A battery management system must estimate how much charge remains even though state of charge (SoC) cannot be measured directly with a single sensor. Two widely used battery state of charge estimation methods are coulomb counting and Kalman filtering. The first tracks charge entering and leaving a cell; the second combines a battery model with noisy voltage and current observations. Understanding their assumptions is more valuable than treating either algorithm as a universal accuracy guarantee.
What State of Charge Means and Why It Is Difficult to Observe
SoC expresses remaining charge relative to a defined usable capacity, commonly as a percentage. It is not identical to state of health, which describes degradation, or state of energy, which accounts for the voltage-dependent energy remaining. A cell displaying 50% SoC does not necessarily contain half its maximum deliverable energy. Definitions must also specify operating limits, including voltage cutoffs and temperature conditions.
Current can be measured, but charge accumulation depends on initialization and available capacity. Terminal voltage responds to resistance, polarization, temperature, load history and relaxation. A voltage sample under dynamic load therefore cannot simply be converted into a precise remaining-charge percentage.
Aging changes capacity, impedance and sometimes the open-circuit voltage (OCV) relationship. Chemistry matters: a flat OCV region provides less voltage-based correction than a steep region. Reliable estimation therefore depends on calibrated sensing, consistent definitions and explicit uncertainty management.
Coulomb Counting: Integrating Current Over Time
Coulomb counting starts with an initial SoC estimate and integrates current over time. With discharge current defined as positive, a simplified update is SoC(new) = SoC(old) minus discharged ampere-hours divided by usable capacity in ampere-hours. Charge current increases the estimate under the same sign convention. Implementations may account for charging efficiency and must handle the current sampling interval consistently.
For a cell with estimated usable capacity of 50 ampere-hours, delivering 5 ampere-hours from 80% SoC yields 70%, neglecting other effects. This is charge bookkeeping, not a runtime prediction: voltage limits and temperature may terminate discharge earlier.
Strengths of Coulomb Counting
The method is transparent, computationally inexpensive and responsive to current changes. It does not need a voltage-versus-SoC curve for every update, making it useful during dynamic loading and over portions of the discharge curve where voltage carries little SoC information. It can also provide a convenient baseline against which more elaborate observers are evaluated.
Where Integration Drift Comes From
Current-sensor offset accumulates even at true zero current. Gain errors, missing samples, timebase errors and incorrect signs add bias. Aging also changes usable capacity. Thus accurate integration still gives a misleading percentage if initial SoC or capacity is wrong.
Offset calibration, synchronized sampling, capacity updates and periodic references help control drift. OCV can anchor an estimate after sufficient chemistry-dependent rest, but loaded voltage is not OCV. Clamping a display between zero and 100% does not restore accuracy.
Kalman Filters: Combining a Model with Measurements
A Kalman filter maintains an internal estimate of battery state and its uncertainty. A common equivalent-circuit model represents OCV as a function of SoC, together with ohmic resistance and one or more polarization elements. Current drives the state prediction, while measured terminal voltage provides a consistency check. Temperature may be an input, a scheduled parameter or part of an expanded state model.
At each sample, the estimator predicts the next SoC and polarization state from the previous estimate and measured current. It then predicts terminal voltage, compares that prediction with the observed voltage and computes an innovation, or residual. The filter uses a gain based on model and measurement uncertainty to correct its states. The updated uncertainty propagates into the next cycle.
A linear Kalman filter assumes linear relationships, whereas battery OCV generally varies nonlinearly with SoC. An extended Kalman filter (EKF) locally linearizes the model; an unscented Kalman filter (UKF) propagates representative points through nonlinear equations. Neither wins universally: model quality and tuning matter.
Why the Correction Is Not Always Reliable
Correction works only when voltage distinguishes SoC from resistance and polarization effects. In a flat OCV region, different SoC values produce similar voltage, so the filter remains reliant on integration. Hysteresis, thermal gradients, cell variation and parameter mismatch also cause bias.
Process-noise covariance represents uncertain model evolution; measurement-noise covariance represents observation uncertainty. Together they influence prediction-versus-measurement weighting. Tuning only for smooth output may conceal bias. Validation should inspect residuals, convergence, numerical stability and parameter sensitivity.
Coulomb Counting Versus Kalman Filtering: Practical Trade-Offs
Required information: Coulomb counting needs current, time, usable capacity and initial SoC. A model-based Kalman estimator generally needs these plus voltage measurements, an OCV curve, dynamic model parameters and uncertainty assumptions. Neither method works well with an uncalibrated current sensor.
Error behavior: Pure integration is vulnerable to accumulated offset and incorrect initialization. A Kalman estimator can reduce some of that drift when voltage contains independent information about SoC, but systematic model errors can produce persistent bias. There is no guarantee that adding a filter improves every operating regime.
Implementation burden: Coulomb counting is easy to audit and inexpensive to execute. EKF and UKF designs require parameter identification, matrix computations, noise tuning and safeguards for abnormal measurements or numerical failure. A more complex algorithm also demands a more comprehensive verification effort.
