Numerical Fitting for SOC Calibration

I designed and evaluated a numerical-fitting approach for accurate state-of-charge (SOC) calibration in an electric vehicle battery pack. The motivation is straightforward: SOC directly influences remaining-range prediction, charge and discharge control, cell balancing, thermal management, and safety protection inside an electric vehicle battery pack. Because lithium-ion cells exhibit strong nonlinear behavior, SOC cannot be measured directly by a single sensor. It has to be inferred from voltage, current, temperature, and capacity-related information. In engineering practice, the ampere-hour integration method is still widely used because it is simple and continuous, but it is vulnerable to initial-value error, current-sampling bias, and capacity degradation. I therefore focused on a calibration design that combines open-circuit-voltage (OCV) and SOC data with polynomial numerical fitting, parameter identification, initial-value correction, and periodic compensation. The objective is not merely to produce a smooth curve, but to create a practical calibration path for an electric vehicle battery pack that can be embedded in a battery management system (BMS) and executed online with acceptable computational cost.

In my design, the OCV-SOC relationship is treated as a continuous function rather than a coarse lookup table. The reason is that discrete lookup methods depend heavily on sampling interval and interpolation accuracy. When the electric vehicle battery pack operates over a wide SOC range, the voltage sensitivity changes significantly. Near the low-SOC and high-SOC ends, the OCV changes relatively quickly with SOC, whereas in the middle SOC plateau the OCV changes slowly. A continuous polynomial model can represent this behavior with a compact parameter set, while also providing a differentiable expression for reverse SOC solving.

The baseline ampere-hour relation I use for comparison can be written as follows:

$$SOC(t)=SOC(t_0)-\frac{1}{Q}\int_{t_0}^{t}\eta I(\tau)\,d\tau$$

Here, \(SOC(t_0)\) is the initial SOC, \(Q\) is the available capacity of the electric vehicle battery pack, \(\eta\) is the coulombic efficiency, and \(I(t)\) is the instantaneous current. This expression shows that SOC calculation is coupled to current integration, initial SOC, and capacity parameters. If the initial value is biased or the capacity parameter is not updated in time, the SOC output will drift continuously. Initial-value deviation and current-sampling error accumulate with operating time. OCV-based methods have a clear physical correspondence, but they are affected by rest time, temperature, and aging of the electric vehicle battery pack. Under vehicle-mounted conditions, the battery is frequently exposed to current fluctuations and temperature changes. If the calibration model cannot reflect local changes in the OCV-SOC relationship, the SOC output will exhibit lag and drift. This is why I propose a numerical-fitting calibration strategy that balances data stability, model accuracy, and embedded implementation constraints.

I define SOC as the ratio of remaining charge to available capacity. For an electric vehicle battery pack, this quantity is one of the most important state indicators. During calibration, the OCV-SOC curve is regarded as a reference when the terminal voltage has stabilized after rest. Because the voltage change rate differs across SOC regions, raw data must be screened and processed before fitting. The data-processing steps I use include abnormal-point removal, sampling-period unification, voltage smoothing, and capacity normalization. Current spikes, voltage jumps, and insufficient rest time can all make the OCV data unreliable. These effects have to be corrected before modeling. After processing, the SOC and OCV data can serve as inputs to the fitting model and provide a reliable basis for numerical calibration.

For an electric vehicle battery pack, the OCV-SOC mapping is not a simple linear relationship. I therefore represent it with a polynomial function in which SOC is the independent variable and OCV is the dependent variable:

$$OCV(SOC)=a_0+a_1 SOC+a_2 SOC^2+\cdots+a_n SOC^n$$

In this expression, \(a_0\) through \(a_n\) are the fitting coefficients, and \(n\) is the highest order of the polynomial. A polynomial model has a clear structure, is convenient to implement in a vehicle BMS, and requires relatively low computation. It is therefore suitable for online calibration in an electric vehicle battery pack. However, when building the model, I cannot pursue local accuracy only through a high order. I also need to consider curve smoothness and the stability of reverse solving. If the order is too low, the model cannot describe voltage changes near the SOC endpoints. If the order is too high, the model may oscillate locally. Therefore, the fitting model must balance accuracy and stability.

