PaperlayerView exact revision on arXiv

Source-linked AI summary

When voltage sensors fail: Electrochemically constrained fault-tolerant state estimation for flat-plateau LFP batteries

Feng Guo, Luis D. Couto, Hamid Hamed, Khiem Trad, Dong Zhang, Ru Hong, Guangdi Hu, Mohammadhosein Safari

arXiv:2610.05906v1eess.SYcs.SYeess.SP

TL;DR

On LFP’s flat voltage plateau, electrochemical-state errors and voltage abnormalities produce similar innovations, complicating SOC estimation. The paper coordinates state recovery and voltage-discrepancy accommodation with an electrochemically constrained RBC-DEKF and uncertain-initialization commissioning. Across the evaluated fault profiles, the RBC-DEKF substantially reduced mean SOC RMSE relative to Single-EKF.

  • Problem

    On LFP’s flat voltage plateau, electrochemical-state errors, model discrepancy, and voltage abnormalities can produce similar innovations, complicating SOC estimation.

    Source: §1 · pp. 3–5

  • Method

    The RBC-DEKF uses paired-electrode electrochemical states and a separately configurable voltage-discrepancy state, suspending residual adaptation during uncertain-initialization commissioning before state replay and handover.

    Source: §1 · p. 5, §3.3.3 · p. 25, §6 · p. 62

  • Results

    Mean SOC RMSE fell from 7.646 to 0.170 percentage points for 216 additive-bias profiles and from 22.646 to 0.356 percentage points for 168 multiplicative-gain profiles, relative to Single-EKF.

    Source: §6 · p. 62

  • Takeaways & Limitations

    The proposed approach coordinates uncertain-state recovery with subsequent voltage-discrepancy accommodation, and its commissioned realization retains frozen healthy-residual calibration after handover.

    Source: §3.3.3 · p. 25, §6 · p. 62

    A small posterior faulty-measurement fit error does not establish recovery of clean terminal voltage or physical sensor bias.

    Source: §5.5 · p. 61

Abstract

from arXiv · show

The flat voltage plateau of lithium iron phosphate (LFP)/graphite cells makes electrochemical-state errors and voltage-measurement abnormalities produce similar innovations, complicating state-of-charge (SOC) estimation. This work proposes an electrochemically constrained residual-bias compensation dual extended Kalman filter (RBC-DEKF) with uncertain-initialization commissioning. A thermal control-oriented parameter-grouped single-particle model (CPG-SPMT) provides paired-electrode dynamics and terminal-voltage prediction, while a separately configurable voltage-discrepancy state accommodates systematic residuals. When the initial SOC is uncertain, residual adaptation is suspended over a verified-healthy startup window. Buffered data support electrochemically constrained trajectory matching, followed by calibrated state replay and residual-channel reactivation. Frozen cross-cycle healthy-residual calibration supports commissioning and remains in the voltage prediction after handover. Evaluation covers 216 additive-bias and 168 multiplicative-gain profiles over 24 A123 temperature-drive-cycle trajectories. Relative to Single-EKF, the original RBC-DEKF reduces mean SOC RMSE from 7.646 to 0.170 percentage points for additive bias and from 22.646 to 0.356 percentage points for gain faults. Separately, 18 nonzero-initial-error profiles on three representative 50%-SOC-anchored segments give a full-record mean SOC RMSE of 1.747 percentage points, including startup, for the commissioned realization, versus 11.192 and 11.400 for the always-on RBC-DEKF and conventional joint EKF. Representative delayed bias, 10-s ramp, and gain tests show that the corrected SOC trajectory remains stable after fault introduction. The central contribution is electrochemically constrained temporal coordination of uncertain-state recovery and subsequent voltage-discrepancy accommodation.

1. Introduction

On LFP’s flat voltage plateau, electrochemical-state errors, model discrepancy, and sensor abnormalities can produce similar voltage innovations, complicating SOC estimation. The paper proposes coordinating electrochemical-state recovery with voltage-discrepancy accommodation, including a commissioning procedure for uncertain initialization.

