pith. sign in

arxiv: 2606.12000 · v1 · pith:LJ4LBTDHnew · submitted 2026-06-10 · 📡 eess.SY · cs.SY

Physics-guided residual Kalman learning for state-of-charge estimation of lithium iron phosphate batteries

Feng Guo , Luis D. Couto , Khiem Trad , Ru Hong , Guangdi Hu , Mohammadhosein Safari This is my paper

Pith reviewed 2026-06-27 08:32 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords state-of-charge estimationlithium iron phosphate batteriesextended Kalman filtergated recurrent unitresidual learningelectrochemical modelhybrid estimationbattery management systems
0
0 comments X

The pith

A hybrid filter merges an electrochemical model with a recurrent network to reduce SOC estimation error by 77 percent for LFP batteries.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a method that merges a physics-based extended Kalman filter using a single-particle electrochemical model with a gated recurrent unit network that learns to correct the filter's structured errors. This hybrid approach is tested on data from lithium iron phosphate batteries across multiple drive cycles, temperatures from -10 to 50 degrees Celsius, and initial state-of-charge offsets up to 20 percent. The combination achieves a global average root mean square error of 1.19 percent in state-of-charge estimates. A sympathetic reader would care because accurate SOC estimation supports safe battery operation in electric vehicles and storage systems, especially given the flat voltage curves of LFP cells. The results show that feeding electrochemical states into the residual learner improves recursive estimation without replacing the physical model entirely.

Core claim

The central claim is that a gated recurrent unit residual learner, trained to compensate structured errors in a single-particle-model-based extended Kalman filter by using electrochemical states and measurement features, yields a global average RMSE of 1.19 percent for SOC estimation across the dataset, a 77 percent reduction relative to the physics-only EKF, while preserving recursive propagation and showing cross-profile robustness when trained on DST and FUDS cycles and tested on US06 within the same cell dataset.

What carries the argument

The physics-guided residual Kalman learning (PRKL) framework, in which the GRU residual learner receives electrochemical states from the EKF to correct its estimation errors while the EKF handles recursive physical state propagation.

If this is right

  • The hybrid method maintains the recursive, real-time state propagation of the EKF while adding correction for temperature-dependent and initialization-sensitive mismatches.
  • Electrochemical state information from the single-particle model guides the residual learner to handle flat OCV-SOC characteristics effectively.
  • Cross-profile testing shows that training on DST and FUDS cycles allows the framework to generalize to the US06 cycle within the studied dataset.
  • The approach supports validation on initialization offsets up to 20 percent and temperatures from -10 to 50 degrees Celsius.
  • Results provide a basis for extending to ageing-aware and embedded-platform implementations as stated in the paper.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same residual-correction idea could be applied to other battery chemistries that exhibit flat voltage regions or strong temperature sensitivity.
  • Embedding the physical states directly into the learner may reduce the volume of data needed compared with fully data-driven SOC estimators.
  • If the GRU correction remains stable over long sequences, the method could lower the frequency of open-circuit voltage resets in battery management systems.
  • Future tests on multi-cell packs would show whether cell-to-cell variation requires retraining the residual learner or only retuning the base EKF.

Load-bearing premise

The assumption that EKF errors remain structured enough for the GRU to learn and correct them reliably when generalizing from training cycles to an unseen test cycle within the same cell dataset without creating new instabilities.

What would settle it

Applying the trained PRKL model to data from a different LFP cell or an aged cell under the same temperature and cycle conditions and observing that the RMSE does not drop below the physics-only EKF level would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2606.12000 by Feng Guo, Guangdi Hu, Khiem Trad, Luis D. Couto, Mohammadhosein Safari, Ru Hong.

Figure 1
Figure 1. Figure 1: Overview of the proposed PRKL framework for SOC estimation of LFP battery. (a) The control-oriented parameter-grouped single particle model (CPG-SPMT) is embedded in an EKF to generate the baseline SOC estimate, voltage innovation, pre-update model voltage, and electrode￾level states from measured current, voltage, and temperature. (b) Residual learning data are constructed using Coulomb-counting reference… view at source ↗
Figure 2
Figure 2. Figure 2: CPG-SPMT model validation and PRKL-enhanced SOC estimation performance. (a, b) Model validation: RMSE (a) and MAE (b) of terminal voltage predictions across temperatures (from −10 to 50 °C) and drive cycles (DST, US06, and FUDS). Dashed lines indicate mean values across all conditions. (c–f) SOC estimation comparison: (c) Temperature-dependent RMSE (%) for EKF (gray) and PRKL (blue); shaded areas represent… view at source ↗
Figure 3
Figure 3. Figure 3: Temporal comparison of SOC estimation across representative operating conditions. Each row corresponds to a distinct drive cycle and temperature: (a, b) DST at −10 °C, (c, d) US06 at 30 °C, and (e, f) FUDS at 50 °C. Left panels show the estimated SOC trajectories by EKF (gray-orange-red dashed lines) and PRKL (light-medium-dark blue solid lines) under initial SOC errors of 0%, 10%, and 20%, compared to the… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of physics information on PRKL for battery SOC estimation. (a) Violin/box plots of absolute SOC RMSE (%) for a classical EKF (grey), an ablated model without physics PRKL-NP (orange), and the physics-guided PRKL (blue). (b) ECDFs of absolute RMSE (%) for the same three methods. (c) Heat map of RMSE reduction (%) of PRKL relative to PRKL-NP across temperature (y-axis) and drive cycle (x-axis; DST, FU… view at source ↗
read the original abstract

