Method and system for predicting a state of health of a battery

Abstract

There is described a method for predicting a state of health, SOH, of a redox flow battery, the method comprising: receiving battery data of the redox flow battery; and performing, using a machine learning agent, an analysis of the battery data to predict the SOH of the redox flow battery.

Description

[0001] LITRICITY GmbH - 1 - 30A-162 925

[0002] Method and system for predicting a state of health of a battery

[0003] Technical Field

[0004] The present disclosure generally relates to a method for predicting a state of health, SOH, of a redox flow battery, a method of managing a redox flow battery, a method for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery, a computer program product comprising program code portions for performing a method according to any one of the example implementations as described herein, a computing unit for predicting a SOH of a redox flow battery, a computing unit for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery, and a system comprising a computing unit for predicting a SOH of a redox flow battery and in particular further comprising a computing unit for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery.

[0005] Background

[0006] Current battery management systems (BMS) operate on the basis of parameters like voltage, current, charge and temperature as a function of charge. Fairly complex models try to mirror the actual state of the battery to provide the basis for control decisions. One way to describe the status of a battery in a better way is to use electrochemical impedance spectroscopy (EIS). Because of its large frequency spectrum, all processes that take place in a battery may practically be reflected in the EIS spectrum. The analysis, however, is hampered by the complexity of the system which usually contains 10-20 elements if the impedance behaviour is described by equivalent circuits which are usually used to explain impedance behaviour. This does not allow for a precise assessment of the status of the battery, although attempts have been made to develop battery models with which interpretation of impedance spectra has been attempted.

[0007] There is therefore a need for improving analysis of the state of a battery, in particular to control the battery using a BMS. Summary

[0008] According to an aspect of the present disclosure, there is provided a method for predicting a state of health, SOH, of a redox flow battery, the method comprising: receiving battery data of the redox flow battery; and performing, using a machine learning agent, an analysis of the battery data to predict the SOH of the redox flow battery. The machine learning agent may, in some examples, predict the SOH and / or remaining useful life of the redox flow battery by comparing one or more electrochemical impedance spectra of the redox flow battery with one or more electrochemical impedance spectra which have been used in order to train the machine learning agent.

[0009] In some examples, the machine learning agent is trained using a training data set comprising an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery as input data, wherein the machine learning agent is trained using a capacity of the redox flow battery and / or the other redox flow battery as output data. In some examples, the training data set is obtained by: measuring the capacity and performing the EIS measurement to obtain capacity data and corresponding EIS data, respectively, of the redox flow battery and / or the other redox flow battery; providing the capacity data and the corresponding EIS data to the machine learning agent; and fitting, by the machine learning agent, a model to the capacity data and the corresponding EIS data. This allows for a particularly precise prediction of the SOH and / or remaining useful life of the redox flow battery. Furthermore, in some examples, the measurement of the capacity and the performing of the EIS measurement are repeated for a plurality of charge-discharge cycles until the capacity diminishes to a predetermined percentage, in particular 90 percent, of an original value of the capacity, wherein the capacity data and the corresponding EIS data are provided to the machine learning agent for each of the charge-discharge cycles, and wherein the machine learning agent fits the model to the capacity data and the corresponding EIS data obtained for all chargedischarge cycles. This may further improve prediction of the SOH and / or remaining useful life of the redox flow battery.

[0010] In some examples, said performing, using the machine learning agent, of the analysis of the battery data to predict the SOH of the redox flow battery comprises: providing the battery data to the machine learning agent, and predicting the SOH based on the model. In some examples, the machine learning agent is trained using a further training data set comprising an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery as input data, and a remaining useful life, RUL, of the redox flow battery and / or the other redox flow battery as output data. In some examples, the further training data set is obtained by: measuring the capacity and performing the EIS measurement to obtain RUL data and corresponding EIS data, respectively, of the redox flow battery and / or the other redox flow battery; providing the RUL data and the corresponding EIS data to the machine learning agent; and fitting, by the machine learning agent, a further model to the RUL data and the corresponding EIS data. This may improve the prediction of the SOH and / or the RUL further.

[0011] In some examples, the measurement of the capacity and the performing of the EIS measurement are repeated for a plurality of charge-discharge cycles until the capacity diminishes to a predetermined percentage, in particular 90 percent, of an original value of the capacity, wherein the RUL data and the corresponding EIS data are provided to the machine learning agent for each of the charge-discharge cycles, and wherein the machine learning agent fits the model to the RUL data and the corresponding EIS data obtained for all charge-discharge cycles. Again, this may improve the prediction of the SOH and / or the RUL further.

