Abstract
The open circuit voltage hysteresis of lithiumion batteries is a phenomenon that, despite intensive research, is still not fully understood. However, it must be taken into account for accurate stateofcharge estimation in battery management systems. Mechanistic models of the open circuit voltage hysteresis previously published are not suitable for deployment in a battery management system. Phenomenological models on the other hand can only superficially represent the processes taking place. To address this limitation, we propose a probability distributed equivalent circuit model motivated by the physical insights into hysteresis. The model incorporates hysteresis effects that are often disregarded for state estimation, while keeping the computational cost low. Although the parameterization is more demanding, the model has the advantage of providing insight into the internal state of the battery and intrinsically incorporating the effect of pathdependent rate capability.
Similar content being viewed by others
Introduction
Lithiumion batteries (LIBs) are used in portable devices, stationary battery energy storage systems, and battery electric vehicles. Accurate knowledge of the current state of charge is essential for safe and efficient operation^{1,2,3}. In the battery management system the stateofcharge (SOC) is often determined by Coulomb counting, i.e., by measuring and integrating the current at the terminals of the battery. This method has the disadvantage that measurement errors cumulate over the time the system is used^{1}. To correct these errors, the cell voltage can be used, as the open circuit voltage (OCV) depends on the SOC. Unfortunately, the OCV can also depend on the charge and discharge history; this particular type of path dependency phenomenon is known as voltage hysteresis^{4}. For a correct SOC estimation it is therefore necessary to consider the voltage hysteresis.
Voltage hysteresis is a shortterm pathdependent phenomenon^{4}, as the OCV depends on the shortterm history of the applied current. However, even after very long relaxation, i.e., an operation with no current applied, voltage hysteresis does not vanish^{5,6,7}. Voltage hysteresis is strongly dependent on the active material. For example, graphite with ~10 mV^{8}, lithium iron phosphate (LFP) with up to 20 mV^{5} and silicon (Si)^{9} with more than 200 mV are known to have pronounced voltage hysteresis, while lithiumtitanate (LTO) shows negligible hysteresis. Moreover, the underlying mechanism may be different, e.g., firstorder phase transition for LFP^{10,11} or charge direction dependent reaction pathways for Si^{12,13}. Hysteresis can also be visualized, for example, by the color of graphite particles, which has been shown to be dependent on current direction^{14}. In addition, voltage hysteresis can be related to other pathdependent phenomena such as the path dependency of cell expansion^{15} as well as rate and power capability^{16,17,18}. Therefore, limiting the assessment to the description of voltage hysteresis may not be sufficient.
To account for the hysteresis of the OCV and thus improve the accuracy of the SOC estimation, a mathematical description of the OCV as a function of the current direction and the stored amount of charge relative to the cell’s capacity is required. Many approaches have been described in the literature, some of which are generally applicable and some of which are specifically designed for certain cell chemistries. Plett discusses state estimation approaches in detail in his work^{19,20,21}. He extends an equivalent circuit model (ECM) for describing the dynamic hysteresis behavior of LIBs with a zerostate response model, where a switch between charge and discharge is considered by adding or subtracting the voltage hysteresis to the OCV equation, respectively. Plett extends this approach by adding a single hysteresis state^{20}. In this model, the voltage hysteresis is accounted for by an additional differential equation. Motivated by hysteresis models in magnetic fields, an adaptation of the Preisach model for LFP is proposed by Tjandra and Jossen^{22} and Baronti et al.^{23}. Some publications developed mathematical functions to model the voltage hysteresis of LFP, especially the transition paths between charge and discharge curve^{24,25}. Other groups have also used a long shortterm memory neural network to learn models that can describe voltage hysteresis of LFP^{26,27}. For materials with very pronounced hysteresis such as Si, more specialized hysteresis models were used. A PrandtlIshlinskii model, based on Preisach modeling, was proposed by Chayratsami et al.^{28}. Baker et al.^{29} developed mathematical formulae to describe the hysteresis of Si, that were later applied by Graells et al.^{30} for the voltage prediction of Si/graphite halfcells. Both models have additional degrees of freedom to design the transition path of the voltage hysteresis, describing the large and lithiationdependent hysteresis path of Si in more detail.
The models discussed so far are useful for the application in battery management systems (BMS), but they are only phenomenological descriptions of the experimentally measured behavior. Mechanistic models are an alternative, but are often associated with high computational cost, which makes their application in BMS difficult, though not impossible^{31}. Farkhondeh et al.^{32} proposed a mesoscopic model consisting of multiple units with an underlying nonmonotonous equilibrium potential and a resistance distribution. A similar approach was followed by Kondo et al.^{33} using multiparticle model with a particle size distribution. Jiang^{13} simulated crystallization and reaction pathways for Si to predict voltage hysteresis. These models provide a deeper insight into the internal state of the cell, e.g., the particlebyparticle phase change. In addition, these models have been shown to predict not only the voltage hysteresis, but also the path dependency of the rate and power capability. As prediction of rate and power capability is another important task for the BMS, these mechanistic models provide an important added value. Furthermore, the estimation of the remaining useful lifetime of the cell is another important topic for battery management, where hysteresis of OCV and impedance are of relevance^{34,35}. However, a significant reduction in the computational cost is required for application. Additionally, mechanistic models are usually limited to certain electrode materials.
In this article, we propose a modeling approach based on a classical equivalent circuit of electrical networks, which, due to its mathematical structure, is applicable in BMS. The model is characterized by distributing the cell capacity into parallel strings to account for stringbystring charging, analogous to the particlebyparticle charging in frequently reported mechanistic models. Thus, this model provides detailed information on the internal state of the voltage hysteresis and the path dependency of the rate and power capability. The described structure is based the mechanism of materials that undergo a firstorder phase transition. Therefore, the model is not suitable for other materials that utilize different lithiumstorage mechanisms, such as silicon. To the best of the authors’ knowledge, this is the first time that the physical principles of hysteresis in materials such as LFP have been modeled using an ECM. The model reaches realtime capability and enables an easier integration of the frequently overlooked impact of rate capability hysteresis into BMS. In this article, the operating principle of this novel hysteresis model is explained, the effect of parameter variations is investigated and the model is applied to simulate the voltage curve of a commercial cell. In addition, the results are compared with wellestablished hysteresis models based on an ECM from the literature.