1. Introduction

  • Voltage-sensor offsets and gain errors can map onto large SOC corrections in weakly observable regions, threatening BMS functions that depend on SOC.The paper identifies calibration drift, reference instability, component aging, and environmental influences as sources of sensor abnormalities.Source: §1 · p. 3
  • Existing residual-learning and region-dependent approaches do not explicitly estimate both voltage-sensor bias and gain online, while joint fault estimation remains dependent on model fidelity and excitation.Sensor-topology reconstruction also depends on measurement redundancy and fault configuration.Source: §1 · pp. 4–5
  • On the flat LFP plateau, state error, model discrepancy, and sensor abnormalities can produce similar voltage innovations, making residual interpretation and SOC correction difficult.Persistent voltage residuals may be explained as large SOC changes even when caused by sensor faults or model mismatch.Source: §1 · pp. 4–6
  • The proposed RBC-DEKF separates electrochemical-state and voltage-discrepancy recursions, suspending residual adaptation during healthy buffered startup before reactivating it after state correction.Its evaluation includes additive-bias and multiplicative-gain faults across 24 temperature–drive-cycle trajectories, plus representative initial-SOC-error and in-operation-abnormality tests.Source: §1 · pp. 5–6

2. Electrochemical and material basis for state estimation

The section develops the electrochemical and material basis for LFP state estimation: a reduced thermal single-particle model represents paired-electrode dynamics, while OCV sensitivity analysis shows why voltage mismatch can imply large SOC error.

2.1. Classical SPM foundation

  • The classical SPM represents solid-phase lithium diffusion and interfacial-reaction dynamics in both electrodes, with terminal voltage formed from electrode OCPs, charge-transfer overpotentials, and lumped ohmic loss.The model’s particle diffusion is governed by Fick’s law.Source: §2.1 · pp. 6–7

2.2. Control-oriented parameter-grouped electrochemical model

  • The CPG-SPMT approximates diffusion with parabolic concentration profiles and grouped parameters, retaining paired-electrode electrochemical structure for recursive estimation with temperature-dependent transport, kinetic, and ohmic terms.Its reduced states include both-electrode diffusion states and normalized concentrations used in the nonlinear voltage calculation.Source: §2.2 · pp. 8–10

2.3. Material-specific voltage sensitivity and SOC ambiguity of LFP

  • Finite OCV inversion yields maximum apparent SOC errors of approximately 16.4, 26.2, 33.5, and 37.3 pp for 5, 10, 20, and 30 mV biases, respectively.The plotted low-sensitivity windows use SU < 1 mV per SOC percentage point as a visualization criterion, not an RBC-DEKF threshold; dynamic hysteresis is not explicitly modeled.Source: §2.3 · pp. 13–15, Figure 1 · p. 14

3. Electrochemically constrained fault-tolerant state estimation

This section presents a fault-tolerant estimation framework that separates electrochemical state tracking from voltage-discrepancy estimation and coordinates their adaptation when initial SOC is uncertain. It describes the sequential residual-bias-compensation design, its distinction from a joint augmented EKF, filter settings, and voltage-fault tests, including healthy-constrained state recovery before residual-loop activation.

3.1. Problem formulation and measurement model

  • Additive-bias and gain faults are treated as voltage-domain discrepancies, but one scalar innovation cannot uniquely distinguish them from state error or model mismatch.The estimated compensation is a lumped discrepancy, not a pure sensor-bias estimate.Source: §3.1 · pp. 17–18

3.2. Residual-bias compensation dual extended Kalman filter (RBC-DEKF) framework

  • RBC-DEKF updates electrochemical states before a separate scalar loop estimates lumped voltage discrepancy, without identifying its physical source.Separate covariance recursions let state and residual adaptation be enabled at different times; how innovations are allocated remains tuning- and excitation-dependent.Source: §3.2 · pp. 20–22

3.3. Commissioning under uncertain initial SOC

  • For uncertain initial SOC, commissioning suspends residual adaptation during a verified-healthy startup window, selects an electrochemically consistent state, then replays it before residual accommodation resumes.A startup sensor abnormality is outside this protocol.Source: §3.3 · p. 23, §3.3.1 · p. 23

3.3.1. Commissioning principle

  • During commissioning, a four-state EKF buffers current, temperature, and voltage while a frozen temperature-specific map predicts healthy-model residuals from SOC and causal current history.The target drive cycle is excluded when fitting the map, which remains distinct from the adaptive discrepancy state.Source: §3.3.1 · p. 23

3.3.2. Electrochemically constrained state correction

  • The correction searches paired-electrode SOC-shift candidates and selects the initial state minimizing calibrated voltage mismatch over CPG-SPMT rollouts.Candidate shifts preserve within-electrode reduced-state differences; the bounded search uses an ordinary L2 loss.Source: §3.3.2 · pp. 23–24