[0012] In some examples, for each charge-discharge cycle, the RUL is calculated by ((Total life cycles - passed cycles / Total life cycles)*100).

[0013] In some examples, said performing, using the machine learning agent, of the analysis of the battery data to predict the SOH of the redox flow battery comprises: providing the battery data to the machine learning agent, and predicting the SOH based on the further model.

[0014] In some examples, the model which is fitted to the capacity data and the corresponding EIS data and the further model which is fitted to the RUL data and the corresponding EIS data are combined into a single model. This may allow for providing an improved model for more precise prediction of the SOH and / or RUL of the redox flow battery.

[0015] In some examples, the battery data comprises an electrochemical impedance spectrum, EIS, of the redox flow battery for which the SOH is predicted. The EIS of the redox flow battery may hereby, in some examples, be obtained for a plurality of frequencies. The frequencies may, in some examples, range from 1 mHz to n MHz, wherein n e [1, 2, ..., 10]. Additionally or alternatively, the EIS of the redox flow battery may, in some examples, be obtained after fully charging and discharging the redox flow battery, in particular obtained periodically after fully charging and discharging the redox flow battery.

[0016] In some examples, receiving battery data of the redox flow battery comprises performing an EIS measurement on the redox flow battery.

[0017] In some examples, the method further comprises predicting, based on the analysis of the battery data by the machine learning agent, a remaining useful life, RUL, of the redox flow battery. The RUL may hereby be defined by a ratio between a maximum capacity of the redox flow battery at a specific point in time and an initial capacity of the redox flow battery at a previous point in time.

[0018] In some examples, predicting the SOH of the redox flow battery comprises predicting a capacity of the redox flow battery based on the analysis, via the machine learning agent, of the EIS. The prediction of the capacity may hereby, in some examples, be further based on a value of the capacity of the redox flow battery obtained at a specific point in time in the past.

[0019] In some examples, the analysis is performed, via the machine learning agent, based on a comparison between the battery data and previously obtained battery data of the redox flow battery and / or another redox flow battery. The comparison comprises, in some examples, analysing, via the machine learning agent, a deviation of the EIS of the redox flow battery from a previously obtained EIS of the redox flow battery and / or the other redox flow battery.

[0020] In some examples, the machine learning agent is an artificial recurrent neural network agent. This has proven to result in particularly precise prediction of the SOH and / or RUL of the redox flow battery.

[0021] In some examples, the machine learning agent is trained based on input data comprising battery data of the redox flow battery and / or another redox flow battery for which impurities have been introduced, at least for a fraction of time during which the battery data has been obtained, into a fluid circulated in the redox flow battery and / or the other redox flow battery. This may mimic degradation of the redox flow battery and may thus allow for further improvements regarding the precision for predicting the SOH and / or RUL of the redox flow battery.

[0022] In some examples, the machine learning agent is comprised in a centralised machine learning system.

[0023] In some examples, the method further comprises displaying the predicted SOH of the redox flow battery on a display.

[0024] There is further provided a method of managing a redox flow battery, the method comprising: predicting a SOH of the redox flow battery according to any one or more of the example methods as outlined above; and managing the redox flow battery based on the precited SOH. In some examples, said managing of the redox flow battery comprises monitoring a capacity state of the redox flow battery.

[0025] There is further provided a method for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery, the method comprising: providing or obtaining a training data set comprising (i) an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery and (ii) a capacity and / or a remaining useful life, RUL, of the redox flow battery and / or the other redox flow battery; and training the machine learning agent using the training data set comprising the electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or the other redox flow battery as input data and the capacity and / or RUL of the redox flow battery and / or the other redox flow battery as output data.

[0026] There is further provided a computer program product comprising program code portions for performing the method of any one of the examples outlined above when the computer program product is executed on one or more computing devices. The computer program product may hereby be stored on a computer-readable recording medium.

[0027] There is further provided a computing unit for predicting a SOH of a redox flow battery, the computing unit comprising at least one processor and at least one memory, the at least one memory containing instructions executable by the at least one processor such that the computing unit is operable to perform the method of any one of the examples outlined above. There is further provided a computing unit for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery, the computing unit comprising at least one processor and at least one memory, the at least one memory containing instructions executable by the at least one processor such that the computing unit is operable to perform the method of any one of the examples outlined above.

[0028] There is further provided a system comprising a computing unit for predicting a SOH of a redox flow batter as outlined above and, optionally, a computing unit for training a machine learning agent as outlined above.