In practical application, after the BMS obtains an approximate OCV during rest, I solve the polynomial equation by the Newton iteration method to obtain the corresponding SOC estimate. The iterative form can be written as:

$$SOC_{k+1}=SOC_k-\frac{OCV(SOC_k)-OCV_{\mathrm{meas}}}{OCV'(SOC_k)}$$

This reverse solution is important for an electric vehicle battery pack because the BMS measures voltage, not SOC. The numerical model must therefore be invertible in a stable and efficient manner. If the derivative \(OCV'(SOC)\) is too small in the middle plateau, the reverse solution becomes sensitive to voltage error. If the derivative changes abruptly near the endpoints, the iteration may become unstable. For this reason, I evaluate not only fitting error but also the smoothness of the first derivative.

The calibration data usually consist of multiple SOC sampling points and their corresponding stable open-circuit voltages. The more completely the data cover the full SOC range, the better the fitting model can reflect the voltage variation of the electric vehicle battery pack. I use least squares for parameter identification. The goal is to minimize the sum of squared residuals between fitted voltage and measured voltage. Let the measured open-circuit voltage at the \(i\)-th sampling point be \(U_i\), and let the fitted voltage be \(\hat{U}_i\). The objective function is:

$$J=\sum_{i=1}^{m}\left(U_i-\hat{U}_i\right)^2$$

When \(J\) reaches a small value, the fitting curve describes the measured data trend well. In matrix form, I write the model as:

$$\mathbf{U}=\mathbf{\Phi}\mathbf{a}+\boldsymbol{\varepsilon}$$

where \(\mathbf{U}\) is the vector of measured OCV values, \(\mathbf{\Phi}\) is the Vandermonde-type design matrix, \(\mathbf{a}\) is the coefficient vector, and \(\boldsymbol{\varepsilon}\) is the residual vector. The least-squares estimate is:

$$\hat{\mathbf{a}}=\left(\mathbf{\Phi}^{\mathsf T}\mathbf{\Phi}\right)^{-1}\mathbf{\Phi}^{\mathsf T}\mathbf{U}$$

For an electric vehicle battery pack, numerical conditioning matters. If the SOC values are not normalized, the higher-order terms can become ill-conditioned. I therefore scale SOC to the interval \([0,1]\) before fitting and then transform the coefficients back if necessary. This step improves numerical stability without changing the physical meaning of the model.

The choice of polynomial order should not be excessively high. A low order cannot express the nonlinear changes at low SOC and high SOC. A high order may produce local oscillation and affect the stability of reverse SOC query. In my design, I compare third-, fifth-, and seventh-order polynomials. The evaluation indicators are root-mean-square error (RMSE), maximum absolute error, and coefficient of determination \(R^2\). These are defined as follows:

$$RMSE=\sqrt{\frac{1}{m}\sum_{i=1}^{m}\left(U_i-\hat{U}_i\right)^2}$$

$$MAE=\max_{1\le i\le m}\left|U_i-\hat{U}_i\right|$$

$$R^2=1-\frac{\sum_{i=1}^{m}\left(U_i-\hat{U}_i\right)^2}{\sum_{i=1}^{m}\left(U_i-\bar{U}\right)^2}$$

For an electric vehicle battery pack, the RMSE reflects overall fitting accuracy, the maximum absolute error reflects worst-case risk, and \(R^2\) reflects how much of the voltage variation is explained by the model. I do not select the order by \(R^2\) alone. A slightly higher \(R^2\) at the seventh order does not necessarily justify the increased risk of local oscillation. The fifth-order model provides a better compromise among fitting accuracy, curve smoothness, and engineering applicability.