3.3.3. Replay and handover

  • The selected state is replayed through buffered samples using the calibrated EKF; afterward, the scalar residual loop restarts from zero while calibration remains in voltage predictions.Retaining calibration avoids reintroducing repeatable healthy-model discrepancy at handover.Source: §3.3.3 · pp. 24–25

3.4. Comparison estimators

  • Comparators distinguish open-loop propagation, Single-EKF voltage correction, always-on sequential RBC-DEKF, conventional joint EKF, and commissioned RBC-DEKF.The commissioned estimator applies the uncertain-initialization procedure before activating residual accommodation.Source: §3.4 · pp. 25–26
  • The comparison baseline is an always-on five-dimensional joint EKF initialized with zero voltage-discrepancy state and using a single voltage measurement covariance.Its measurement covariance is set to Rx, matching the voltage weighting used by the electrochemical-state loop.Source: §3.4 · pp. 26–27

3.5. Sequential RBC-DEKF versus joint EKF: covariance coupling and commissioning capability

  • Sequential RBC-DEKF exposes independently controllable state and residual updates, enabling healthy-state recovery before discrepancy-loop activation, unlike the covariance-coupled joint EKF.The two separately tuned updates use the same voltage sample, so RBC-DEKF is not a covariance-equivalent factorization of the joint EKF.Source: §3.5 · pp. 27–28

3.6. Filter settings and implementation details

  • Filter covariances were empirically selected and fixed across operating conditions and fault levels, with uncertain-startup covariance enlarged only along the SOC-aligned direction.The large initial residual covariance supports early adaptation; Qθ supports response to systematic voltage deviations, while Rθ regularizes updates against measurement noise.Source: §3.6 · p. 30

3.7. Fault injection protocol

  • Fault tests inject additive bias or multiplicative gain into noisy voltage measurements, with additional mid-trajectory profiles evaluating accommodation rather than automatic detection.The voltage noise has 1 mV standard deviation; a ramp reaches +30 mV after 10 s, and commissioned RBC-DEKF activates its residual loop after the verified-healthy 300 s interval without receiving fault-onset time.Source: §3.7 · pp. 30–31

4. Results

Results assess RBC-DEKF SOC accuracy across additive-bias and gain-fault matrices and evaluate commissioning that suspends residual adaptation during uncertain-initialization recovery before accommodating voltage abnormalities. Tests span 24 temperature–drive-cycle trajectories and include representative faults, cross-condition robustness, initialization stress tests, and commissioning-window evaluations.

4.1. Experimental settings and evaluation scenarios

  • Evaluation used 24 trajectories spanning eight temperatures and three drive cycles, with 216 additive-bias and 168 gain-fault profiles per applicable estimator.Initialization tests used three 50%-SOC-anchored segments; voltage measurements included 1 mV standard-deviation white noise. Cross-cycle healthy-residual calibration had a 9.902 mV equal-weight macro-mean held-out RMSE, distinct from estimator voltage-fit error.Source: §4.1 · pp. 32, 33, 35, Table 7 · p. 36

4.2. Overall fault-tolerant performance under bias and gain faults

  • SOC RMSE fell from 7.646 pp for Single-EKF to 0.170 pp for original RBC-DEKF across 216 additive-bias profiles, and from 22.646 to 0.356 pp across 168 gain profiles.These are mean SOC RMSE comparisons over the static fault matrices; posterior faulty-measurement fit RMSE is not a clean-voltage recovery metric.Source: §4.2 · p. 37, Figure 4 · p. 38, Table 9 · p. 39

4.3. Representative results under additive bias faults

  • At 25°C under DST with +30 mV bias, SOC RMSE fell from 17.043 to 0.062 pp and posterior faulty-measurement fit RMSE from 20.98 to 0.12 mV.Across DST, US06, and FUDS at this temperature and bias, original RBC-DEKF SOC RMSE values were 0.062, 0.060, and 0.075 pp, respectively.Source: §4.3 · pp. 40–41, Figure 5 · p. 41, Table 11 · p. 44

4.4. Performance under multiplicative gain faults

  • At 25°C under DST with +3% gain, SOC RMSE fell from 47.402 to 0.052 pp and posterior faulty-measurement fit RMSE from 45.95 to 0.16 mV.The scalar residual accommodates the gain-induced voltage contribution without explicit gain-coefficient identification; under US06, compensated SOC RMSE was 3.899 pp versus 47.547 pp for Single-EKF.Source: §4.4 · p. 42, Figure 6 · p. 43, Table 11 · p. 44