[0029] Brief Description of the Drawings

[0030] Further aspects, details and advantages of the present disclosure will become apparent from the detailed description of exemplary embodiments below and from the drawings, wherein:

[0031] Figure 1 shows a schematic block diagram of a system according to some example implementations as described herein;

[0032] Figure 2 shows EIS spectra of a Li-ion battery;

[0033] Figures 3a and b show an online capacity estimation and a remaining useful life prediction of a Li-ion battery, respectively, according to some example implementations as described herein;

[0034] Figures 4a-k show measurements, capacity prediction and remaining useful life prediction of redox flow batteries according to some example implementations as described herein;

[0035] Figure 5 shows a flow diagram of a method according to some example implementations as described herein;

[0036] Figure 6 shows a flow diagram of a further method according to some example implementations as described herein; Figure 7 shows a flow diagram of a further method according to some example implementations as described herein; and

[0037] Figures 8a and b show schematic block diagrams of computing units according to some example implementations as described herein.

[0038] Detailed Description

[0039] In the following description, for purposes of explanation and not limitation, specific details are set forth in order to provide a thorough understanding of the present disclosure. It will be apparent to one of skill in the art that the present disclosure may be practiced in other embodiments that depart from these specific details.

[0040] Those skilled in the art will further appreciate that the steps, services and functions explained herein below may be implemented using individual hardware circuitry, using software functioning in conjunction with a programmed micro-processor or general purpose computer, using one or more Application Specific Integrated Circuits (ASICs) and / or using one or more Digital Signal Processors (DSPs). It will also be appreciated that when the present disclosure is described in terms of a method, it may also be embodied in one or more processors and one or more memories coupled to the one or more processors, wherein the one or more memories are encoded with one or more programs that perform the steps, services and functions disclosed herein when executed by the one or more processors.

[0041] In order to rationalise the problem of the state of the art in which the precise assessment of the status of a battery has not been possible, machine learning is used in example implementations as described herein to identify degradation patterns in redox flow batteries. In particular repeatedly recording impedance spectra upon cycling with, in some examples, ten or more spectra in a chargedischarge cycle and cycling several thousand times may allow creating a data base which can be used for the development of algorithms that assess the properties of the redox flow battery. The application allows for an assessment of the state of a redox flow battery by just measuring a few EIS spectra (or even just one EIS spectrum). This is a basis for a battery management system superior to previous approaches since it mirrors all properties of the redox flow battery. The present disclosure therefore generally relates to a machine learning approach which can be used for real-time state of health and / or remaining useful life prediction for redox flow battery systems.

[0042] Figure 1 shows a schematic block diagram of a system 100 according to some example implementations as described herein.

[0043] In this example, the system 100 comprises a redox flow battery 102. The redox flow battery 102 is coupled to a sampling module 104 which is configured to receive and process battery data relating to the redox flow battery 102.

[0044] Furthermore, in this example, a control module 106 is coupled to the sampling module 104. The control module 106 is configured to analyse the battery data and train a machine learning model to predict the state of health (for example (online) capacity) and / or the remaining useful life of the redox flow battery 102. The control module 106 is used, in this example, to control the redox flow battery 102 based on the predicted state of health and / or the remaining useful life of the redox flow battery 102.

[0045] In this example, the control module 106 is coupled to a display module 108 which is configured to visualise the prediction of the state of health and / or the remaining useful life of the redox flow battery 102.

[0046] In some examples, the system 100 may be a centralised system, in particular a centralised machine learning-based system, which is configured to predict the state of health and / or the remaining useful life of the redox flow battery 102. This may, in some examples, entail battery data of the redox flow battery 102 being uploaded to, for example, an external server and / or into the cloud (that is using cloud computing, in which on-demand availability of computer system resources, such as, for example, data storage and / or computer power, is provided, without the requirement of direct active management by a user of the cloud computing environment).

[0047] Throughout the present disclosure, the state of health of the redox flow battery may refer to one or more of a capacity of the redox flow battery, a power capability of the redox flow battery, deposition of used material at the electrode in an intermediate layer between the electrode of the redox flow battery and a fluid circulating within the redox flow battery, an internal resistance and / or impedance and / or conductance of the redox flow battery, a voltage of the redox flow battery, a self-discharge of the battery, an ability of the redox flow battery to accept a charge, a number of (for example full) charge-discharge cycles the redox flow battery has undergone, an age of the battery, a temperature of the redox flow battery when it was used previously, a total energy charged and discharged, and other parameters. The state of health may hereby be a figure of merit (measured in percent) of the condition of the redox flow battery compared to its ideal conditions, whereby the ideal condition may be 100% at the time of manufacture of the redox flow battery. In some examples, however, the state of health of the redox flow battery may, in some examples, be considered as being less than 100% at its time of manufacture if the redox flow battery does not meet its (theoretical) specifications at the time of manufacture.