Polynomial order RMSE / V Maximum absolute error / V \(R^2\) Reverse-solve stability Embedded cost
Third order 0.0186 0.0421 0.9913 Moderate Low
Fifth order 0.0094 0.0217 0.9978 Good Moderate
Seventh order 0.0087 0.0196 0.9984 Fair High

The third-order polynomial has a simple form and low computational cost, but it does not describe voltage changes well at low SOC and high SOC. Its maximum absolute error is relatively large, which indicates that a low-order model cannot capture the nonlinear characteristics of the OCV-SOC relationship in an electric vehicle battery pack. The seventh-order polynomial has the best goodness of fit, and its RMSE is slightly smaller than that of the fifth-order model. However, high-order fitting tends to amplify local sampling noise and can cause local oscillation, which reduces the stability of reverse SOC solving. The fifth-order polynomial achieves a balanced performance between error control and curve smoothness. Its RMSE is smaller than that of the third-order model, and its maximum absolute error remains within a small range without obvious local oscillation. Considering fitting accuracy, model stability, and computational requirements in a vehicle BMS, I select the fifth-order polynomial as the calibration model for the electric vehicle battery pack.

Initial SOC calibration is used for state recovery after vehicle power-up, long rest, or maintenance reset. If the initial SOC is biased, subsequent ampere-hour integration will continue to recurse from that biased value and may not return to the true state in time. During initial-value calibration, the BMS first determines whether the electric vehicle battery pack is in a rest state. This can be judged from the absolute current value, the voltage change rate, and the rest time. When the current is close to zero and the terminal voltage change rate is below a threshold, I use the corrected terminal voltage as an approximate OCV input to the fitting model and obtain the initial SOC. If the rest time is insufficient, the terminal voltage still contains polarization effects. In that case, I estimate the voltage drop from the equivalent internal resistance and compensate for it. Initial calibration should not be executed immediately after a high-current event, because the polarization voltage has not fully decayed and the SOC calibration result may be too high or too low. In a low-temperature environment, I call the correction coefficient for the corresponding temperature range. Temperature and aging change both the capacity and the terminal-voltage response of an electric vehicle battery pack, so initial calibration should be conditional and should avoid forced correction under unstable operating conditions.

During driving, SOC is mainly propagated by ampere-hour integration. Numerical fitting is used to compensate for long-distance error. I avoid frequent calibration correction under transient conditions. Instead, I trigger correction when the current is small, the temperature change is mild, or the vehicle is briefly stopped. The method is as follows. First, I estimate OCV from the current terminal voltage, current, and internal resistance. Then I use the fitting model to obtain a calibrated SOC. I compare it with the recursive SOC and obtain the difference. When the difference is smaller than the allowable threshold, I apply only a small smooth correction. When the difference is larger than the threshold, I combine voltage stability and sensor status for a second judgment to prevent abnormal voltage from causing an incorrect correction. The correction coefficient can be set in the form of a weight so that the SOC output remains continuous. This process preserves the real-time capability of ampere-hour integration and uses the OCV-SOC fitting relationship to limit long-term drift. It is suitable for deployment as a periodic task in BMS software for an electric vehicle battery pack.

The voltage estimate used for correction can be written as:

$$OCV_{\mathrm{est}}=V_t+I R_0+V_{\mathrm{pol}}$$

Here, \(V_t\) is the measured terminal voltage, \(I\) is the current, \(R_0\) is the ohmic internal resistance, and \(V_{\mathrm{pol}}\) is the polarization voltage. In a practical electric vehicle battery pack, \(R_0\) and \(V_{\mathrm{pol}}\) vary with temperature, SOC, and aging. I therefore treat them as correction terms rather than fixed constants. The OCV estimate is more reliable when the current is small and the polarization has decayed. If the current is large, the uncertainty of \(I R_0+V_{\mathrm{pol}}\) can become comparable to the OCV change caused by a small SOC change. This is especially problematic in the middle SOC plateau, where \(dOCV/dSOC\) is small. Therefore, the correction trigger must include current and stability conditions.

The SOC difference and correction can be expressed as:

$$\Delta SOC=SOC_{\mathrm{fit}}-SOC_{\mathrm{AH}}$$

$$SOC_{\mathrm{new}}=SOC_{\mathrm{AH}}+\alpha\Delta SOC$$

where \(\alpha\) is a weighting factor between 0 and 1. A small \(\alpha\) gives smooth but slow correction. A large \(\alpha\) gives fast correction but may cause visible jumps in the SOC display of the electric vehicle battery pack. I prefer a variable weighting strategy. When the electric vehicle battery pack is in a stable rest state and the measured voltage is reliable, \(\alpha\) can be relatively large. When the vehicle is under dynamic load, \(\alpha\) should be small or zero. This avoids incorrect correction caused by transient voltage behavior.

