METHOD FOR ESTIMATING THE STATE OF AN ENERGY STORE

Abstract

The invention relates to a method for estimating the state of an energy store comprising at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) using a battery management system (BMS) which comprises an impedance spectroscopy chip, having at least the following steps: a) determining the frequency-dependent impedance of the at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) using a data set recording taken in real-time, b) training an artificial neural network (60) with temperature-based training spectra as the input and a specification for temperature values belonging to each training spectrum as the output, c) taking into consideration a battery cell-to-battery cell variance (30) between the electrochemical battery cells (12, 14, 16, 18, 20, 22, 24, 26, 28) when testing the artificial neural network (60) using weighting functions ascertained during step b) and test spectra and estimating the temperature values belonging to the test spectra according to the weighting functions ascertained in step b), and d) estimating at least one internal state (SoC, SoH, T.sub.int) of the at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) of the energy store using the trained artificial neural network (6).

Claims

1. A method for estimating the state of an energy store comprising at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) by means of a battery management system (BMS), which comprises an impedance spectroscopy chip, said method comprising at least the following method steps: a) determining the frequency-dependent impedance of the at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) by means of a data set recording performed in real time, b) training an artificial neural network (60) with temperature-dependent training spectra as input and a preset for a temperature value belonging to each training spectrum as output, c) taking into consideration a battery cell-to-battery cell variance (30) between the electrochemical battery cells (12, 14, 16, 18, 20, 22, 24, 26, 28) during testing of the artificial neural network (60) with weighting functions, determined during method step b), and test spectra and estimating the temperature values belonging to the test spectra in accordance with the weighting functions determined in method step b), and d) estimating at least one internal state (SoC, SoH, Tint) of the at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) of the energy store by means of the trained artificial neural network (60).

2. The method as claimed in claim 1, wherein the data set in accordance with method step a) is given to the artificial neural network (60) at an input layer (64) in accordance with method step b) as training data set, a temperature value is passed on as output, and, in the training phase in accordance with method step b), weighting functions are determined, optimized and stored.

3. The method as claimed in claim 1, wherein the temperature of the at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) is known to the artificial neural network (60) in relation to the data set in accordance with method step a).

4. The method as claimed in claim 1, wherein weightings of a plurality of nodes (66, 72, 78) are parameterized within the artificial neural network (60) by a training algorithm until a desired input/output ratio is set and input values of the artificial neural network (60) are correctly assigned to output values.

5. The method as claimed in claim 1, wherein the artificial neural network (60) is provided with an input layer (64), at least one hidden layer (70) and an output layer (76) having at least one node (66, 72, 78).

6. The method as claimed in claim 5, wherein the data sets determined in accordance with method step a) are given at the input layer (64) to the artificial neural network (60).

7. The method as claimed in claim 6, wherein the output of the at least one internal state (SoC, SoH, T.sub.int) of the at least one electrochemical battery cell (12, 14, 16, 18, 20, 22, 24, 26, 28) is performed in the output layer (76).

8. The method as claimed in claim 6, wherein at least one variable determined by impedance spectroscopy is passed on to the input layer (64) of the artificial neural network (60).

9. The method as claimed in claim 1, wherein data sets of impedance data in the frequency range of between 0.1 Hz and 10 kHz are given to the artificial neural network (60) for training the artificial neural network (60) in accordance with method step b).

10. The method as claimed in claim 1, wherein a number of nodes (66) which corresponds to a supporting number of measurement frequencies in the frequency range during the data recording is provided in the input layer (64) in the artificial neural network (60) in accordance with method step b).

11. The method as claimed in claim 1, wherein up to 15 nodes (72) are arranged in the hidden layer (70) of the artificial neural network (60) for learning temperature predictions by means of measured impedance data sets.

12. The method as claimed in claim 1, wherein, for the temperature estimation by impedance measurement, an estimation of the influence of at least one load current (86, 88) or a load current (86) of 0 A is performed.

13. The use of the method as claimed in claim 1 for estimating the internal state (SoC, SoH, T.sub.int) of an energy store in electrically driven vehicles.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The invention will be described in more detail below with reference to the drawings, in which:

[0027] FIG. 1 shows a schematic illustration of a battery cell-to-battery cell variance for a plurality of battery cells of identical design,

[0028] FIG. 2 shows an exemplary network structure of an artificial neural network for temperature estimation with the aid of impedance data, and

[0029] FIG. 3 an estimation accuracy for an internal cell temperature for three used battery cells under load in a frequency range of between 0.1 Hz and 10 kHz.