[0048] Furthermore, throughout the present disclosure, the remaining useful life of the redox flow battery may relate to a drop in capacity compared to an initial capacity of the redox flow battery. For example, the end of a useful life of the redox flow battery may, in some examples, be considered to have been reached when the capacity has dropped to 80% (or another percentage, e.g. 90%) of the initial capacity of the redox flow battery. The remaining useful life may thus, for example, be defined as a difference between the total number of charge-discharge cycles when the actual capacity of the redox flow battery drops to the threshold value (e.g. 80% or 90% or another value) and the number of charge-discharge cycles of the current redox flow battery. The number of charge-discharge cycles may hereby refer to cycles of fully charging and discharging the redox flow battery.

[0049] The battery data which is received and processed by the sampling module 104 comprises, in this example, one or more electrochemical impedance spectra. The one or more electrochemical impedance spectra may hereby relate to spectra of the redox flow battery 102, the state of health of which is to be analysed, and / or to spectra of one or more other redox flow batteries. The one or more other redox flow batteries may hereby, preferably, relate to batteries of the same type as the redox flow battery 102, in particular comprising the same materials being used as in the redox flow battery 102. This may allow for a more precise training of the machine learning agent, thereby providing for a more precise prediction of the state of health and / or the remaining useful life of the redox flow battery 102.

[0050] The machine learning agent may predict the state of health and / or the remaining useful life of the redox flow battery 102 in particular by comparing one or more electrochemical impedance spectra of the redox flow battery 102 with one or more electrochemical impedance spectra which have been used in order to train the machine learning agent. In some examples, the one or more electrochemical impedance spectra of the redox flow battery 102 may also be used in order to train the machine learning agent.

[0051] The machine learning agent / model is trained, in this example, with one or more electrochemical impedance spectra as input data. Furthermore, the machine learning agent / model is trained, in this example, with a capacity and / or remaining useful life of the redox flow battery 102 and / or of the other one or more redox flow batteries as output data.

[0052] Different machine learning approaches, such as supervised learning, unsupervised learning or reinforcement learning may be used. An artificial recurrent neural network architecture has proven to allow for accurate prediction of the state of health and / or remaining useful life of the redox flow battery, as will be shown further below in relation to figures 3a and b.

[0053] Figure 2 shows EIS spectra 200 of a Li-ion battery which may form the basis of the analysis of the state of health and / or remaining useful life of the Li-ion battery, and / or which may be used in order to train the machine learning agent.

[0054] The EIS spectra are collected in the sampling module 104. The EIS spectra are mapped along cycle numbers of the Li-ion battery. The EIS spectra are obtained at different stages of charging / discharging under galvanostatic conditions in the frequency range of 10 mHz to 100 kHz (or 1 MHz) with an excitation current of 20 mA. The applied voltage is kept constant, with an excitation voltage of 10 mV for voltage modulation.

[0055] Figure 3a shows an online capacity estimation 300 of a Li-ion battery according to some example implementations as described herein.

[0056] The measured capacity and the estimated (that is predicted) capacity of the Li-ion battery are shown as a function of cycle number. As can be seen, while the measured capacity and the estimated (predicted) capacity of the Li-ionbattery slightly deviate between about 50 and 3000 charging-discharging cycles, the prediction of the remaining useful life of the Li-ionbattery (defined, in this example, as 80% of the initial capacity) matches the measured remaining useful life (reached, in this example, after approximately 5000 cycles) within the margin of error for predicting the remaining useful life using the machine learning agent according to some example implementations as described herein. The margin of error of the measured capacity is hereby negligible.

[0057] Figure 3b shows the remaining useful life prediction 350 of a Li-ion battery according to some example implementations as described herein.

[0058] The predicted remaining useful life is hereby plotted against the actual (measured) remaining useful life. The prediction of the remaining useful life and the measurement of the actual remaining useful life are hereby obtained from respective EIS measurements at the current cycle, without requiring EIS measurements from previous cycles. As can be seen, the predicted remaining useful life matches the actual remaining useful life within the margin of error for predicting the remaining useful life using the machine learning agent according to some example implementations as described herein.

[0059] Figure 4a shows the 3D Nyquist Impedance plot 400 for ten redox flow batteries, based on which the information in the following figures has been measured and / or extracted.

[0060] Figure 4b shows capacity measurements 401 for ten batteries based on which capacity prediction and remaining useful life prediction was performed. It is to be noted that some of the data was estimated based on known characteristics of redox flow batteries, given that some measurements were unstable after an extended period of time. In figure 4b, the normalised capacity is plotted over the cycle number for the ten batteries. The measurements were performed, in this example, until the capacity of the batteries dropped to 80% of their original values, which is considered, in this example, as the dead point for a battery.