4.5. Robustness across temperatures and dynamic drive cycles

  • Across the tested temperature–fault plane, additive-bias SOC RMSE stayed below 0.7 pp; for +3% gain it reached 4.604 pp at 20°C but was at most 1.398 pp at other temperatures.For +3% gain, the cycle-wise values were 0.220, 0.403, and 3.161 pp under DST, FUDS, and US06; the results describe in-matrix robustness, not out-of-sample generalization.Source: §4.5 · pp. 44–45, Figure 7 · p. 46, Table 13 · p. 48

4.6. State recovery and fault tolerance under initial-SOC uncertainty

  • Full-record mean SOC RMSE was 1.747 pp for commissioned RBC-DEKF, versus 11.192 pp for original RBC-DEKF and 11.400 pp for conventional joint EKF.The metric includes startup across six nonzero initial-error profiles on three 50%-SOC-anchored segments. During a verified-healthy window, the commissioned method keeps the residual state at zero, then applies calibrated correction and causal state-EKF replay; the illustrated post-300-s SOC RMSE is 0.671 pp.Source: §4.6 · pp. 48, 50, Table 14 · p. 49, Figure 9 · p. 49
  • For uncertain initial SOC, commissioning suspends the residual state at zero until the 300-s handover; reliable initial priors bypass commissioning, and replay does not backfill earlier outputs.In the three zero-error controls, one trajectory received a 0.5-pp correction from 180 s onward.Source: Figure 10 · p. 51, §4.6.1 · p. 53, §4.7 · p. 54

4.6.1. Commissioning-window sensitivity

  • Correction-error RMSE falls to 0.277 pp at 180 s and remains there at 300 and 600 s; 300 s was retained as a practical, non-optimal setting.The 180-s window is the earliest tested to reach that stable level; the same 300-s window was used in comparative experiments without case-wise retuning.Source: §4.6.1 · pp. 51, 53
  • Across 180–600 s commissioning windows, SOC-aligned correction accuracy is stable, while commissioned-estimator RMSE ranges from 0.965 to 1.122 pp over the common t ≥600 s horizon.The sensitivity study assesses robustness rather than optimizing the window length.Source: §4.6.1 · p. 53

4.7. Fault accommodation after uncertain-state commissioning

  • 0.697–0.715 percentage points: after correcting the startup SOC offset, commissioned RBC-DEKF maintained stable SOC across delayed-step, 10-s ramp, and gain faults.The representative tests used a +10-percentage-point initial-SOC perturbation; the fault onset was not supplied to the estimators.Source: §4.7 · pp. 53–54, Table 16 · p. 54, Figure 12 · p. 55

5. Discussion

The discussion contrasts joint and dual filter behavior, then explains how commissioning addresses uncertain initialization without creating intrinsic SOC observability. It also frames the discrepancy state as a voltage-domain accommodation whose interpretation and validation remain bounded by model mismatch, chemistry, and test scope.

5.1. Why a dual realization: functional decoupling versus joint augmentation

  • Across 384 static condition–fault profiles, the conventional joint EKF is more accurate than the original RBC-DEKF (0.092811 versus 0.251308 percentage points), while sequential updates can pass discrepancy-dominated innovation into SOC correction.The commissioned realization can suspend the residual loop while retaining the four-state EKF recursion, supporting the combined startup procedure rather than a universal ranking of dual and joint filters.Source: §5.1 · pp. 57–58

5.2. Initial-state uncertainty and intrinsic observability

  • Commissioning constrains startup correction to model-generated, SOC-aligned trajectories under a verified-healthy window; it does not create instantaneous SOC observability or guarantee unique recovery.A fault present at startup or inadequate healthy-model calibration leaves sensor and model contributions non-identifiable from the single voltage channel; finite current-driven trajectories help rank candidate states.Source: §5.2 · pp. 58–59

5.3. Equivalent-residual interpretation of multiplicative gain faults

  • The scalar discrepancy state can represent a gain fault’s voltage-domain effect without identifying its coefficient; additive and gain faults are accommodated similarly when equivalent residuals are comparable.Remaining errors cluster where gain residuals are larger and more time-varying, while sign asymmetry is associated with nonlinear voltage–state sensitivity and fixed covariance tuning.Source: §5.3 · pp. 59–60

5.4. Implications for flat-plateau battery chemistries

  • A separate voltage-discrepancy channel complements electrochemical state estimation for LFP’s flat plateau, but quantitative compensation depends on chemistry-specific voltage–composition slope, polarization, hysteresis, and model fidelity.Extension to other flat-voltage or phase-transition-dominated systems requires chemistry-specific parameterization and independent validation.Source: §5.4 · p. 60