DETAILED DESCRIPTION

[0030] In the description below of the embodiments of the invention, identical or similar elements are denoted by the same reference symbols, wherein these elements have not been described again in individual cases. The figures represent the subject of the invention only schematically.

[0031] FIG. 1 shows a battery cell-to-battery cell variance 30 for a plurality of battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 of identical design. In the illustration shown in FIG. 1, impedance data for a first battery cell 12, a second battery cell 14, a third battery cell 16, a fourth battery cell 18, a fifth battery cell 20, a sixth battery cell 22, a seventh battery cell 24, an eighth battery cell 26 and a ninth battery cell 28 are plotted in a Nyquist plot 10.

[0032] In the Nyquist plot 10 shown in FIG. 1, a real part 48 (Z.sub.real in mΩ) is plotted on an x axis while an imaginary part 50 (Zima.sub.imag in mΩ) of the corresponding impedance values is plotted on the y axis.

[0033] The battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 compared with one another here are, for example, lithium-iron-phosphate (LFeP) cells. Using the Nyquist plot 10 it is also possible for other electrochemical system cells, i.e. other lithium-ion battery cells, to be investigated. The Nyquist plot 10 can be used for all impedance spectra irrespective of that which is being investigated. The plot itself has nothing to do with the battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 compared with one another here and is merely a form of illustration which makes it possible to illustrate relationships. The method proposed according to the invention using an artificial neural network 60 enables the use of all electrochemical cells which can have the widest variety of chemical compositions. Lithium-ion battery cells merely describe a small section of this range.

[0034] In the Nyquist plot 10 shown in FIG. 1, accordingly different battery cell-to-battery cell variances 30 are recorded. These are given to an artificial neural network 60 (cf. illustration shown in FIG. 2) for training. In the training phase, in which temperatures belonging to the input spectra, for example, are preset as output values, the artificial neural network 60 determines weighting functions. The artificial neural network 60 adapts the internal parameters, i.e. the weighting functions, during the training phase in such a way that, during processing of the input data with the artificial neural network 60, a calculated value largely corresponds to the previously preset output value. In the test phase, on the other hand, the computation parameters are fixed within the artificial neural network 60. Therefore, input data are given in the artificial neural network 60, and these data are processed within the artificial neural network 60 in such a way, which takes place as part of computation operations, that an output value is achieved which specifies, for example, with a very high degree of accuracy, the actual temperature of a battery cell 12, 14, 16, 18, 20, 22, 24, 26, 28.

[0035] As can furthermore be seen from the Nyquist plot 10 shown in FIG. 1, the impedance data which are set in the case of eight different temperature values are recorded in each case for the nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28. Thus, for example, a first temperature 32 is −10° C., a second temperature 34 is 0° C., a third temperature 36 is 10° C., while a fourth temperature 38 is 20° C. (approximately room temperature). Furthermore, in the Nyquist plot 10 shown in FIG. 1, a fifth temperature 40 is plotted, which is 30° C., in addition a sixth temperature 42, which corresponds to 40° C., a seventh temperature 44, which corresponds to 50° C., and finally an eighth temperature 46, which corresponds to 60° C. It is apparent from the illustration in accordance with the Nyquist plot 10 in FIG. 1 that the scatter of the cell data, i.e. the battery cell-to-battery cell variance 30, is not inconsiderable.

[0036] Accordingly, as the temperature of the battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 increases, the internal resistance is reduced owing to quicker electrochemical reactions at the electrodes, the electrolyte and their boundary layers. The data set illustrated in FIG. 1 as Nyquist plot 10 is given to an artificial neural network 60 shown in FIG. 2.

[0037] The illustration in FIG. 2 shows a network architecture for an artificial neural network 60 for estimating the temperature with the aid of impedance data. An input vector 62 comprises impedance data as the real part 48 Re_Z and/or imaginary part 50 Im_Z. Here, other data could also be used, such as the impedance and the phase, which can be calculated from the real part 48 and the imaginary part 50. A hidden layer 70 can comprise, for example, up to 15 neurons and, with the aid of impedance training data, learns a pattern for temperature prediction. However, other network architectures could also be used which comprise a plurality of hidden layers 70 in which in each case more than 15 nodes 72 are implemented. In addition to the temperature, further states such as the state of charge (SoC) and a state of health (SoH) can also be estimated.