[0061] Without wishing to be bound by theory, it is presumed that in particular the increase of the capacity for some of the batteries at around 3100 charging-discharging cycles may be due to fluctuations in charge and / or any inaccuracies of the capacity measurement system, not least due to breaks in between any measurements.

[0062] Figure 4c shows a comparison 402 between actual values and predicted values of the capacity of the measured batteries of figure 4b. In this example, a cross-validation for the ten batteries was performed and an average value of the capacity for the ten batteries is calculated to obtain the actual and the predicted capacity values plotted over the charging-discharging cycle numbers as depicted in figure 4c. One can see that, when using the method according to example implementations as described throughout the present disclosure, a value of R2of 0.99453 was obtained for the training dataset (using nine of the ten batteries) and a value of R2of 0.92931 was obtained for the test dataset (using the remaining battery of the ten batteries) in the method for predicting the capacity. These results show that the capacity (and thus the state of health, which may be defined based on a ratio between the capacity at a certain charging-discharging cycle and the initial capacity) of the batteries can be predicted using the method according to example implementations as described throughout the present disclosure with high accuracy.

[0063] Figure 4d shows the comparison 404 between the predicted normalised values of the capacity and the actual normalised values of the capacity, depicted in figure 4c. The straight line represents the curve y = x. This graph again shows high accuracy when predicting the capacity of the batteries using the method according to example implementations as described herein.

[0064] In the EIS measurements of the ten batteries, the frequency applied to the batteries was varied between 100 kHz and 100 mHz (with a precision in the nHz range). In figure 4e, the importance score (in arbitrary units) is plotted over specific frequencies, whereby frequency 1 amounts to 100 kHz and frequency 80 amounts to 100 mHz. The frequency x-axis is plotted using a logarithmic scale. The importance score hereby indicates how accurately the capacity was predicted for a particular frequency using methods according to example implementations as described herein.

[0065] As can be seen from figure 4e, the capacity can be predicted particularly accurately for specific frequencies for which the importance score is relatively high. In this example, the most accurate prediction of the capacity was obtained at a frequency of approximately 16.47 Hz (freq54). The frequencies which allow for a particularly accurate prediction of the capacity are the most important ones for the description of the remaining useful life. These particular frequencies relate to certain processes in the battery in terms of battery fading.

[0066] The result in figure 4e shows that one may not need to measure all frequencies over a range of frequencies to allow for an accurate prediction of the capacity, but one may only need to measure certain frequencies based on which the prediction may accurately be made. The result of figure 4e is depicted in figure 4f, where the features at the particular frequencies are plotted over the feature importance. In this example, one may use the 20 most important frequencies, or even fewer frequencies (e.g. ten frequencies, five frequencies, three frequencies, two frequencies, or even just the most important frequency), based on which the prediction of the capacity may be made.

[0067] Figure 4g shows graphs 410 of the actual remaining useful life (in arbitrary units) and the predicted remaining useful life (in arbitrary units), respectively, which are plotted over the numbers of charging-discharging cycles, for the average of the ten batteries. In this example, the remaining useful life corresponds to the number of charging-discharging cycles after which the battery reaches the dead point, i.e. where the capacity drops to 80% of its initial value.

[0068] One can see that, when using the method according to example implementations as described throughout the present disclosure, a value of R2of 0.99781 was obtained for the training dataset (using nine of the ten batteries) and a value of R2of 0.95072 was obtained for the test dataset (using the remaining battery of the ten batteries) in the method for predicting the remaining useful life. These results show that the remaining useful life of the batteries can be predicted using the method according to example implementations as described throughout the present disclosure with high accuracy.

[0069] As was the case for the prediction of the capacity, for the remaining useful life prediction, it was found that some frequencies are more important than other frequencies (see figure 4h, where the most important frequencies are plotted over the feature importance in arbitrary units). The corresponding feature importance scatter plot 414 is shown in figure 4i. Therefore, one may not need to measure all frequencies over a range of frequencies to allow for an accurate prediction of the remaining useful life, but one may only need to measure certain frequencies (e.g. 20 frequencies, ten frequencies, five frequencies, three frequencies, two frequencies, or even just the most important frequency (in this example 403.065 Hz)) based on which the prediction may accurately be made.