5.5. Model mismatch, hysteresis, and residual interpretation

  • The estimated residual is lumped, not a source-isolated sensor bias: hysteresis, reduced-model relaxation error, parameter error, and sensor abnormalities can enter the same scalar state.A low posterior faulty-measurement fit error does not establish recovery of clean voltage or physical sensor bias; source isolation needs additional information.Source: §5.5 · pp. 60–61

5.6. Practical implications and future extensions

  • A one-state discrepancy loop improves electrochemical-model SOC robustness for the tested additive-bias and gain abnormalities without sensor redundancy or SOC-region partitioning; explicit fault-parameter identification is unnecessary for this accommodation.The estimated residual alone does not establish or classify sensor faults; diagnostic decisions require a separately validated baseline, thresholds, or classification layer.Source: §5.6 · p. 61
  • Validation is bounded by synthetic abnormalities, fixed-health assumptions, unevaluated current-sensor and joint faults, three commissioning trajectories, and in-matrix rather than out-of-sample static coverage.The commissioning study also uses one noise realization and a supervised cross-cycle healthy-residual map.Source: §5.6 · pp. 61–62

6. Conclusion

The electrochemically constrained RBC-DEKF coordinates paired-electrode state estimation with a separately configurable voltage-discrepancy loop, and commissioning coordinates uncertain-state recovery with later residual adaptation. Across the reported fault and initialization evaluations, it reduced SOC RMSE relative to the stated baselines.

6. Conclusion

  • Mean SOC RMSE fell from 7.646 to 0.170 percentage points for 216 additive-bias profiles and from 22.646 to 0.356 percentage points for 168 multiplicative-gain profiles.The evaluation covered 24 temperature–drive-cycle trajectories.Source: §6 · p. 62
  • Full-record mean SOC RMSE, including startup, was 1.747 pp for the commissioned realization, versus 11.192 pp for always-on RBC-DEKF and 11.400 pp for conventional joint EKF.These results cover 18 paired nonzero-initial-error profiles on three 50%-SOC-anchored segments.Source: §6 · p. 62
  • Representative delayed step, 10-s ramp, and gain tests yielded post-fault RMSE values of 0.697–0.715 pp, with stable correction over tested 180–600-s windows.The reported window results support the retained 300-s setting.Source: §6 · p. 62
  • Commissioning suspends residual estimation, selects electrochemically consistent paired-electrode initial states from verified-healthy buffered data, then replays the calibrated state EKF before handover.Residual adaptation is reactivated after the calibrated state replay.Source: §6 · pp. 62–63
  • The reported commissioned accuracy includes frozen healthy-model calibration; physical sensor-fault validation, aging-aware calibration, and fault-source classification remain future work.The frozen cross-cycle calibration remains in the voltage prediction after handover.Source: §6 · p. 63

Appendix A.1. Data matrix and parameterization

The CPG-SPMT parameterization was identified from the complete A123 temperature–drive-cycle matrix. It combines temperature-linked parameters with electrode-specific OCP expressions.

Appendix A.1. Data matrix and parameterization

  • Identification used 24 measured trajectories: DST, FUDS, and US06 at −10, 0, 10, 20, 25, 30, 40, and 50°C.The parameter vector includes αn,1, αp,1, bn, bp, dn,1, dp,1, sn,0, sp,0, E1–E5, and Rini,1.Source: Appendix A.1 · p. 63
  • The five activation energies E1–E5 couple reference transport, kinetic, and ohmic parameters across temperature through Arrhenius relations.The universal gas constant, Faraday’s constant, and electrolyte concentration are treated as known and fixed.Source: Appendix A.1 · p. 63
  • The positive- and negative-electrode OCP expressions follow the CPG-SPMT formulation in Ref..Source: Appendix A.1 · p. 63

Appendix A.2. Bound-constrained multi-start optimization

Appendix A.2 describes a bound-constrained multi-start refinement and cross-condition selection workflow for the A123 LFP parameterization. The fit uses the available measured matrix, which is not an independent held-out validation set.