Results and discussion
The proposed hysteresis model is suitable for battery materials where the hysteresis is caused by a firstorder phase transition. In this case, the Gibbs free energy has two minima, so there are two stable phases in the twophase region, a lithiumpoor phase (α) and a lithiumrich phase (β)^{36}. The electrode potential is proportional to the change in Gibbs free energy during insertion/extraction of lithium ions, so the open circuit potential (OCP) without phase separation is a nonmonotonic function of the degree of lithiation^{10}. The OCP has a local maximum at a slightly higher lithiation state than the αphase and a local minimum at a slightly lower degree of lithiation than the βphase. In the case of phase separation, the α and β phases can coexist in equilibrium because they have the same potential. In a multiparticle system this can lead to a particlebyparticle phase change during lithiation/delithiation. During lithiation, all particles are filled with lithium until they reach the local maximum of the nonmonotonic potential curve. As soon as the first particle crosses this maximum, it is lithiated until it reaches a stable state in the Lirich phase with the same potential. This leads to sequential lithiation of the particles. The potential of the multiparticle system with phase separation is the horizontal line at the local maximum of the nonmonotonic potential curve. During delithiation, the potential of the multiparticle system with phase separation is again a horizontal line, now starting from the local minimum of the nonmonotonic potential curve^{5}. The difference of the maximum and minimum is observed as voltage hysteresis. The sequential order of particle lithiation/delithiation depends on the overpotentials of the individual particles. For example, poorly connected particles have a higher ohmic resistance and will therefore be lithiated/delithiated later.
The hysteresis theory described above is now transferred into a physically motivated ECM. Therefore, a nonmonotonic OCP, i.e., OCP_{n}, and a distribution of cell capacity in multiple parallel strings are considered, which allows to model the voltage hysteresis of phaseseparating materials such as LFP. In addition, a distribution of electrical resistances within the multiple strings is used. The detailed description of the model, its mathematical representation, and the solving algorithm are given in the Methods section. The model is used to simulate three basic tests, (1) a pseudo OCV test, (2) a pseudo OCV test with partial cycles and (3) a path dependency test for rate capability. The tests are also described in detail in the Methods section. The model is parameterized with experimental data and simulation errors are analyzed and compared with another hysteresis models from the literature.
Characteristics of the probability distributed ECM
Figure 1 shows simulation results for the pseudo OCV test. An ECM with only N = 10 parallel strings and normally distributed resistance values is used to illustrate the model behavior. The underlying nonmonotonous curve is shown in Fig. 1a. It has a maximum at a SOC of 0.4 and a minimum at a SOC of 0.7. The simulated voltage is shown as the green line in Fig. 1a. Voltage versus time and applied Crate is shown in Fig. 1b. It can be seen that the voltage of the pseudo OCV test differs from the considered OCP_{n}. For the charging phase, the voltage remains almost constant after the maximum of the OCP_{n} at 40% SOC has been exceeded. During the subsequent discharge, the voltage shows a similar behavior, this time though after passing the local minimum of the OCP_{n} at 70% SOC. The model can therefore represent the voltage hysteresis. However, it can also be seen that there are significant ripples in the voltage as it moves horizontally.
For a deeper insight into the characteristics of the model, the course of the OCP_{n} and the SOC within the parallel strings, called SOC_{n}, are shown over time in Fig. 1c, d, respectively. The number of parallel strings is indicated in the color bar, where larger string numbers n, correspond to a higher assigned resistance value r_{n}. In Fig. 1d, initially the SOC_{n} almost simultaneously increases for all parallel strings. Slight differences can be seen, as strings with a smaller assigned resistance have a slightly higher SOC_{n}. Therefore, the parallel string with the smallest resistance, i.e., n = 1, reaches the SOC_{n} with the local maximum of the potential curve first. As soon as a string exceeds the SOC_{n} of 0.4 the SOC_{n} quickly rises to 0.8. Thus, a sequential charging of the parallel strings, i.e., stringbystring charging, is observed. The sequential order is determined by the assigned resistance, while strings with lower resistance are charged earlier. Figure 1c shows, that as soon as the first string crosses the local maximum, the OCP_{n} in this string is decreased, leading to a fast charging of a single string. Furthermore, in Fig. 1d it becomes apparent that during these phases, the SOC_{n} of all the other strings decreases slightly, indicating that the charge is mainly redistributed and supplied by the other parallel strings. This behavior also explains the observed ripples in the simulated voltages as it is associated with significant changes of OCP_{n} and current within the parallel strings. A similar behavior is observed during discharge, with stringbystring discharge starting at the potential of the local minimum of the OCP_{n}. Overall, stringbystring charging and discharging results in a horizontal voltage curve. The voltage is determined by the direction of the applied current, showing that the proposed ECM is capable of modeling the voltage hysteresis and representing the system state through the SOC_{n} within the considered parallel strings.
Effects of parameter variation
The probability distributed equivalent circuit model (PDECM) can be adjusted in a number of ways. These include the definition of the OCP_{n} function as well as the number, the assigned resistance, resistance range, and shape of the capacity distribution function of the parallel strings.
A larger number of parallel strings results in a smaller capacity in each string. One effect of an increasing number of parallel strings is that the simulated voltage in the phase transition region becomes smoother. This can be seen by comparing Fig. 2a with Fig. 1a. The simulation in Fig. 2a uses a PDECM with 50 parallel strings, which significantly reduces the height of the voltage ripples compared to the results shown in Fig. 1a where a PDECM with only 10 parallel strings is used. However, since the number of states in the statespace representation increases with the number of strings, the computational cost also increases with the square of N (see the Methods section).