[0070] Figure 4j shows a graph 416 in which the relative feature (frequency) importance (in arbitrary units) is plotted over the feature frequency in Hz. The corresponding Bode plots 418 are depicted in figure 4k. One can see that, in this example, the most important features are found around 2 Hz, 20 Hz and 600 Hz for the prediction of the remaining useful life. The importance of features around 2 Hz finds its origin in transport properties of the species, for example based on a change in the cell, while interfacial properties play a role in frequencies around 20 Hz. Within this frequency range from 2 Hz to 20 Hz, Faraday reactions may occur at the electrodes, resulting in these approximate frequencies being particularly important when predicting the remaining useful life. Around 600 Hz, the importance of such frequencies for the prediction of the remaining useful life stems from surface properties of the electrodes, which may degrade with an increasing number of charging-discharging cycles.

[0071] Figure 5 shows a flow diagram of a method 500 according to some example implementations as described herein.

[0072] In this example, the method 500 for predicting a state of health, SOH, of a redox flow battery comprises, at step S502, receiving battery data of the redox flow battery. The battery data may comprise, in some examples, EIS spectra of the redox flow battery.

[0073] At step S504, the method 500 comprises performing, using a machine learning agent, an analysis of the battery data to predict the SOH of the redox flow battery.

[0074] In this example, the method 500 further comprises, at step S506, predicting, based on the analysis of the battery data by the machine learning agent, a remaining useful life, RUL, of the redox flow battery. The end of the RUL may hereby be defined as the capacity of the redox flow battery dropping to 80% of its initial value. As will be appreciated, other values may be used to define the end of life of the redox flow battery. The RUL may thus, in some examples, be defined as a difference between the total number of (e.g. full) charge-discharge cycles when the actual capacity of the redox flow battery drops to the threshold value (e.g. 80% or another value) and the number of charge-discharge cycles of the current redox flow battery.

[0075] In this example, the method 500 further comprises, at step S508, displaying the predicted SOH of the redox flow battery on a display (e.g. the display module 108 shown in figure 1). The display may hereby display, for example, the capacity state of the redox flow battery. Figure 6 shows a flow diagram of a further method 600 of managing a redox flow battery according to some example implementations as described herein.

[0076] In this example, the method 600 comprises predicting a state of health, SOH, of the redox flow battery using the method 500 as shown in figure 5. At step S602, the method 600 comprises managing the redox flow battery based on the precited SOH. The managing of the redox flow battery comprises, in some examples, in particular monitoring a capacity state of the redox flow battery.

[0077] Figure 7 shows a flow diagram of a further method 700 for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery according to some example implementations as described herein.

[0078] At step S702, the method 700 comprises providing or obtaining a training data set comprising (i) an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery and (ii) a capacity and / or a RUL of the redox flow battery and / or the other redox flow battery. At step S704, the method 700 comprises training the machine learning agent using the training data set comprising the electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or the other redox flow battery as input data and the capacity and / or RUL of the redox flow battery and / or the other redox flow battery as output data. The machine learning agent may then be used for predicting the SOH and / or remaining useful life of a redox flow battery.

[0079] Figures 8a and b show schematic block diagrams of computing units 800 and 806 according to some example implementations as described herein.

[0080] Figure 8a schematically illustrates a computing unit 800 for predicting a SOH of a redox flow battery. The computing unit 800 comprises, in this example, at least one processor 802 and at least one memory 804, wherein the at least one memory 804 contains instructions executable by the at least one processor 802 such that the computing unit 800 is operable to perform the method of predicting a SOH of a redox flow battery as outlined throughout the present disclosure according to any one or more example implementations.

[0081] Figure 8b schematically illustrates a computing unit 806 for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery. The computing unit 806 comprises, in this example, at least one processor 808 and at least one memory 810, wherein the at least one memory 810 contains instructions executable by the at least one processor 808 such that the computing unit 806 is operable to perform the method for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery as outlined throughout the present disclosure according to any one or more example implementations.

[0082] The following examples are also encompassed by the present disclosure and may fully or partly be incorporated into embodiments:

[0083] 1. A centralized machine learning system for predicting the state of health (online capacity) and remaining useful life of redox flow batteries, comprising: a sampling module configured to receive and process battery data, a control module configured to analyze the battery data and train a machine learning model to make the prediction, a display module configured to visualize the prediction.

[0084] 2. The centralized machine learning system of example 1, wherein the battery data received by sampling module is the electrochemical impedance spectrum (EIS) of redox flow batteries.

[0085] 3. The centralized machine learning system of example 1 or 2, wherein an entire EIS spectrum over a wide range of frequency (e.g. 1 mHz to 1 or several MHz) is periodically collected after fully charging and discharging conditions.

[0086] 4. The centralized machine learning system of any one of examples 1 to 3, wherein the sampling module includes a plurality of agents, each agent including a controller configured to save EIS data from a training dataset, receive EIS data from the new redox flow battery system and periodically transmit the EIS data to the machine learning model.