[0038] As can be seen from FIG. 2, the artificial neural network 60 illustrated in this figure comprises an input layer 64, which contains a number of nodes 66. In the network architecture illustrated in FIG. 2, the nodes 66 of the input layer 64 are identified by N.sub.11, N.sub.12, N.sub.13, N.sub.14, . . . N.sub.1.sub.2n.

[0039] An input vector 62, which is given to the input layer 64 of the artificial neural network 60 in accordance with the illustration in FIG. 2, is reproduced by individual impedance data, or, for example, the imaginary and real parts 50 and 48, respectively, thereof, thus, for example, Z.sub.Re(f1), Z.sub.Im(f1), which are given to the nodes 66 N.sub.13 and N.sub.12, respectively, in addition, for example Z.sub.Re(f2), Z.sub.Im(f2), which are both given to the nodes 66 N.sub.13 and N.sub.14, and so on. It is apparent from the network architecture illustrated in FIG. 2 that the individual nodes 66 of the input layer 64 are connected to the hidden layer 70, for example, via edges 68. The at least one hidden layer 70, which is part of the artificial neural network 60, likewise comprises a plurality of nodes 72 identified as N.sub.21, N.sub.22, N.sub.23, N.sub.24, ... N.sub.2.sub.2n. Edges 74 again run around the individual nodes 72 of the at least one hidden layer 70 to an output layer 70. There is only one node 78, identified as N31, in the output layer 76 illustrated in FIG. 2. This produces an output, i.e. an internal cell temperature at the output, i.e. the estimated value 80 T.sub.int of the respective battery cell 12, 14, 16, 18, 20, 22, 24, 26, 28.

[0040] If, for example, the abovementioned nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 are investigated by virtue of their electrochemical impedance spectra being recorded and at the same time their cell temperature being detected via a separate sensor, the determined data set, i.e. the electrochemical impedance spectrum, is passed on as input and the respective temperature is passed on as output to the artificial neural network 60 in order to perform the training phase of the artificial neural network 60, for example, for seven of the nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28. This means that, now, respective data sets in the form of impedance spectra are passed on as input and temperatures for a battery cell 12, 14, 16, 18, 20, 22, 24, 26, 28 are passed on as output for each of seven selected ones of the in total nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 to the artificial neural network 60 in order to perform the training phase at the artificial neural network 60. The artificial neural network 60 is accordingly told that the respective impedance spectra represent the input and the respective output value corresponds to the associated temperature and must assume this value. Now, computation operations take place within the artificial neural network 60, which determines internal weighting functions during the training phase such that, after mathematical processing of the input data taking into consideration determined network weighting functions, the output value appears. When the weighting functions within the artificial neural network 60 are so good that the output value is met, storage thereof takes place.

[0041] In the subsequent test phase of the artificial neural network 60, the data sets of the two still remaining ones of the in total nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 are now used. In this case, only the impedance spectra are passed on as input to the artificial neural network 60. These are now mathematically processed with the weighting functions previously determined and then stored during the training phase, and, at the end, an output value is estimated. This output value, within the context of the test phase of the artificial neural network 60, reflects the temperature estimated thereby of the respective battery cell 12, 14, 16, 18, 20, 22, 24, 26, 28. This estimated temperature for the respective battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 is compared with the actually measured temperature of these two battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 at the end, cf. illustration shown in FIG. 3.