Simulation results for the second test, i.e., pseudo OCV test with partial cycles, are shown in Fig. 2b. The test includes partial cycles at low rates between 0% and three different upper SOCs, i.e., 75, 60 and 45%. The simulation results show the expected behavior for the voltage transition when switching from charge to discharge, as a transition between the upper and lower voltage of the full cycle is observed. While phenomenological models implement this behavior explicitly by using inverse exponential functions^{20} or quadratic functions^{25}, this is achieved implicitly here, since in the PDECM the voltage transition depends on the state of the parallel strings. Moreover, it can be seen that the slope of the voltage transition depends on the depth of the partial cycle. For example, the voltage transition for the 75% SOC partial cycle is significantly steeper in the beginning compared to the other partial cycles. Similar changes in the voltage transition are reported for experiments in the literature^{37}, indicating an advantage of this model approach.
Figure 2c shows the effect of hysteresis on the path dependency test for rate capability. The applied current is shown in Fig. 2d. The rate capability of a cell is determined by the amount of charge or energy that can be extracted from a cell for higher rates. The rate capability is affected by the internal resistance of the cell. To test the effect of path dependency, a constant current discharge of 50% of the cell’s capacity is performed for two cases. First, after the cell has been charged to 0.5 SOC. Second, after the cell has been discharged to 0.5 SOC. The first case is shown in Fig. 2c with a dashed line. The second case is shown with a dotted line. The solid line shows the simulated voltage curve for the charge and discharge phases at 1C. Comparing the two cases, it can be seen that the voltage is lower in the second case. This also leads to a decrease in the energy of the discharge step and indicates a dependency of the rate capability on operation history. This path dependency phenomenon is often reported for LFP cells^{16,17,18} and can be represented by the PDECM, highlighting another advantage of this model over phenomenological approaches.
The reason for the observed path dependency of the rate capability can be illustrated using Fig. 2e, which shows the SOC_{n} of the parallel strings. The color of the graphs represents the string number and corresponds to the resistance of the string, with blue corresponding to the lowest resistance and pink to the highest. In the relaxation just before the two highrate discharge phases, clear differences in the system state can be observed. For the first discharge, the strings with the lowest resistance are at high SOC_{n}. In contrast, for the second discharge, the strings with the highest resistance are at high SOC_{n}. Therefore, in the first case during discharge, low resistance strings are discharged first until all strings reach a common SOC at the local maximum of OCP_{n}. Then all strings are discharged together until they reach 0% SOC. In contrast, in the second case, the strings with the highest resistance are at the higher SOC_{n} and the strings with low resistance are already at low SOC_{n}. As a result, during discharge, the current flows mainly through the highresistance strings, resulting in higher polarization, which explains the lower voltage observed. These results show that once the cell is in the hysteresis region, the strings with the lowest resistance are charged/discharged first. The PDECM can represent this dependency of the history of operation by modeling the state in the parallel strings instead of using a lumped SOC of the whole system. As a lumped SOC is common in conventional ECM, they cannot represent this phenomenon.
In several publications, e.g., by Li et al.^{38} or Bai et al.^{39}, it is shown, that the phase separation of LFP can be suppressed under certain circumstances. For nanoparticles charged with high current rates the lithiated iron phosphate changes from particletoparticle to a concurrent lithiation, forming a quasisolidsolution or solidsolution. As shown by Katrasnik et al.^{40} this leads to very specific behavior during pulse load and the subsequent low load or relaxation phase. The PDECM does not incorporate mechanisms aimed to simulate the transition from phase separation to solid solution. However, the behavior shown by Katrasnik et al.^{40} that is attributed to the forming of a quasisolidsolution, can nevertheless be modeled using the PDECM without changes in the model structure, as shown in the Supplementary Material (Supplementary Fig. 1). In the PDECM the change from particletoparticle to a concurrent (de)lithiation caused by a high load pulse can be explained by the establishment of a new multiparticle equilibrium, due to the increased voltage drop at the network of resistors. To which extend the results of Katrasnik et al.^{40} can be explained by a multiparticle equilibrium and to which by forming of a quasisolidsolution is not scope of this work and remains a topic for future research.
In the following, the influence of three model parameters are investigated:

(1)
the resulting ohmic resistance of the network \({R}_{{{{{{{{\rm{res}}}}}}}}}\).

(2)
the range of the normalized resistances ΔR_{norm}.

(3)
the function for the probability density of the discrete resistances p(r_{n}).
For this purpose, these parameters are varied and the results are presented for the path dependency test of the rate capability in Fig. 3. To examine the influence of the shape of the probability distribution function, the Gaussian distribution is replaced by a Weibull distribution, where the parameter k modifies the shape of the distribution while maintaining the distribution width. The parameter values listed in Table 1 are used for the simulations. The analysis is conducted by changing one parameter through all listed values, while the rest is kept at the reference value.
Figure 3a shows simulation results for the variation of the resulting ohmic resistance of the ECM. As can be seen, the polarization increases with higher \({R}_{{{{{{{{\rm{res}}}}}}}}}\) over the whole SOC range. The differences between the simulated voltage curves are slightly higher at higher SOCs, emphasizing that the strings assigned with a high resistance are utilized last. For the highrate discharges, it can be seen that the polarization is similarly affected by the higher \({R}_{{{{{{{{\rm{res}}}}}}}}}\) for both discharges.
This is in contrast to the influence of the range of normalized resistances ΔR_{norm} shown in Fig. 3b. Here, the effects are predominantly visible during the highrate discharge phases. The increased inhomogeneity in the distribution of resistances leads to a significant reduction in rate capability during the second highrate discharge. For the first highrate discharge the effects are less significant, since the strings with low resistances are discharged. Consequently, the effect of polarization differences is less significant.