[0087] 5. The centralized machine learning system of any one of examples 1 to 4, wherein the machine learning model estimates the online capacity and predicts the remaining useful life. 6. The centralized machine learning system of any one of examples 1 to 5, wherein the control module includes a plurality of agents, each agent including a controller configured to record the machine learning model trained by temporary data, update the machine learning model and transmit the prediction of the machine learning model to display module.

[0088] 7. The centralized machine learning system of any one of examples 1 to 6, wherein the machine learning model is an artificial recurrent neural network architecture used in the field of deep learning, the method further comprising training, validating and predicting, using one controller to transmit data to display module.

[0089] 8. The centralized machine learning system of any one of examples 1 to 7, wherein the display module shows the online capacity and remaining useful life of the redox battery systems.

[0090] 9. The centralized machine learning system of any one of examples 1 to 8, wherein the machine learning model can run in real-time as part of the battery management system.

[0091] 10. Analyzing the estimation of online capacity and the prediction of remaining useful life enables diagnosis of redox flow battery issues months in advance.

[0092] It will be appreciated that the present disclosure has been described with reference to exemplary embodiments that may be varied in many aspects. As such, the present invention is only limited by the claims that follow.

Claims

Claims1. A method for predicting a state of health, SOH, of a redox flow battery, the method comprising: receiving battery data of the redox flow battery; and performing, using a machine learning agent, an analysis of the battery data to predict the SOH of the redox flow battery.

2. A method as claimed in claim 1, wherein the machine learning agent is trained using a training data set comprising an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery as input data and a capacity of the redox flow battery and / or the other redox flow battery as output data.

3. A method as claimed in claim 2, wherein the training data set is obtained by: measuring the capacity and performing the EIS measurement to obtain capacity data and corresponding EIS data, respectively, of the redox flow battery and / or the other redox flow battery; providing the capacity data and the corresponding EIS data to the machine learning agent; and fitting, by the machine learning agent, a model to the capacity data and the corresponding EIS data.

4. A method as claimed in claim 3, wherein the measurement of the capacity and the performing of the EIS measurement are repeated for a plurality of chargedischarge cycles until the capacity diminishes to a predetermined percentage, in particular 90 percent, of an original value of the capacity, wherein the capacity data and the corresponding EIS data are provided to the machine learning agent for each of the charge-discharge cycles, and wherein the machine learning agent fits the model to the capacity data and the corresponding EIS data obtained for all chargedischarge cycles.

5. A method as claimed in claim 3 or 4, wherein said performing, using the machine learning agent, the analysis of the battery data to predict the SOH of the redox flow battery comprises: providing the battery data to the machine learning agent, andpredicting the SOH based on the model.

6. A method as claimed in any one of the preceding claims, wherein the machine learning agent is trained using a further training data set comprising an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery as input data and a remaining useful life, RUL, of the redox flow battery and / or the other redox flow battery as output data.

7. A method as claimed in claim 6, wherein the further training data set is obtained by: measuring the capacity and performing the EIS measurement to obtain RUL data and corresponding EIS data, respectively, of the redox flow battery and / or the other redox flow battery; providing the RUL data and the corresponding EIS data to the machine learning agent; and fitting, by the machine learning agent, a further model to the RUL data and the corresponding EIS data.

8. A method as claimed in claim 7, wherein the measurement of the capacity and the performing of the EIS measurement are repeated for a plurality of chargedischarge cycles until the capacity diminishes to a predetermined percentage, in particular 90 percent, of an original value of the capacity, wherein the RUL data and the corresponding EIS data are provided to the machine learning agent for each of the charge-discharge cycles, and wherein the machine learning agent fits the model to the RUL data and the corresponding EIS data obtained for all charge-discharge cycles.

9. A method as claimed in claim 8, wherein, for each charge-discharge cycle, the RUL is calculated by((Total life cycles-passed cycles / Total life cycles)*100).

10. A method as claimed in any one of claims 7 to 9, wherein said performing, using the machine learning agent, the analysis of the battery data to predict the SOH of the redox flow battery comprises: providing the battery data to the machine learning agent, and predicting the SOH based on the further model.

11. A method as claimed in any one of claims 3 to 5 in combination with any one of claims 7 to 10, wherein the model which is fitted to the capacity data and the corresponding EIS data and the further model which is fitted to the RUL data and the corresponding EIS data are combined into a single model.

12. A method as claimed in any preceding claim, wherein the battery data comprises an electrochemical impedance spectrum, EIS, of the redox flow battery for which the SOH is predicted.

13. A method as claimed in claim 12, wherein the EIS of the redox flow battery is obtained for a plurality of frequencies.