Response to operating changes: Current integration follows rapid charge transfer naturally. Model-based filters can account for voltage transients when their dynamic states are representative, but temperature changes, aging and unusual duty cycles may require adaptive parameters or scheduled recalibration.
Why Hybrid Estimation Is Often the Engineering Choice
The methods are complementary: many Kalman estimators integrate current for prediction, then use voltage for correction. A hybrid observer preserves charge tracking while limiting drift when voltage contains meaningful independent information.
Selection begins with requirements: cell chemistry, current profile, available sensing, embedded processing budget, target error metric, service lifetime and conditions for initialization. A simple counter may be adequate where periodic, trustworthy reference measurements exist. A model-based estimator becomes more attractive when current profiles vary and continuous uncertainty management matters, provided the model is validated for those conditions.
Pack-level reporting requires additional care. Series cells carry approximately the same current but may differ in capacity, impedance and SoC. A pack average can hide the cell that reaches its voltage limit first. For some systems, the most constrained cell matters more for usable pack capability than a single aggregate percentage. Estimation must remain distinct from cell balancing, protection thresholds and thermal supervision.
How to Calibrate and Validate an SoC Estimator
Begin with a documented test definition: initial conditioning, charge and discharge limits, ambient and cell temperature, rest periods and the usable-capacity reference. Establish OCV-versus-SoC data using an appropriate controlled procedure, recognizing relaxation and hysteresis. Identify equivalent-circuit resistance and polarization parameters at relevant temperatures and SoC regions rather than fitting a single trace and assuming universal validity.
Collect synchronized current, voltage and temperature measurements over realistic loading patterns. Include charge, discharge, idle periods, rapid transients, temperature variation and aged-cell conditions where appropriate. Preserve independent test sequences; evaluating a filter only on the same data used to tune parameters overstates confidence. Also challenge the estimator with sensor offset, missing samples and incorrect initial SoC.
Compare predictions against a defensible laboratory reference, such as charge throughput measured with calibrated instrumentation under a well-defined capacity test. Report mean absolute error and root-mean-square error in SoC percentage points, alongside maximum error and convergence time. State the reference method and its uncertainty. An estimator that has low average error may still be unsuitable if it fails near important operating limits.
Inspect voltage residuals, rest-period bias and long-duration drift. For Kalman implementations, check covariance behavior and ensure outlier handling does not conceal faults. Retest as cells age; report chemistry, test conditions and assumptions rather than universal accuracy claims.
Safety, Installation and System Boundaries
SoC is an estimate, not a safety interlock. It must not replace independent voltage, current and temperature protections or a correctly engineered battery management system. Unexpected voltage behavior, swelling, overheating or damaged wiring require established safety procedures rather than an attempt to recalibrate an algorithm during operation.
Battery packs can expose personnel to hazardous voltage, high fault current, fire and thermal-runaway risks. Cell testing, pack wiring, sensor installation and changes to protective controls should be performed only by qualified or licensed professionals using appropriate equipment and procedures. Do not bypass interlocks, dismantle energized packs or rely on a displayed SoC percentage to establish that a system is electrically safe.
Frequently Asked Questions
Is a Kalman filter more accurate than coulomb counting?
It can be, particularly when its voltage model offers independent information that corrects integration drift. Accuracy is conditional on parameter quality, sensor calibration, chemistry and duty cycle. In flat-OCV regions or with a poor model, the expected advantage may be small or absent.
Can voltage alone reveal battery SoC?
Sometimes an adequately rested cell’s OCV offers a useful estimate through a chemistry-specific calibration curve. Under load, resistance and polarization alter terminal voltage. Temperature, hysteresis and aging create further ambiguity, so voltage alone is not a universal real-time SoC gauge.
What happens when the battery ages?
Usable capacity and internal impedance change. Coulomb counting needs a refreshed capacity estimate, while model-based filters may need updated OCV or circuit parameters. State-of-health estimation can support those updates, but SoC and state of health remain different quantities.
What is the difference between an EKF and a UKF?
An EKF approximates nonlinear behavior using local derivatives; a UKF uses selected sample points to propagate nonlinear effects. Both require a suitable state-space model and realistic noise assumptions. Processor cost, numerical robustness and validation results should determine the choice.
Can an SoC estimator replace protective battery controls?
No. Even a well-validated estimate contains uncertainty and can fail under sensor faults or unmodeled conditions. Dedicated protection and monitoring functions must remain active and be designed and reviewed by qualified professionals.
Key takeaway: Coulomb counting is a sound foundation for tracking charge, while Kalman filtering adds model-based correction when measurements are informative. The useful engineering question is not which label sounds more advanced, but how each estimator behaves under the particular battery, sensors, aging profile and safety requirements being evaluated.
For company information and updates, visit the PhotoBattery Inc. homepage.