Accurate state of charge (SOC) estimation of lithium iron phosphate (LFP) batteries remains challenging because of their flat open-circuit-voltage (OCV)-SOC characteristics, temperature-dependent dynamics, and sensitivity to initialization errors. Here, we propose a physics-guided residual Kalman learning (PRKL) framework for electrochemical-model-based SOC estimation. PRKL combines a control-oriented single-particle-model-based extended Kalman filter (EKF), which provides recursive physical state propagation, with a gated recurrent unit (GRU) residual learner that compensates structured EKF errors using electrochemical states and measurement features. The framework is evaluated on a public graphite/LFP dataset covering three dynamic drive cycles, eight temperatures from -10 to 50 degrees C, and initialization offsets up to 20 percent. Using dynamic stress test (DST) and federal urban driving schedule (FUDS) cycles for training and the supplemental federal test procedure (US06) cycle for cross-profile testing within the same cell dataset, PRKL achieves a global average root mean square error (RMSE) of 1.19 percent, corresponding to a 77 percent reduction relative to the physics-only EKF. These results show that electrochemical state information can guide residual learning and improve recursive SOC estimation for LFP batteries. The present validation supports cross-profile robustness within the studied dataset and provides a basis for future cross-cell, ageing-aware, and embedded-platform validation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a physics-guided residual Kalman learning (PRKL) framework that augments a single-particle-model-based extended Kalman filter (EKF) with a gated recurrent unit (GRU) residual learner to compensate structured EKF errors using electrochemical states and measurements for SOC estimation in LFP batteries. It reports evaluation on a public graphite/LFP dataset across eight temperatures and initialization offsets up to 20%, with training on DST/FUDS cycles and cross-profile testing on the US06 cycle within the same cell, achieving a global average RMSE of 1.19% (77% reduction relative to physics-only EKF).

Significance. If the reported improvement is shown to be robust, the work illustrates how embedding electrochemical state information into residual learning can enhance recursive SOC estimation for LFP cells with flat OCV-SOC curves. The use of a public dataset and explicit cross-profile split provides a concrete, reproducible baseline for extensions to cross-cell or ageing-aware settings.

major comments (2)
  1. [Abstract] Abstract: the 77% RMSE reduction claim is load-bearing for the central contribution, yet the abstract supplies no information on training procedure, hyperparameter selection, error bars, statistical significance testing, or checks for overfitting; without these, it is impossible to determine whether the reported gain is reliable or an artifact of the single held-out profile.
  2. [Evaluation] Evaluation description: the cross-profile test trains the GRU on DST/FUDS and evaluates on US06 within the same cell; the manuscript provides no analysis of whether feeding GRU-corrected states back into the EKF recursion can destabilize the filter (e.g., via altered covariance propagation or divergence), which directly affects whether the 1.19% RMSE generalizes beyond the reported split.
minor comments (1)
  1. [Abstract] The phrase 'global average' is used without clarifying whether it is taken across all temperatures, all initialization offsets, or both; a brief clarification would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comments point-by-point below and indicate the revisions we will make to strengthen the paper.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the 77% RMSE reduction claim is load-bearing for the central contribution, yet the abstract supplies no information on training procedure, hyperparameter selection, error bars, statistical significance testing, or checks for overfitting; without these, it is impossible to determine whether the reported gain is reliable or an artifact of the single held-out profile.

    Authors: The abstract is intentionally concise to highlight the key contribution and results. Detailed information on the training procedure (DST/FUDS for training, US06 for testing), hyperparameter selection via cross-validation, and the public dataset is provided in the Methods section. To address the concern, we will revise the abstract to include a short phrase on the evaluation protocol and note that the improvement is observed consistently across eight temperatures. We will also add error bars (standard deviation over multiple random seeds) to the reported RMSE in the results section and discuss the lack of overfitting due to the cross-profile split. These changes will make the reliability of the 77% reduction more transparent. revision: yes

  2. Referee: [Evaluation] Evaluation description: the cross-profile test trains the GRU on DST/FUDS and evaluates on US06 within the same cell; the manuscript provides no analysis of whether feeding GRU-corrected states back into the EKF recursion can destabilize the filter (e.g., via altered covariance propagation or divergence), which directly affects whether the 1.19% RMSE generalizes beyond the reported split.