The effect of the shape of the distribution function is shown in Fig. 3c. The corresponding shares of capacity with the corresponding resistance values r_{n} are shown in Fig. 3d. The visible changes in area covered by the bar plots is caused by changing the distribution. For example, for k = 1 a large contribution of smaller resistance values needs to be balanced by few large resistors. This leads to a larger range of r_{n} in the parallel network to keep \({R}_{{{{{{{{\rm{res}}}}}}}}}\) constant. The sum of shares of capacity over all resistance remains one. Since the resulting ohmic resistance of the PDECM is kept constant, comparable overpotentials are observed at the beginning of each charge or discharge process. As can be seen, the distributions with smaller shape parameters, i.e., a wide distribution of the highresistance strings, show an increasing polarization when the high resistivity strings are used, i.e., at the end of the charging process. A similar effect can be observed for the highrate discharges. The simulations with smaller k show a slightly lower polarization during the first discharge due to the higher proportion of low resistance strings. For the second discharge curve a significant influence can be observed. Here, a smaller k leads to significantly increased polarization and thus to a loss of dischargeable energy.
Overall, the parameter variation shows that the model is not only able to simulate the voltage hysteresis during a firstorder phase transition, but also allows the parameters to be adjusted to modify the path dependency of the rate capability.
Halfcell simulation
The next step is setting up the simulation of a cylindrical, commercial A123 battery with LFP cathode and graphite anode using the PDECM. The hysteresis of the fullcell depends on the hysteresis of each of the two electrodes. Therefore, each electrode must be simulated separately. To adapt the model parameters to the halfcells, the cylindrical A123 cell is opened, and the electrodes are extracted. These are then used in halfcells to measure the OCV of each electrode. The OCP_{n} of each electrode cannot be measured directly and therefore must be approximated and parameterized. The whole procedure is described in the Methods section.
The resulting lowrate full cycle for the graphite halfcell with the according error is plotted in Fig. 4a, c, while the LFP halfcell with error is shown in Fig. 4b, d.
As can be seen, the model can be used to simulate the pseudo OCV test including the voltage hysteresis for both electrodes. For both electrodes the error is less than 5 mV for most of the SOC range. However, for LFP there is a significant deviation at very high and very low SOCs. In this range the measured voltage hysteresis is underestimated in the model. This suggests a limitation of the hysteresis model, where the slopes at beginning and end of the plateau are not just implemented but rather a result of the chosen network and the underlying OCP_{n}. However, these observations could also be caused by residual overpotentials in the pseudo OCV measurement that are not correctly accounted for in the model. These could be changes in the charge transfer resistance or diffusion overpotentials not considered in the halfcell simulation using the PDECM.
Fullcell simulation
For the fullcell simulation the halfcells are simulated separately and the results are subtracted. In addition, a serial connection of two RC elements is added to the model to account for further polarization. The ohmic resistor connected in serial that is commonly used in ECM for batteries is replaced by the distributed resistors in the PDECM. These further model modifications and model parameterization are described in the Methods section.
To evaluate the performance of the model, the results are compared with an established hysteresis model from the literature. The commercial LFP cell from A123 is used for the measurement. Figure 5 shows the results for three current profiles. Figure 5a shows the results for the pseudo OCV test. Figure 5b, c depicts the results for the path dependency test for rate capability. The corresponding errors of the simulated voltages are shown in Fig. 5d–f. The measured current is used as the input for the simulations. The simulation results of the PDECM are compared with the singlestate hysteresis model proposed by Plett^{20}. The implementation and parameterization of the singlestate hysteresis model is described in the Methods section. The model uses the same SOCdependent parameters for the preconnected RCmodel as the PDECM and differs in the modeling of the voltage hysteresis and an additionally implemented serial resistor. The results of additional simulations of partial cycles with lowrate are given in the Supplementary Methods in Supplementary Fig. 2.
The pseudo OCV test and the corresponding errors of both models are shown in Fig. 5a, d respectively. As can be seen, the error of the singlestate hysteresis model is with a relative root mean square error (RMSE) of 0.432% only half of the RMSE for the PDECM with 0.9323%. This is not surprising as the measured data from the pseudo OCV test is fed directly into the model as the upper and lower OCV curves. The visible errors are only caused by polarization errors, which are very small for the low Crates used. Nevertheless, the PDECM results are of sufficient precision, when comparing with the reported RMSE for the simulation of LFP using a Preisach model by Baronit et al.^{23} ranging from 0.6% to 2.5%. Still, the error for the fullcell simulation is significantly higher than the sum of the errors of the halfcell measurement. This indicates a problem arising from the scaling from halfcell to fullcell measurement. When taking a closer look at the plateaus following the stage transition of graphite two different slopes for the PDECM simulation and the fullcell measurement are visible. This may be an indicator for more inhomogeneous lithiation in the fullcell compared to the halfcell. Due to the parameterization of the PDECM using halfcells these inhomogeneities are not taken into account. This problem is avoided in the singlestate hysteresis model as it can be directly parameterized with full cell measurements.
The results of the rate capability tests are shown in Fig. 5b, c with the corresponding errors in 5e, f. For the first test the 0.5 SOC was set from a fully discharged state (Fig. 5b), while for the second test the 0.5 SOC was set from a fully charged state (Fig. 5c). In both cases the analyzed models have comparable errors during the 1C charge or discharge phases. However, during the highrate discharge the PDECM shows an advantage. The singlestate hysteresis underestimates the voltage in the first test and overestimates the voltage in the second test. The PDECM can account for the path dependency of the rate capability and follows the slope of the measured voltage better. This also results in smaller errors at the beginning of the discharge. However, at the end of the discharge, both models have significant errors, which may be due to insufficient modeling of the cell polarization. The RMSE of the PDECM simulation for the rate capability test on the charge curve is with 2.856% and with 1.97% for the rate capability test on the discharge curve significantly higher than the slow rate simulation. Comparing these to dynamic tests and ECM simulation conducted by Guenther et al.^{25}, who reach a median deviation of 0.06%, shows worse final results for the PDECM. Nevertheless, the simulation error for the dynamic test cases is still comparable to the simulation of Baronti et al.^{23}. Since Guenther et al. did not simulate the edge regions of the SOC, where the largest model errors are occurring, this may explain the different results.