Trigger condition Current criterion Voltage criterion Temperature criterion Correction weight \(\alpha\)
Long rest after power-up \(|I|<0.01C\) \(|dV/dt|<0.5\) mV/min Stable for 30 min 0.80-1.00
Short stop \(|I|<0.02C\) \(|dV/dt|<1.0\) mV/min Stable for 10 min 0.30-0.60
Low-current cruise \(|I|<0.05C\) Voltage residual small Slow change 0.05-0.20
High-current dynamic \(|I|\ge0.05C\) Not stable Any 0

For the electric vehicle battery pack, the most difficult region is the middle SOC plateau. In this region, a relatively large change in SOC produces only a small change in OCV. Therefore, a small voltage measurement error can produce a large SOC error when reverse solving. I handle this by reducing the correction weight in the plateau and by relying more on ampere-hour integration when the current is stable. Near the low-SOC and high-SOC ends, the voltage sensitivity is higher, so the fitting model can identify SOC changes more effectively. This region-dependent behavior is an important reason why a single fixed correction strategy is not sufficient for an electric vehicle battery pack.

To evaluate the method, I set up an experimental platform for a single cell with a rated capacity of 50 Ah. The platform included a battery charge-discharge tester, a thermal chamber, a temperature acquisition module, and a host computer recording system. Before the test, I performed a standard charge-discharge cycle at 25 °C to eliminate the influence of the previous cycle on the capacity state. For OCV-SOC calibration, I used a constant-current stepped discharge method. The discharge rate was set to 0.05C. After the cell was fully charged, it was rested for 2 h. Then I discharged it in steps of 5% SOC. After each step, the cell was rested until the voltage change rate fell below 0.5 mV/min, and the stable voltage corresponding to that SOC point was recorded. The dynamic validation stage included constant-current discharge and variable-current discharge. The first was used to examine the calibration model under stable load, and the second was used to examine the influence of current variation on model calibration during vehicle operation. After data acquisition, I removed voltage jump points, current spike points, and data with insufficient rest time. I then smoothed the valid data so that the fitting result was closer to the OCV-SOC mapping law of the electric vehicle battery pack.

Parameter Value Purpose
Rated capacity 50 Ah Cell-level calibration baseline
Ambient temperature 25 °C Reference temperature
Discharge rate for OCV test 0.05C Reduce polarization and rate effect
Rest time after full charge 2 h Allow voltage relaxation
SOC step 5% Cover full SOC range
Rest criterion \(|dV/dt|<0.5\) mV/min Ensure quasi-equilibrium OCV
Dynamic validation Constant current and variable current Evaluate online calibration behavior

The test described above was completed at 25 °C. High-temperature and low-temperature validation will be part of follow-up work. Even at the reference temperature, the results show clear differences among fitting orders and calibration methods. For the electric vehicle battery pack, the calibration model must be robust not only to fitting error but also to input data quality. A model that fits the training data extremely well but oscillates between sample points is not suitable for online SOC calibration. This is why I emphasize smoothness and derivative behavior in addition to RMSE.

To evaluate the application effect of the numerical-fitting SOC calibration method, I compared it with ampere-hour integration and OCV lookup. Ampere-hour integration has good continuous calculation capability, but its result depends strongly on initial SOC, current sensor accuracy, and capacity parameters. As the operating time of the electric vehicle battery pack increases, cumulative error is likely to appear. OCV lookup can use open-circuit voltage to correct SOC, but this method has high requirements for rest conditions, and the lookup interval affects the refinement of SOC calibration. The numerical fitting method transforms discrete OCV-SOC data into a continuous functional relationship. It can reduce lookup error and achieve relatively stable SOC initial-value updating after voltage correction. The comparison results are summarized in the following table.

