductive material (e.g., x percent graphite) is added. Furthermore, additives added for the processability of the cathode material and / or its service life may be included in the information about the cathode material.

[0017] The anode material can also include the active material (e.g. graphite, silicon, lithium (metal)), its exact composition (e.g. Li x C 6 ), the type of graphite (e.g. needles coke, MCMB, graphene) and / or additives added for processability, lifetime, etc.

[0018] Layer thicknesses and / or porosities and / or tortuosity can also be specified for the cathode material and / or anode material.

[0019] The physical structure may in particular include information on the battery cell size and / or electrode area and / or the number of arresters and / or the position and / or the design.

[0020] The designs of the battery cells can include, for example, round cells, which can also be referred to as cylindrical cells, prismatic cells and / or pouch cells, which are also referred to as coffee bag cells.

[0021] Further information that can generally be taken into account are external factors such as cooling systems and their topologies, tensions in the modules and the system, and / or a measured pressure in the system as well as possibly installed phase change materials (PCM).

[0022] Furthermore, a method for determining a battery system state is proposed, which comprises: providing a constant battery description vector and a sequence of temporally successive measurement vectors, and processing the battery description vector and the sequence of temporally successive measurement value vectors in a machine model described above, whereby the battery state is obtained.

[0023] Furthermore, the present disclosure relates to a device for determining a battery system state, which comprises processing means suitable for carrying out one of the methods described above.

[0024] Examples of embodiments are illustrated below with the help of the drawings.

[0025] Fig. 1 shows an embodiment of a method for training a machine model. The method comprises providing training data 101, which comprises constant battery description vectors and, for each of these battery description vectors, a sequence of temporally successive measured value vectors. The training data 101 is processed in a machine model 102, obtaining a simulated battery system state 103. Furthermore, a cost function 105 is provided, which describes a deviation of the simulated battery system state 103 from a reference battery state 104. The machine model 102 is trained, wherein the cost function 105 is minimized by adapting parameters of the machine model.

[0026] Fig. 2shows an example of a method for determining a battery system state. A constant battery description vector 201 and a sequence of temporally successive measured value vectors 202 are provided, which are processed in a machine model 203, whereby a battery system state 204 is obtained. The machine model 203 can in particular be Fig. 1 described procedures have been trained.

[0027] Fig. 3 shows a device for determining a battery system state 310. The device 310 comprises a computing unit 311 and a memory 312. The computing unit 311 is configured to carry out the method according to Fig. 1 or Fig. 2 to carry out.

[0028] Fig. 4 , Fig. 5 and Fig. 6 illustrate exemplary courses of measured value vectors for two different battery description vectors, which characterize two different battery systems.

[0029] Both battery systems consist of single round cells, specifically 18650 round cells. The battery systems do not differ in their design, but only in the cell chemistry used, particularly in the differences in the anode and cathode materials used.

[0030] The Figures 4 to 6 The dashed curves refer to a lithium iron phosphate cell (LFP cell) and the measured values ​​shown with a solid line refer to a lithium nickel manganese cobalt cell (NMC cell).

[0031] Fig. 4 shows a discharge curve for the two different cells. The voltage in volts is plotted against time in minutes. The two cells were discharged at a constant discharge rate of 1C, which corresponds to a full discharge of the cell within one hour.

[0032] The characteristic curves clearly show that the slopes of the two curves differ significantly. The voltage limits also differ, ranging from 2.0 V to 3.6 V for the LFP cell and from 2.5 V to 4.2 V for the NMC cell.

[0033] In the Fig. 5 The voltage in volts is plotted against the charge in Ah drawn from the respective cell. Fig. 5 It can be seen that the LFP cell has a capacity of 1.1 Ah and the NMC cell has a capacity of 2.5 Ah. It can also be seen that the voltage curve differs depending on the discharged charge, which corresponds to the cell's state of charge (SOC).

[0034] Fig. 6 illustrated how much charge change corresponds to which voltage change. The charge change dQ / dV in Ah per volt is plotted against the voltage in volts.

[0035] The Figures 4 to 6It can be seen that the measured value vectors and the corresponding derived quantities differ significantly for different constant battery description vectors.

[0036] The proposed method makes it possible to train a machine model for many different battery systems together, and thus to make predictions even for battery systems for which only a few measured value vector curves are available.

Claims

1. Method for training a machine model for describing a battery system, comprising: - Provision of training data, which comprise constant descriptive battery vectors and, in each case, for these descriptive battery vectors, a series of temporally sequential measurement value vectors; - Processing of training data in the machine model, wherein a simulated battery system state is obtained; - Provision of a cost function, which describes a deviation of the simulated battery system state from a reference battery state; - Training of the machine model wherein, by the adjustment of parameters of the machine model, the cost function is minimized.

2. Method according to patent Claim 1, wherein the adjustment of parameters of the machine model is executed under marginal value conditions.

3. Method according to patent Claim 1 or 2, wherein parameters of the machine model comprise at least one parameter which describes a temporal distribution of at least one measurement value of the measurement value vector.

4. Method according to patent Claim 2 or 3, wherein marginal value conditions comprise a minimum and / or a maximum parameter value.

5. Method according to one of patent claims 2 to 4. wherein marginal value conditions comprise known relationships between different physical variables of measurement value vectors.

6. Method according to one of patent Claims 1 to 5, wherein measurement value vectors comprise: current values, and / or voltage values, and / or temperature values.

7. Method according to one of patent Claims 1 to 6, wherein the descriptive battery vectors describe: a cell chemistry of a battery cell of the battery system, and / or a physical structure of a battery cell of the battery system, and / or a number of battery cells of the battery system, and / or a circuit architecture of batteries of the battery system, and / or a rated energy density of the battery system, and / or a rated power density of the battery system.

8. Method according to patent Claim 7, wherein the cell chemistry comprises data: on the cathode material, and / or on the anode material, and / or on the stoichiometry of the cathode and / or anode material, and / or on the electrolyte.

9. Method according to patent Claim 7 or 8, wherein the physical structure comprises data: on the battery cell size, on the electrode surface area, on the number of diverters, on the position of diverters, on the structural design.

10. Method for determining a state of a battery system, comprising: - Provision of a constant descriptive battery vector and of a series of temporally sequential measurement vectors, and - Processing of the descriptive battery vector and of the series of temporally sequential measurement vectors in a machine model according to one of patent Claims 1 to 9, wherein the battery system state is obtained.