Conclusion
The presented results show the capability of the probability distributed equivalent circuit model comprising a nonmonotonous potential curve OCP_{n}, controlled by the state of charge in each string SOC_{n}, to simulate the hysteresis behavior of materials undergoing a firstorder phase transition. For the first time, the physical principles of hysteresis were successfully transferred into an ECM. The additional model parameters allow for a detailed tuning of the voltage behavior under load, including hysteresis effects often neglected in conventional modeling approaches for BMS application, such as rate capability and power prediction. The model design shows some disadvantages as well: a low number of simulated parallel strings lead to high ripples in the simulated voltage curve. Moreover, the system does not prove stable for all conditions, especially for high values of ΔR_{norm}. Due to the need to use halfcell data to obtain a matching fullcell voltage simulation, the effort to set up the model is comparatively high. In addition, approximating the OCP_{n} curve along with the set of optimal model parameters is a challenging optimization task due to its nonlinearity. This problem is exacerbated when the two halfcell simulations are to be combined for the full cell. Establishing a workflow for the OCP_{n} curve approximation and parameter validation for fullcells is therefore a topic for further research.
Methods
In this section, the mathematical description of the statespace representation of the PDECM is derived. This includes the description of the governing equations and the solution algorithm. In addition, the implementation of an established hysteresis model is described, the test procedure, and the parameterization workflow is presented.
Model idea
The ECM hysteresis model is based on the model introduced by Dreyer et al.^{5}. They show that in a multiparticle system, particlebyparticle charge occurs, if a nonmonotonic potential curve is applied. In the work of Kondo et al.^{33} it was further shown that by using a distributed particle size the sequence of the phase transition is determined and the path dependency of the rate capability can be described. They assumed that the differences in charge transfer resistance caused by varying particle size is the dominant influence on the sequence of particlebyparticle lithiation, while electrical and ionic conductivity are negligible. ThomasAlyea^{41}, though, proposed a resistivereactant model accounting for holes in the conductive matrix around LFP particles as reason for inhomogeneous particle lithiation. EDXRD measurements by Paxton et al.^{42} showed good agreement with the model prediction. To account for both effects the PDECM uses distributed resistances with a constant particle size. While an additional variation of the particle size is closer to the actual electrode structure, particle size variation is only represented by its influence on the resistance in the PDECM. In this work the concepts of nonmonotonic potential curves and a variation of the particle resistance are combined in an equivalent circuit model, which is the most common modeling approach used in BMS.
In Fig. 6a, the underlying idea of the PDECM is illustrated. It shows the network of N_{ini} partitions of capacity:
with the overall cell capacity C in connection with a specific resistor for each string. The parallel resistances add up to the overall resistance \({R}_{{{{{{{{\rm{res}}}}}}}}}\) by:
The resistors are assigned to N bins, which in turn represent uniformly distributed resistance values in the fixed range \([{R}_{\min },{R}_{\max }]\). The probability function p defines the frequency of each discrete resistance value in the network. Exemplary, the probability density of a normal distribution is shown in Fig. 6b. Here, the probability of being assigned to bin R_{2} is three times higher compared to R_{1}. Looking at Fig. 6a this would translate into R_{I} belonging to bin r_{1}, while R_{II} = R_{III} = R_{IV} are all assigned to bin r_{2}.
The model idea is translated into an ECM, as depicted in Fig. 6c. The ECM consists of N_{ini} resistors connected in parallel and a charge controlled voltage source in each string, representing the partition of capacity. For the hysteresis model to work as described above, a nonmonotonous potential curve is assigned to the voltage sources. The potential OCP_{n} of each voltage source is depending on the amount of charge stored in each parallel string relatively taken to the assigned capacity ΔQ. An exemplary OCP_{n} curve over the SOC_{n} of each string is shown in Fig. 6d, as well as the expected voltage curve for the whole network, depending on the direction of the current.
Model implementation
At first, a uniformly distributed vector of resistance values is defined. The length of the vector is determined by the number of bins N. The maximum and minimum values of the vector are in relation to the center of the interval defined to be:
ΔR is constrained to ΔR ∈ (0, R_{center}) to prevent negative values.
Now, the relative frequencies P_{n} of the resistors associated with each bin are calculated. The frequency is given by the area under the curve of the probability density function. The function of the probability density p(r) as a function of continuous resistance values r must be discretized to obtain a value for each bin. For the Gaussian distribution, with the probability density function:
this is done by limiting the function to 3σ and splitting the range into N uniformly sized intervals:
With the resistance vector and the definition of ΔR the expectation value and standard deviation of the Gaussian distribution are μ = R_{center} and σ = 1/_{3} ⋅ ΔR, respectively. The relative frequency of resistors in each bin is then given by:
For the analysis of the influence of changes on the distribution function, a Weibull distribution is implemented. The Weibull distribution is defined by the shape parameter k and the inverse scaling parameter λ. The probability density function is given by:
Similar to the Gaussian distribution the relative frequency of resistors in each bin is calculated by:
The last design parameter of the PDECM is the OCP_{n}. The OCP of the nth string can be any nonmonotonous analytical function of the SOC_{n}. In this publication the OCP_{n} is defined as:
For the model introduction and parameter variation the variables of Eqs. (10) and (11) were chosen to be x_{1} = −0.8, x_{2} = 0.8, y_{1} = 1.6, and y_{2} = 0.9.
In the Results and discussion section ΔR is given normalized to the center value of the vector of resistances to improve comparability between results. Therefore, ΔR_{norm} = ΔR/R_{center} is used in the range from (0, 1).
Model reduction
The model described above is reduced to improve the performance of the simulation. This is done by pooling resistors, i.e., strings with the same resistor get combined. The relative frequency P_{n} of each resistance value is multiplied with the total number of resistors G. The resulting resistance R_{n} of a number of N_{P} = P_{n} ⋅ G equal resistors with resistance r_{n} connected in parallel is:
Consequently, the capacity assigned to the bin is:
With Eq. (12), the resulting resistance of the network is:
Model reduction is crucial to the applicability of the PDECM in a BMS environment. The electrode’s physical properties with thousands of particles per mm^{2} result in an almost continuous variation of resistivity values in the electrical pathways. Binning these values simplifies the network but still involves high computational cost. Reducing the model to fewer equivalent resistances and capacities allows fast calculation while preserving the basic assumptions of the model. For example, by reducing the number of elements from 1000 to 50, the calculation time is reduced by approximately 400, with a quadratic decrease as illustrated in Table 2.