Calibration method Initial SOC error / % Average calibration error / % Maximum calibration error / % Long-term drift tendency Online suitability
Ampere-hour integration 3.20 2.86 5.14 High Good continuity, weak correction
OCV lookup 2.45 2.18 4.36 Moderate Depends on rest conditions
Numerical fitting 1.18 1.05 2.31 Low Good balance

The numerical fitting method has lower initial SOC error, average calibration error, and maximum calibration error than the other two methods. This indicates that the method has advantages in calibration accuracy and output continuity. In the low-SOC and high-SOC regions, the voltage changes significantly with SOC, so the numerical fitting model has a strong ability to identify SOC changes. In the middle SOC region, the voltage plateau is flat, and there is insufficient sensitivity in reverse SOC calculation. In practical application, I can combine current thresholds and voltage stability conditions to limit the correction timing and reduce false correction under dynamic conditions.

From the perspective of an electric vehicle battery pack, the calibration problem is not isolated from other BMS functions. SOC calibration affects remaining-range estimation, power derating, charge control, and fault diagnosis. If the SOC estimate is too high, the vehicle may overestimate the available energy and underestimate the risk of deep discharge. If the SOC estimate is too low, the vehicle may unnecessarily limit power or trigger early charging. Both cases degrade user experience and may affect battery safety. Therefore, the calibration design must not only be accurate on average but also robust under edge conditions. The fifth-order polynomial model I selected provides a practical compromise for the electric vehicle battery pack because it captures endpoint nonlinearity without introducing excessive oscillation.

I also examined the computational cost of the calibration model. For an embedded BMS, the SOC task usually runs periodically, for example every 10 ms, 100 ms, or 1 s depending on the software architecture. A fifth-order polynomial requires a small number of multiplications and additions. The Newton iteration for reverse solving usually converges within a few steps if the initial guess is reasonable. By contrast, a lookup table with fine resolution requires more memory, and interpolation between points may still introduce quantization error. The polynomial model therefore offers a favorable trade-off for an electric vehicle battery pack with limited microcontroller resources.

Implementation option Memory requirement Computation per evaluation Interpolation error Smoothness
Fine lookup table High Low to moderate Depends on grid size Limited by interpolation
Third-order polynomial Very low Low High at endpoints Good
Fifth-order polynomial Low Moderate Low Good
Seventh-order polynomial Low Moderate to high Very low in-sample Risk of oscillation

The SOC calibration strategy I propose includes several layers. The first layer is data preprocessing. The second layer is polynomial fitting and parameter identification. The third layer is initial SOC calibration. The fourth layer is online correction and compensation. The fifth layer is validation and fault suppression. Each layer has a specific role in the electric vehicle battery pack. Data preprocessing prevents bad measurements from entering the model. Parameter identification provides a compact representation of the OCV-SOC relationship. Initial calibration recovers the SOC state after rest or reset. Online correction limits long-term drift during operation. Validation and fault suppression prevent abnormal voltage or current from producing an incorrect SOC update.

For data preprocessing, I use the following steps. First, I remove points where the current exceeds a threshold, because these points are not in a quasi-equilibrium state. Second, I remove points where the voltage changes abruptly, because these may be caused by contact resistance, sensor noise, or cell connection issues. Third, I unify the sampling period so that the data used for fitting are evenly spaced. Fourth, I smooth the voltage with a moving average or a low-pass filter. Fifth, I normalize the capacity and SOC data. These steps improve the reliability of the OCV-SOC calibration data and reduce the influence of outliers on the polynomial coefficients.

Preprocessing step Purpose Risk if omitted
Outlier removal Remove voltage jumps and current spikes Biased polynomial coefficients
Sampling-period unification Make data comparable Uneven weighting in least squares
Voltage smoothing Reduce sensor noise Local oscillation in high-order fit
Capacity normalization Separate SOC from capacity variation Confusion between SOC and aging effects
Rest-time validation Ensure quasi-equilibrium OCV Polarization error in OCV

In my calibration design, the OCV-SOC polynomial is not used as a standalone SOC estimator. It is used as a correction reference. The primary real-time SOC propagation is still ampere-hour integration, because it provides continuity and fast response. The polynomial model provides a stable anchor when the electric vehicle battery pack is at rest or under low-current conditions. This hybrid structure is important because neither method alone is sufficient. Ampere-hour integration alone drifts, while OCV-based methods alone cannot provide continuous online SOC. The combination gives the electric vehicle battery pack a more reliable SOC output.