Appendix A.2. Bound-constrained multi-start optimization

  • The workflow samples initial candidates within prescribed bounds, refines each with a bound-constrained trust-region reflective least-squares solver, and ranks refined candidates by the unweighted mean of condition-wise voltage MAEs.The refinement objective uses an equal-condition mean so longer trajectories do not receive greater weight solely from having more samples.Source: Appendix A.2 · pp. 63–64
  • The optimizer is chemistry-independent, but model outputs, OCP functions, stoichiometric windows, bounds, and selected parameters remain LFP-specific; multi-start refinement does not prove global convergence or parameter uniqueness.Source: Appendix A.2 · p. 64
  • Table A.1 reports the supplied bounds for the 14-parameter A123 LFP/graphite identification and the selected values used in the study.Source: Table A.1 · p. 64
  • The selected bn value is 10999.93 C, near its 11000 C upper bound, so it is not a tightly isolated physical parameter without additional excitation or profile-likelihood analysis.Source: Appendix A.2 · p. 64
  • All 24 measured trajectories enter refinement and/or candidate selection, so the fit is cross-condition consistency over the available matrix, not independent held-out validation.Fault-injection tests reusing these trajectories assess estimator behavior under controlled measurement perturbations, not out-of-sample parameter generalization.Source: Appendix A.2 · p. 64

Appendix A.3. SOC-dependent raw voltage-residual audit

The SOC-dependent residual audit recomputes raw voltage residuals across the complete temperature–cycle records and summarizes them by SOC interval. Figure A.1 displays residual statistics with trajectory coverage indicated.

Appendix A.3. SOC-dependent raw voltage-residual audit

  • Raw residuals were recomputed at every sample of all 24 complete temperature–cycle records to examine SOC dependence.Source: Appendix A.3 · p. 65
  • Reference SOC comes from coulomb counting from the 100% record initialization; samples are grouped into ten left-closed SOC intervals, with the final interval including 100%.Source: Appendix A.3 · p. 65
  • Figure A.1 plots equal-weight trajectory averages of signed residual statistics and RMSE by SOC bin, with descriptive inter-trajectory standard deviations and contributing-trajectory counts.Coverage below 20% SOC is incomplete because several low-temperature records terminate earlier; lower RMSE whiskers are truncated at zero.Source: Figure A.1 · p. 65, Appendix A.3 · p. 65

Appendix B.1. Calibration implementation

The auxiliary calibration learns repeatable healthy CPG-SPMT voltage discrepancy from other drive cycles at the target temperature. Its frozen sample-by-SOC lookup is queried at estimated SOC during estimator operation.

Appendix B.1. Calibration implementation

  • Healthy samples from the other two drive cycles at the same temperature train a regressor to predict the target cycle’s voltage discrepancy in mV.Training uses open-loop CPG-SPMT propagation, SOC and current-history features, and a histogram-gradient-boosting regressor.Source: Appendix B.1 · p. 66
  • The fitted model is evaluated over a 0–100% SOC grid at 1-percentage-point spacing to create a sample-by-SOC lookup table.The measured target-cycle current through sample k supplies current features; only the SOC coordinate is swept, and the table is interpolated at estimated SOC during operation.Source: Appendix B.1 · p. 66
  • The frozen healthy calibration is distinct from the adaptive online voltage-discrepancy state; Coulomb-counted reference SOC is used only for offline fitting and audit.During estimator operation, the lookup uses model-estimated SOC.Source: Appendix B.1 · p. 66
  • The implemented EKF gain uses neighboring-grid slope times the paired-electrode SOC Jacobian, retaining the nearest interior slope outside the lookup range as an approximation.The lookup value is held at the nearest endpoint outside range; the exact derivative there would be zero, and the features can be constructed causally from current measurements through sample k.Source: Appendix B.1 · pp. 66–67

Appendix B.2. Calibration contribution

The calibration ablation compares two otherwise matched commissioned configurations over 18 nonzero-initial-error profiles, with full-record RMSE including startup. Calibration contributes to lower mean SOC RMSE, though the ablation does not isolate its search-stage effect.

Appendix B.2. Calibration contribution

  • 1.747 pp versus 8.665 pp mean SOC RMSE: adding healthy-residual calibration reduces error across the 18 paired profiles, including startup.The comparison retains suspension, trajectory correction, causal replay, and state-first handover in both configurations.Source: Table B.1 · p. 67, Appendix B.2 · p. 67
  • On the flat LFP plateau, a repeatable healthy-model voltage bias can shift candidate SOC and distort initial-state ranking; accounting for it reduces that influence.The full-record ablation measures calibration’s combined contribution in replay and subsequent updates, not its search-stage effect alone.Source: Appendix B.2 · p. 67

CRediT authorship contribution statement

The contribution statement assigns Guo a broad set of research and writing roles, with methodology shared by three coauthors and supervision and resources credited to Safari.

CRediT authorship contribution statement

Loading 2610.05906v1…