Solving algorithm
For the fullcell model two RC elements are added in series to the hysteresis ECM to model the dynamic behavior of the battery. The resulting differentialalgebraic system of equations (DAE) is given in its statespace representation. The state vector x is defined as:
with the voltage drop over each RC element U_{RC1} and U_{RC2} respectively, the line currents I and the stored charge in each string Q. The discretetime form of the statespace representation with algebraic equations and isolated nonlinearities is:
with the system response y_{k} given by:
The definition for each matrix and vector is given in the Supplementary Methods in Eqs. (S.1)–(S.5). With reference to Hahn et al.^{43} the system of equations is solved with a linearimplicit Euler method:
with the time step h, and the system function f. The utilization of the LUfactorization enables the necessary inversion of the matrix:
without iterative calculations, assuring stable computation times for solving the DAE. To show the realtime capability of the algorithm, the simulation of a slow rate full cycle with more than 40 h simulation time is conducted on a PC with Intel i710510U processor and 16 GB RAM. By changing N, the influence of the number of parallel strings on the calculation time t_{calc} is demonstrated. The results of the study are given in Table 2, showing an increasing calculation time with the square of N. Even for large numbers of N, the calculation time is two magnitudes smaller than the simulation time.
Singlestate hysteresis model
Plett et al.^{20} derive their model for the dynamic hysteresis voltage from a function f(SOC, t) of SOC and time as well as the maximum polarization \(M({{{{{{{\rm{SOC}}}}}}}},\dot{{{{{{{{\rm{SOC}}}}}}}}})\) that has its origin in hysteresis. With the simplification of \(M({{{{{{{\rm{SOC}}}}}}}},\dot{{{{{{{{\rm{SOC}}}}}}}}})=M\cdot {{{{{{{\rm{sign}}}}}}}}(I)\) and further mathematical operations the difference equation:
for the unitless hysteresis state x_{hys} is derived. The rate of alteration in the hysteresis voltage is linked to how far the current hysteresis value deviates from the main hysteresis loop. This results in an exponential decline of voltage toward the primary loop. The positive constant γ controls the pace of this decline. Additionally to the dynamic hysteresis that changes with SOC, instantaneous hysteresis can be modeled by M_{0} ⋅ s_{k}, with:
Equation (20) can easily be implemented in a statespace representation by including x_{hys} as an additional state. Together with M, x_{hys} and an SOCdependent mean OCV curve the resulting hysteresisaffected voltage is described by:
and is added to the output equation of the statespace representation.^{44}
Full and halfcell measurements and tests
For the validation of the hysteresis model a cylindrical A123 18650M1A cell in the format 18650 is chosen. The cells consists of a graphite anode and a LFP cathode and therefore show a significant hysteresis dominated by firstorder phase transitions. All tests are carried out using a BaSyTec CTS battery tester in a Binder MK115 climate chamber at 25 °C. After preconditioning the cell with ten 1C fullcycles, it undergoes a twopart testing procedure: first, the cell’s characterization, and second, the conduction of test cases for the validation of the hysteresis model.
The cell’s characterization involves conducting pulse tests to determine the parameters of the R2RC model. Each pulse test consists of the following steps: The cell is charged to a specific SOC and allowed to rest for 2 h. Afterwards, a 1C, 10 s charge pulse is applied to the cell and its relaxation behavior is observed for 1 h. This procedure is repeated for every 1% SOC step between 0% to 5% and 95% to 100% SOC, as well as for every 5% SOC increment between 5% and 95%.
The first test case for the model validation is a full cycle with a low current rate of 0.05C. To analyze the transition behavior from the charging hysteresis curve to the discharging curve, expanding hysteresis loops starting at 50% SOC are measured in the second test. Finally, the influence on the path dependency of the rate capability is demonstrated by simulating a fast discharge with 3C at 50% SOC once starting on the charge voltage curve and once starting from the discharge voltage curve.
For the determination of the OCPs of the electrodes, the cell is opened after the tests. Samples of the electrodes are punched out and assembled in ELCell PAT cells with lithium metal counter electrode, glass fiber separator, lithium reference electrode, and LP57 electrolyte. The cells are rested for 24 h in a temperature chamber at 25 °C. Afterwards a BaSyTec CTSLAB is used to cycle the cells three times with 0.05C to assure a stable cell setup. The OCP is determined by performing a galvanostatic intermittent titration technique (GITT) test as well as a slow rate full cycle with 0.05C with both electrodes. In the GITT for each 1% SOC step, the potential measured at the end of the 10 min rest phase is read out as OCP.
Parameterization
First step of the parameterization of both models is the determination of the SOCdependent values of the R2RC model, depicted in Fig. 7. Since the same preconnected elements are used in both models, the parameterization procedure is identical. For the determination of the values the rest phase after each charge pulse in the pulse characterization is used. At first the difference between the last data point of the pulse and the first of the rest phase is divided by the current amplitude I_{0} of the pulse to calculated the resistance R_{0} of the processes with very low time constants, as shown in Fig. 7. Afterwards, the optimization problem:
is solved using a leastsquares optimization algorithm (Matlab’s lsqnonlin function). U_{meas} is the measured voltage from t_{0} to t_{1} as indicated in Fig. 7. The resulting parameter used in both models are given in the Supplementary Methods in Supplementary Table 1.