The correction trigger logic can be described by a set of inequalities. Let \(|I|\) be the absolute current, \(|dV/dt|\) the absolute voltage change rate, and \(T_{\mathrm{stable}}\) the duration of stable conditions. A correction is allowed when:

$$|I|<i_{\mathrm{th}},\quad \left|\frac{dv}{dt}\right|T_{\mathrm{th}}$$

Here, \(I_{\mathrm{th}}\), \(V_{\mathrm{th}}\), and \(T_{\mathrm{th}}\) are thresholds that depend on the cell chemistry, the size of the electric vehicle battery pack, and the BMS sampling capability. For a large electric vehicle battery pack, the current threshold may be scaled by the rated capacity, for example \(0.01C\) to \(0.05C\). The voltage threshold may be set to a few tenths of a millivolt per minute. The stable duration may range from several minutes to tens of minutes. These thresholds should be calibrated for the specific electric vehicle battery pack rather than copied from another system.

The temperature and aging compensation can be written in a general form:

$$OCV_{\mathrm{corr}}=OCV_{\mathrm{meas}}+k_T(T-T_{\mathrm{ref}})+k_{\mathrm{age}}(SOH-SOH_{\mathrm{ref}})$$

where \(k_T\) and \(k_{\mathrm{age}}\) are compensation coefficients, \(T_{\mathrm{ref}}\) is the reference temperature, and \(SOH_{\mathrm{ref}}\) is the reference state of health. This expression is a simplified representation. In a real electric vehicle battery pack, the compensation may be a function rather than a constant coefficient. Temperature affects the OCV curve and internal resistance, while aging affects capacity and the shape of the OCV-SOC curve. Therefore, the calibration model should be updated as the electric vehicle battery pack ages. If the model is fixed for the entire life of the battery, the SOC calibration accuracy will degrade over time.

I also considered the effect of current-sampling error. In ampere-hour integration, the current error enters the integral and accumulates over time. If the current sensor has a bias \(b_I\), the SOC error after time \(t\) is approximately:

$$\Delta SOC_I=-\frac{1}{Q}\int_{t_0}^{t}\eta b_I\,d\tau=-\frac{\eta b_I(t-t_0)}{Q}$$

This shows that even a small current bias can produce a significant SOC error over a long driving period. For an electric vehicle battery pack, this is a major reason why open-loop ampere-hour integration cannot be used without correction. The numerical fitting model provides a periodic reference that limits this drift. However, the correction should not be applied blindly. If the voltage measurement is noisy or the electric vehicle battery pack is not in a stable state, the correction may introduce additional error. Therefore, the trigger conditions and weighting factors are essential.

The capacity parameter \(Q\) also changes with aging. If the BMS uses a fixed nominal capacity, the SOC error will grow as the electric vehicle battery pack ages. A capacity update strategy can be combined with the numerical fitting calibration. For example, when the battery is fully charged or fully discharged under known conditions, the integrated charge can be used to update the capacity estimate. This is beyond the scope of the present calibration design, but it is an important direction for a complete BMS. I mention it because SOC and capacity estimation are coupled in an electric vehicle battery pack.

To further validate the design, I analyzed the error distribution across SOC regions. The following table summarizes the typical behavior I observed. The exact values depend on the cell and the test conditions, but the trend is representative of an electric vehicle battery pack.

SOC region Voltage sensitivity Main challenge Calibration strategy
0%-20% High Steep voltage change, risk of deep discharge Use fitting model with conservative correction
20%-80% Low to moderate Flat plateau, poor reverse-solve sensitivity Reduce correction weight, rely on ampere-hour integration
80%-100% High Steep voltage change, risk of overcharge Use fitting model with voltage stability check

In the low-SOC region, the electric vehicle battery pack is close to the discharge limit. Accurate SOC calibration is important for power limiting and safety protection. In the high-SOC region, accurate calibration is important for charge termination and regenerative braking acceptance. In the middle region, the voltage plateau makes OCV-based correction less sensitive, but this is also where the electric vehicle battery pack operates most of the time. Therefore, the calibration strategy must be region-dependent. A single fixed correction weight is not sufficient.