14. A method as claimed in claim 13, wherein the frequencies range from 1 mHz to n MHz, wherein n e [1, 2, ..., 10].

15. A method as claimed in claim 13 or 14, wherein the EIS of the redox flow battery is obtained after fully charging and discharging the redox flow battery, in particular obtained periodically after fully charging and discharging the redox flow battery.

16. A method as claimed in any preceding claim, wherein receiving battery data of the redox flow battery comprises performing an EIS measurement on the redox flow battery.

17. A method as claimed in any preceding claim, further comprising predicting, based on the analysis of the battery data by the machine learning agent, a remaining useful life, RUL, of the redox flow battery.

18. A method as claimed in claim 17, wherein the RUL is calculated by((Total life cycles-passed cycles / Total life cycles)*100).

19. A method as claimed in any preceding claim, in combination with claim 12, wherein predicting the SOH of the redox flow battery comprises predicting a capacity of the redox flow battery based on the analysis, via the machine learning agent, of the EIS.

20. A method as claimed in claim 19, wherein the prediction of the capacity is further based on a value of the capacity of the redox flow battery obtained at a specific point in time in the past.

21. A method as claimed in any preceding claim, wherein the analysis is performed, via the machine learning agent, based on a comparison between the battery data and previously obtained battery data of the redox flow battery and / or another redox flow battery.

22. A method as claimed in claim 21, in combination with claim 12, wherein the comparison comprises analysing, via the machine learning agent, a deviation of the EIS of the redox flow battery from a previously obtained EIS of the redox flow battery and / or the other redox flow battery.

23. A method as claimed in any preceding claim, wherein the machine learning agent is an artificial recurrent neural network agent.

24. A method as claimed in any preceding claim, wherein the machine learning agent is trained based on input data comprising battery data of the redox flow battery and / or another redox flow battery for which impurities have been introduced, at least for a fraction of time during which the battery data has been obtained, into a fluid circulated in the redox flow battery and / or the other redox flow battery.

25. A method as claimed in any preceding claim, wherein the machine learning agent is comprised in a centralised machine learning system.

26. A method as claimed in any preceding claim, further comprising displaying the predicted SOH of the redox flow battery on a display.

27. A method as claimed in any preceding claim, wherein the battery data comprises an EIS measurement of the redox flow battery and / or another redox flow battery performed at a plurality of frequencies, wherein the method further comprises determining one or more frequencies of the plurality of frequencies for which an importance for the prediction of the SOH of the redox flow battery and / or a capacity of the redox flow battery and / or a remaining useful life of the redox flow battery is higher compared to an importance of one or more remaining frequencies of the plurality of frequencies for the prediction of the SOH of the redox flow batteryand / or the capacity of the redox flow battery and / or the remaining useful life of the redox flow battery.

28. A method as claimed in claim 27, wherein performing the analysis of the battery data to predict the SOH of the redox flow battery is based on the EIS measurement performed at the one or more determined frequencies.

29. A method as claimed in claim 27 or 28, further comprising predicting the capacity of the redox flow battery and / or the remaining useful life of the redox flow battery based on the EIS measurement performed at the one or more determined frequencies.

30. A method of managing a redox flow battery, the method comprising: predicting a SOH of the redox flow battery according to the method as claimed in any preceding claim; and managing the redox flow battery based on the precited SOH.

31. A method as claimed in claim 30, wherein said managing of the redox flow battery comprises monitoring a capacity state of the redox flow battery.

32. A method for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery, the method comprising: providing or obtaining a training data set comprising (i) an electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or another redox flow battery and (ii) a capacity and / or a remaining useful life, RUL, of the redox flow battery and / or the other redox flow battery; and training the machine learning agent using the training data set comprising the electrochemical impedance spectrum, EIS, measurement of the redox flow battery and / or the other redox flow battery as input data and the capacity and / or RUL of the redox flow battery and / or the other redox flow battery as output data.

33. A computer program product comprising program code portions for performing the method of any one of claims 1 to 32 when the computer program product is executed on one or more computing devices.

34. The computer program product of claim 33, stored on a computer-readable recording medium.

35. A computing unit for predicting a SOH of a redox flow battery, the computing unit comprising at least one processor and at least one memory, the at least one memory containing instructions executable by the at least one processor such that the computing unit is operable to perform the method of any one of claims 1 to 31.

36. A computing unit for training a machine learning agent configured to perform an analysis of battery data of a redox flow battery to predict a SOH of the redox flow battery, the computing unit comprising at least one processor and at least one memory, the at least one memory containing instructions executable by the at least one processor such that the computing unit is operable to perform the method of claim 32.

37. A system comprising a computing unit according to claim 35 and, optionally, a computing unit according to claim 36.

Images