Since the OCP_{n} curve of the materials cannot be directly measured, it must be approximated. For both electrodes the incremental OCP is taken and the first derivative used to determine the SOCs of the phase transitions. Afterwards Eq. (10) is used to model the nonmonotonous part of the OCP_{n} curve. x_{1} and y_{1} are taken at the start of the phase transition from the upper voltage curve while x_{2} and y_{2} are taken from the lower voltage curve. The part of the curve between the transition areas is interpolated using a piecewise cubic Hermite interpolating polynomial (Matlab’s pchip function). Afterwards a leastsquares optimization approach for nonlinear problems (Matlab’s lsqnonlin function) is used to identify the optimized parameter of Eq. (10). In our simulations, N = 50 turned out to be a satisfactory compromise between calculation speed and ripples in the voltage. \({R}_{{{{{{{{\rm{res}}}}}}}}}\) was oriented at the measured R_{0} values determined in the pulse characterization. ΔR_{norm} and the shape parameter k of the Weibull distribution are varied in the range between 0.7 and 0.999 for ΔR_{norm} and between 1 and 3 for k. For the fullcell simulation the values are chosen to represent the behavior during the rate capability test with small error. Since both electrodes are simulated separately, for the simulation of the commercial fullcell the electrode balancing is needed. The balancing is obtained by analyzing the position of stage transition peaks in the differential voltage of the full cell. These peaks are then aligned with the features in the electrodes.
For the Plett hysteresis model, M is computed directly as the difference between the mean OCV and its respective charging and discharging curves. The deviation is averaged over both instances and the values are stored at each SOC sampling point. This is not the standard procedure proposed by Plett^{44}, but yields the results with the smallest deviations for our application. Finally, the value for γ is optimized within a range of 0 and 300 until the smallest deviation between a modeled and measured OCV loop is reached. In this publication 25 is identified as an optimal result for this parameter.
Data availability
The data that supports the findings of this study are available on Zenodo (https://doi.org/10.5281/zenodo.10852930).
Code availability
The code for the model that supports the findings of this study is available on Zenodo (https://doi.org/10.5281/zenodo.10852695).
References
Waag, W., Fleischer, C. & Sauer, D. U. Critical review of the methods for monitoring of lithiumion batteries in electric and hybrid vehicles. J. Power Sources 258, 321–339 (2014).
Farmann, A., Waag, W., Marongiu, A. & Sauer, D. U. Critical review of onboard capacity estimation techniques for lithiumion batteries in electric and hybrid electric vehicles. J. Power Sources 281, 114–130 (2015).
Farmann, A. & Sauer, D. U. A comprehensive review of onboard stateofavailablepower prediction techniques for lithiumion batteries in electric vehicles. J. Power Sources 329, 123–137 (2016).
Röder, F. & Ramasubramanian, S. A review and perspective on path dependency in batteries. Energy Technol. 10, 2200627 (2022).
Dreyer, W. et al. The thermodynamic origin of hysteresis in insertion batteries. Nat. Mater. 9, 448–453 (2010).
Zhu, Y. & Wang, C. Strain accommodation and potential hysteresis of lifepo4 cathodes during lithium ion insertion/extraction. J. Power Sources 196, 1442–1448 (2011).
Kao, Y.H. et al. Overpotentialdependent phase transformation pathways in lithium iron phosphate battery electrodes. Chem. Mater. 22, 5845–5855 (2010).
Mercer, M. P., Peng, C., Soares, C., Hoster, H. E. & Kramer, D. Voltage hysteresis during lithiation/delithiation of graphite associated with metastable carbon stackings. J. Mater. Chem. A 9, 492–504 (2021).
McDowell, M. T., Lee, S. W., Nix, W. D. & Cui, Y. 25th anniversary article: understanding the lithiation of silicon and other alloying anodes for lithiumion batteries. Adv. Mater. 25, 4966–4985 (2013).
van der Ven, A., See, K. A. & Pilon, L. Hysteresis in electrochemical systems. Battery Energy 1, 20210017 (2022).
Allart, D., Montaru, M. & Gualous, H. Model of lithium intercalation into graphite by potentiometric analysis with equilibrium and entropy change curves of graphite electrode. J. Electrochem. Soc. 165, A380–A387 (2018).
Li, J. & Dahn, J. R. An in situ Xray diffraction study of the reaction of Li with crystalline Si. J. Electrochem. Soc. 154, A156 (2007).
Jiang, Y., Offer, G., Jiang, J., Marinescu, M. & Wang, H. Voltage hysteresis model for silicon electrodes for lithium ion batteries, including multistep phase transformations, crystallization and amorphization. J. Electrochem. Soc. 167, 130533 (2020).
Grimsmann, F. et al. Hysteresis and current dependence of the graphite anode color in a lithiumion cell and analysis of lithium plating at the cell edge. J. Energy Storage 15, 17–22 (2018).
Grimsmann, F. et al. Hysteresis and current dependence of the thickness change of lithiumion cells with graphite anode. J. Energy Storage 12, 132–137 (2017).
Srinivasan, V. & Newman, J. Existence of pathdependence in the LiFePo_{4} electrode. Electrochem. Solid State Lett. 9, A110 (2006).
Christophersen, J. P., Hunt, G. L., Ho, C. D. & Howell, D. Pulse resistance effects due to charging or discharging of highpower lithiumion cells: a path dependence study. J. Power Sources 173, 998–1005 (2007).
Roscher, M. A., Vetter, J. & Sauer, D. U. Cathode material influence on the power capability and utilizable capacity of next generation lithiumion batteries. J. Power Sources 195, 3922–3927 (2010).
Plett, G. L. Extended Kalman filtering for battery management systems of LiPBbased HEV battery packs: part 1. Background. J. Power Sources 134, 252–261 (2004).
Plett, G. L. Extended Kalman filtering for battery management systems of LiPBbased HEV battery packs: part 2. Modeling and identification. J. Power Sources 134, 262–276 (2004).
Plett, G. L. Extended Kalman filtering for battery management systems of LiPBbased HEV battery packs: part 3. State and parameter estimation. J. Power Sources 134, 277–292 (2004).
Tjandra, R., Thanagasundram, S., Tseng, K. J. & Jossen, A. Improved lithiumion battery model with hysteresis effect. In 2014 IEEE Transportation Electrification Conference and Expo (ITEC 2014) 1–8 (2014).
Baronti, F., Femia, N., Saletti, R., Visone, C. & Zamboni, W. Preisach modelling of lithiumironphosphate battery hysteresis. J. Energy Storage 4, 51–61 (2015).