The numerical fitting method also has limitations. It depends on the quality of the OCV-SOC data. If the data are collected at the wrong temperature or after insufficient rest, the fitted curve will be biased. It also assumes that the OCV-SOC relationship is stable over the fitting window. In reality, the relationship changes with aging and temperature. Therefore, the model should be periodically updated. For an electric vehicle battery pack, this update can be performed when the battery is at rest for a long period, such as overnight parking. The BMS can collect OCV data at several SOC points over time and update the polynomial coefficients using recursive least squares or a batch least-squares method.

In my view, the fifth-order polynomial is a good starting point for an electric vehicle battery pack. It captures the main nonlinear features of the OCV-SOC curve without excessive complexity. However, the optimal order may vary with cell chemistry and pack configuration. For a lithium iron phosphate electric vehicle battery pack, the middle plateau is very flat, and a higher-order model or a piecewise model may be needed. For a nickel-manganese-cobalt electric vehicle battery pack, the curve may be smoother, and a lower-order model may be sufficient. Therefore, the calibration design should be adaptable rather than fixed.

I also compared the proposed method with a piecewise linear interpolation approach. Piecewise linear interpolation is simple and can approximate a nonlinear curve with enough breakpoints. However, it is not differentiable at the breakpoints, and the derivative discontinuity can affect Newton iteration. A polynomial model is smooth and differentiable, which is an advantage for reverse solving. The trade-off is that a single polynomial may not fit all regions equally well. A piecewise polynomial or spline model can combine smoothness with local flexibility, but it increases implementation complexity. For a resource-constrained BMS in an electric vehicle battery pack, the fifth-order polynomial is a practical compromise.

The calibration compensation strategy can be summarized as follows. When the electric vehicle battery pack is in a stable rest state, the BMS estimates OCV, solves the polynomial model, and obtains a reference SOC. If the difference between the reference SOC and the recursive SOC is small, the BMS applies a gentle correction. If the difference is large, the BMS checks the validity of the measurement and may request a longer rest or additional data. During dynamic operation, the BMS relies mainly on ampere-hour integration and disables or reduces OCV correction. This prevents transient voltage behavior from corrupting the SOC estimate. The result is a continuous SOC output that is periodically anchored to the OCV-SOC relationship.

Operating state Primary SOC source OCV correction Reason
Long rest OCV fitting model Strong Voltage is close to equilibrium
Short stop Ampere-hour integration Moderate Partial relaxation, some polarization remains
Low-current cruise Ampere-hour integration Weak Small current, but not fully equilibrated
High-current dynamic Ampere-hour integration Disabled Polarization and ohmic drop dominate

For an electric vehicle battery pack, the BMS software architecture usually includes a real-time task for current integration and a slower task for state estimation and calibration. The numerical fitting model can be executed in the slower task. The polynomial coefficients can be stored in non-volatile memory and updated during service or long rest. The Newton iteration can be limited to a maximum number of steps to ensure bounded execution time. If the iteration does not converge, the BMS can fall back to a lookup table or retain the previous SOC estimate. This fail-safe behavior is important for an electric vehicle battery pack because an incorrect SOC update can affect vehicle control.

The experimental results show that the numerical fitting method reduces the initial SOC error to about 1.18%, the average calibration error to about 1.05%, and the maximum calibration error to about 2.31%. These values are lower than those of ampere-hour integration and OCV lookup in the same test. The improvement is not only in average accuracy but also in the maximum error, which is important for safety and user experience. A lower maximum error means that the electric vehicle battery pack is less likely to experience extreme SOC misestimation. This is especially valuable near the low-SOC and high-SOC limits.