[0042] The data shown in FIG. 1 are given to the artificial neural network 60 embodied in a network architecture shown in FIG. 2 for training, for example, seven of the nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 and, for testing, the data or impedance spectra of two of the nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 as part of the input vector 62 at the input layer 64. In relation to this data set, i.e. the impedance spectra, the corresponding temperature is known as output value to the artificial neural network 60. The data set of the input vector 62 is given to the input layer 64 of the artificial neural network 60 illustrated in FIG. 2 as part of a feedforward structure, for example, as training data or teaching data. In the present context, feedforward should be understood to mean that there is an input into the function or the artificial neural network 60 on the input side, i.e. on the input layer 64 of the artificial neural network 60, and an output, which arises at the output layer 76 or at the node 78, is calculated. Within the artificial neural network 60, weightings of individual neurons in the artificial neural network 60 are adapted until a desired input-to-output ratio is present. This takes place within a training algorithm. In the training algorithm of the artificial neural network 60, as part of an error feedback or backpropagation, an error is minimized until a desired input-to-output ratio is present. The impedance data of the impedance measurement are present as input vector 62 at the input layer 64. The number of nodes 66 contained in the input layer 64 is dependent on, among other factors, the supporting number used of measurement frequencies in the frequency range. In the present example of the artificial neural network 60, a node 78, which reflects an estimated temperature T.sub.int of the recorded impedance measurement, is positioned at the output layer 76. The number of layers between the input layer 64 and the output layer 76 can vary, depending on the number of hidden layers 70 therebetween. This may be at least 1 and can be optimized corresponding to the input data used.

[0043] Impedance data recorded under laboratory conditions train the artificial neural network 60. Now, new unknown measurement data can be presented to the trained artificial neural network 60. These data are used in the test phase of the artificial neural network 60. In this test phase, output values are estimated, wherein weighting functions which were determined, optimized and weighted in the training phase are used.

[0044] For the temperature estimation by way of impedance spectroscopy under load, an estimation of the influence of a load current is necessary. For temperatures above 0° C., it has become apparent that a frequency range for the measurement of from 0.1 Hz to 10 kHz results in a negligible error.

[0045] FIG. 3 shows an illustration of an estimation accuracy 84 for an internal cell temperature for used battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 illustrated therein under load within a frequency range of from 0.1 Hz to 10 kHz.

[0046] In the illustration shown in FIG. 3, the characteristic of the estimation accuracy 84 for an internal cell temperature for three used ones of the in total nine battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 under load within a frequency range of from 0.1 Hz to 10 kHz is recorded. The temperature characteristic 82 extends from −10° C. to approximately 40° C. The individual values for the estimation accuracy 84 are indicated by a cross for a first load current 86 and by a circle for a second load current 88. The illustration shown in FIG. 3 shows an error to be expected in the temperature estimation of the internal cell temperature T.sub.int taking into consideration the battery cell-to-battery cell variance 30. A superimposed direct current as artificial load results in a temperature error of <8° C. The error range is primarily defined by the cell scatter, i.e. the battery cell-to-battery cell variance 30, since comparison measurements without load current 86, 88 are approximately of the same order of magnitude.

[0047] It is possible to read, for example, from the illustration shown in FIG. 3 how great an error in a temperature estimation is at different temperatures for various battery cells 12, 14, 16, 18, 20, 22, 24, 26, 28 under different loads. For example, the temperature of 20° C. marks the highest point on the x axis. Compared with the y axis, this demonstrates that an error of +4° C. has arisen in the temperature estimation. This means that the artificial neural network 60 has predicted a temperature of 24° C. although the actual temperature was only 20° C.; accordingly an error of +4° C. has occurred. The conclusion from FIG. 3 is that in each case one temperature for a battery cell 12, 14, 16, 18, 20, 22, 24, 26, 28 can be estimated for various actual temperatures with the artificial neural network 60 via the impedance spectra, and that the discrepancy, i.e. the error (y axis) in the estimation accuracy 84 for the estimated temperature is only a few degrees Celsius from the actual temperature (x axis).

[0048] The solution proposed according to the invention can advantageously be used in energy storage systems which have electrochemical battery cells, in particular lithium-ion battery cells. In a particularly advantageous manner, the battery management system assigned to this energy storage system can use an impedance measurement chip. In addition, the method proposed according to the invention can be used in energy storage systems in which information on the cell temperature of the battery cell 12, 14, 16, 18, 20, 22, 24, 26, 28 is required and no temperature sensors are available. Battery management systems which are used in particular in plug-in hybrids, in hybrid vehicles and in electric vehicles can be equipped with the method proposed according to the invention, and in particular battery management systems to be expected in the future can be correspondingly retrofitted or prepared.

[0049] The invention is not restricted to the exemplary embodiments described here and the aspects highlighted therein. Rather, a large number of modifications which are within the scope of a person skilled in the art is possible within the scope specified in the claims.