Roscher, M. A. & Sauer, D. U. Dynamic electric behavior and opencircuitvoltage modeling of LiFePO_{4}based lithium ion secondary batteries. J. Power Sources 196, 331–336 (2011).
Guenther, C., Barillas, J. K., Stumpp, S. & Danzer, M. A. A dynamic battery model for simulation of batterytogrid applications. In 2012 3rd IEEE PES Innovative Smart Grid Technologies Europe (ISGT Europe) 1–7 (2012).
Li, F. et al. Exploring the hysteresis effect of Liion batteries: a machine learning based approach. In 2023 International Joint Conference on Neural Networks (IJCNN) 1–8 (2023).
Xu, Z., Wang, J., Fan, Q., Lund, P. D. & Hong, J. Improving the state of charge estimation of reused lithiumion batteries by abating hysteresis using machine learning technique. J. Energy Storage 32, 101678 (2020).
Chayratsami, P. & Plett, G. L. Hysteresis modeling of lithium–silicon half cells using Prandtl–Ishlinskii model. In 2020 IEEE 16th International Conference on Control & Automation (ICCA) 1578–1583 (2020).
Baker, D. R., Verbrugge, M. W. & Xiao, X. An approach to characterize and clarify hysteresis phenomena of lithiumsilicon electrodes. J. Appl. Phys. 122, 165102 (2017).
Graells, C. P., Trimboli, M. S. & Plett, G. L. Differential hysteresis models for a siliconanode Liion battery cell. In 2020 IEEE Transportation Electrification Conference & Expo (ITEC) 175–180 (2020).
Galuppini, G. et al. Efficient computation of safe, fast charging protocols for multiphase lithiumion batteries: a lithium iron phosphate case study. J. Power Sources 580, 233272 (2023).
Farkhondeh, M., Pritzker, M., Fowler, M. & Delacourt, C. Mesoscopic modeling of a LiFePO_{4} electrode: experimental validation under continuous and intermittent operating conditions. J. Electrochem. Soc. 164, E3040–E3053 (2017).
Kondo, H., Sasaki, T., Barai, P. & Srinivasan, V. Comprehensive study of the polarization behavior of LiFePO_{4} electrodes based on a manyparticle model. J. Electrochem. Soc. 165, A2047–A2057 (2018).
Wang, S. et al. Improved singular filteringGaussian process regressionlong shortterm memory model for wholelifecycle remaining capacity estimation of lithiumion batteries adaptive to fast aging and multicurrent variations. Energy 284, 128677 (2023).
Wang, S., Fan, Y., Jin, S., TakyiAninakwa, P. & Fernandez, C. Improved antinoise adaptive long shortterm memory neural network modeling for the robust remaining useful life prediction of lithiumion batteries. Reliab. Eng. Syst. Saf. 230, 108920 (2023).
Yamada, A. et al. Roomtemperature miscibility gap in Li_{x}FePO_{4}. Nat. Mater. 5, 357–360 (2006).
Baronti, F., Zamboni, W., Femia, N., Roncella, R. & Saletti, R. Experimental analysis of opencircuit voltage hysteresis in lithiumironphosphate batteries. IECON 2013 6728–6733 (2013).
Li, Y. et al. Currentinduced transition from particlebyparticle to concurrent intercalation in phaseseparating battery electrodes. Nat. Mater. 13, 1149–1156 (2014).
Bai, P., Cogswell, D. A. & Bazant, M. Z. Suppression of phase separation in LiFePO_{4} nanoparticles during battery discharge. Nano Lett. 11, 4890–4896 (2011).
Katrašnik, T. et al. Entering voltage hysteresis in phaseseparating materials: revealing the electrochemical signature of the intraparticle phaseseparated state. Adv. Mater. 35, e2210937 (2023).
ThomasAlyea, K. E. Modeling resistivereactant and phasechange materials in battery electrodes. ECS Trans. 16, 155–165 (2008).
Paxton, W. A., Zhong, Z. & Tsakalakos, T. Tracking inhomogeneity in highcapacity lithium iron phosphate batteries. J. Power Sources 275, 429–434 (2015).
Hahn, M., Schiela, A., Mößle, P., Katzer, F. & Danzer, M. A. Revealing inhomogeneities in electrode lithiation using a realtime discrete electrochemical model. J. Power Sources 477, 228672 (2020).
Plett, G. Battery Management Systems, Vol. I (Artech House, 2015).
Crameri, F., Shephard, G. E. & Heron, P. J. The misuse of colour in science communication. Nat. Commun. 11, 5444 (2020).
Acknowledgements
The authors would like to thank Srivatsan Ramasubramanian from the Bavarian Center of Battery Technology (BayBatt) for the scientific discussion and critical review of the manuscript. Furthermore, the authors thank the graduate school of the BayBatt for ongoing support. Support by the BayBatt Cell Technology Center is gratefully acknowledged, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)—INST 91/4521 LAGG. The color palette used in this report was developed by Crameri et al.^{45} and is publicly available on Zenodo (https://doi.org/10.5281/zenodo.8035877). This work is funded by the German Federal Ministry of Education and Research (BMBF) under funding reference 03XP0321A.
Author information
Authors and Affiliations
Contributions
Leonard Jahn: methodology, software, investigation, data curation, validation, writing—original draft, visualization. Patrick Mößle: methodology, software, validation, writing—original draft. Fridolin Röder: conceptualization, methodology, writing—review & editing, supervision. Michael A. Danzer: methodology, writing—review & editing, supervision, funding acquisition.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Communications Engineering thanks Shunli Wang, Dominik Wycisk and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Jiangong Zhu, Mengying Su, Miranda Vinay, Anastasiia Vasylchenkova, Saleem Denholme. A peer review file is available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Jahn, L., Mößle, P., Röder, F. et al. A physically motivated voltage hysteresis model for lithiumion batteries using a probability distributed equivalent circuit. Commun Eng 3, 74 (2024). https://doi.org/10.1038/s44172024002214
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s44172024002214