However, I do not claim that numerical fitting alone solves all SOC calibration problems. It is a calibration method, not a full state estimator. It should be combined with ampere-hour integration, temperature compensation, aging compensation, and fault diagnosis. In a complete BMS for an electric vehicle battery pack, SOC estimation may also use Kalman filtering or other advanced observers. The numerical fitting model can serve as the OCV-SOC reference for those observers. It can also provide a fallback method when the observer is not available or when the battery is at rest. This modular view is important for practical deployment.

For future work, I plan to extend the tests to high and low temperatures. Temperature has a strong effect on the OCV-SOC curve and on internal resistance. A temperature-dependent polynomial model or a temperature-compensated polynomial model may be needed. I also plan to test the method under aging conditions. As the electric vehicle battery pack ages, the OCV-SOC curve shifts and the capacity decreases. The calibration model should adapt to these changes. A joint SOC and state-of-health estimation framework may be more effective than treating them separately. Finally, I plan to evaluate the method on a full-scale electric vehicle battery pack rather than a single cell. Cell-to-cell variation, balancing current, and pack-level thermal gradients can introduce additional challenges that are not fully captured in cell-level tests.

In summary, I designed a numerical-fitting SOC calibration method for an electric vehicle battery pack. The method uses OCV-SOC calibration data to build a polynomial model, identifies the parameters by least squares, calibrates the initial SOC after rest, and applies weighted correction during operation. The fifth-order polynomial model provides a good balance among fitting accuracy, curve smoothness, and engineering applicability. It reduces initial SOC error and improves the stability of SOC estimation in the BMS. Compared with ampere-hour integration and OCV lookup, the numerical fitting method achieves lower calibration error and better output continuity. It is suitable for online implementation in an electric vehicle battery pack, provided that the trigger conditions, temperature compensation, and aging adaptation are properly designed.

The main equations and their roles in the calibration design are summarized in the following table.

Equation Role in calibration Key input
\(SOC(t)=SOC(t_0)-\frac{1}{Q}\int_{t_0}^{t}\eta I(\tau)d\tau\) Real-time SOC propagation Current, capacity, efficiency
\(OCV(SOC)=a_0+a_1SOC+\cdots+a_nSOC^n\) Continuous OCV-SOC mapping SOC, OCV calibration data
\(J=\sum_{i=1}^{m}(U_i-\hat{U}_i)^2\) Parameter identification objective Measured and fitted voltage
\(\hat{\mathbf{a}}=(\mathbf{\Phi}^{\mathsf T}\mathbf{\Phi})^{-1}\mathbf{\Phi}^{\mathsf T}\mathbf{U}\) Least-squares coefficient estimate Design matrix, voltage vector
\(SOC_{k+1}=SOC_k-\frac{OCV(SOC_k)-OCV_{\mathrm{meas}}}{OCV'(SOC_k)}\) Reverse SOC solving Measured OCV, polynomial derivative
\(OCV_{\mathrm{est}}=V_t+I R_0+V_{\mathrm{pol}}\) OCV estimation under load Terminal voltage, current, resistance
\(SOC_{\mathrm{new}}=SOC_{\mathrm{AH}}+\alpha\Delta SOC\) Weighted SOC correction Recursive SOC, fitted SOC, weight

For an electric vehicle battery pack, the calibration task is ultimately a balance between accuracy, stability, and computational cost. A model that is accurate in the laboratory but unstable in the vehicle is not acceptable. A model that is computationally cheap but inaccurate near the SOC limits is also not acceptable. The numerical fitting approach with a fifth-order polynomial provides a middle path. It is accurate enough for the tested conditions, smooth enough for reverse solving, and cheap enough for an embedded BMS. With proper temperature and aging compensation, it can support reliable SOC calibration throughout the life of an electric vehicle battery pack.

I conclude that the proposed design is technically feasible and practically meaningful. The OCV-SOC data provide the physical basis. The polynomial model provides a continuous and differentiable representation. The least-squares method provides a systematic way to identify the coefficients. The initial calibration and weighted correction provide the online mechanism. The experimental comparison shows lower error than the baseline methods. The remaining work is to extend the validation to more temperatures, aging states, and pack-level configurations. These extensions will make the method more robust for a production electric vehicle battery pack.

Scroll to Top