    Authors: We appreciate this important point regarding potential filter instability. In our implementation, the GRU residual is added to the SOC estimate after the EKF update step, without modifying the internal covariance propagation, which helps maintain stability. However, we acknowledge that an explicit analysis of the closed-loop behavior was not included. In the revised manuscript, we will add a discussion on the integration approach and include additional results showing the innovation statistics and filter consistency metrics over the test cycle to demonstrate no divergence occurred. This will provide stronger evidence for the generalizability of the 1.19% RMSE. revision: yes

Circularity Check

0 steps flagged

No significant circularity in PRKL derivation or evaluation

full rationale

The paper's claimed improvement derives from training a GRU residual learner on DST/FUDS cycles to correct EKF residuals, then evaluating the hybrid estimator on a held-out US06 cycle within the same cell dataset. The reported 1.19% RMSE is an empirical test-set metric, not a quantity defined in terms of itself or forced by the training fit. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the abstract or described framework; the physics EKF and data-driven residual components remain distinct, with explicit train/test splits. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that EKF errors are structured and learnable from the listed features; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption EKF errors in the single-particle model are structured and can be compensated by a GRU using electrochemical states and measurement features.
    This premise enables the residual learner to improve upon the physics-only baseline.

pith-pipeline@v0.9.1-grok · 5796 in / 1384 out tokens · 23585 ms · 2026-06-27T08:32:08.639006+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 1 canonical work pages

  1. [1]

    Physics-guided residual Kalman learning for state-of-charge estimation of lithium iron phosphate batteries Feng Guoa,b,c, Luis D. Coutoa,b, Khiem Trada,b, Ru Hongd,∗, Guangdi Hue,∗, Mohammadhosein Safarib,c,f,∗ aWET, VITO, Boeretang 200, Mol, 2400, Belgium bEnergyVille, Thor Park 8310, Genk, 3600, Belgium cInstitute for Materials Research (IUMAT), Hasselt...

    work page doi:10.1016/j.jechem.2026.05.040 2026

  2. [2]

    Key hyperparameters and configuration of the physics-guided GRU residual learner. Component Configuration/value Input features 9 (electrochemical, EKF, and measurement-based features) Window length/stride 60 steps/30 steps (50% overlap), yielding a total of 11,277 training windows Hidden layer type Single-layer GRU Hidden units 32 Output dimension 1 (resi...

    2080

  3. [3]

    J. Zhao, X. Han, Y . Wu, Z. Wang, A.F. Burke, J. Energy Chem. 102 (2025) 463–496

    2025

  4. [4]

    Schoberl, M

    J. Schoberl, M. Ank, M. Schreiber, N. Wassiliadis, M. Lienkamp, eTransportation 19 (2024) 100305

    2024

  5. [5]

    H. Liu, C. Li, X. Hu, J. Li, K. Zhang, Y . Xie, R. Wu, Z. Song, Nat. Commun. 16 (2025)

    2025

  6. [6]

    K. Luo, X. Chen, H. Zheng, Z. Shi, J. Energy Chem. 74 (2022) 159–173

    2022

  7. [7]

    Z.P. Cano, D. Banham, S. Ye, A. Hintennach, J. Lu, M. Fowler, Z. Chen, Nat. Energy 3 (2018) 279–289

    2018

  8. [8]

    Harper, R

    G. Harper, R. Sommerville, E. Kendrick, L. Driscoll, P. Slater, R. Stolkin, A. Walton, P. Christensen, O. Heidrich, S. Lambert, A. Abbott, K. Ryder, L. Gaines, P. Anderson, Nature 575 (2019) 75–86

    2019

  9. [9]

    Zhang, Y

    Z. Zhang, Y . Li, H. Wang, L. Lu, X. Han, D. Li, M. Ouyang, J. Energy Storage 75 (2024) 109696

    2024

  10. [10]

    Guo, L.D

    F. Guo, L.D. Couto, G. Mulder, K. Trad, G. Hu, O. Capron, K. Haghverdi, J. Energy Storage 101 (2024) 113850

    2024

  11. [11]

    Haghverdi, D.L

    K. Haghverdi, D.L. Danilov, G. Mulder, L.D. Couto, F. Guo, R.-A. Eichel, J. Power Sources 652 (2025) 237483

    2025

  12. [12]

    L. Ma, T. Zhang, J. Energy Chem. 80 (2023) 48–57

    2023

  13. [13]

    Y . Che, L. Xu, R. Teodorescu, X. Hu, S. Onori, ACS Energy Lett. 10 (2025) 741–749

    2025

  14. [14]

    Nekahi, K

    A. Nekahi, K. Zaghib, Energy Storage Mater. 84 (2026) 104759

    2026

  15. [15]

    Safari, C

    M. Safari, C. Delacourt, J. Electrochem. Soc. 158 (2011) A562–A571

    2011

  16. [16]

    C. Wang, M. Yang, X. Wang, Z. Xiong, F. Qian, C. Deng, C. Yu, Z. Zhang, X. Guo, J. Energy Storage 110 (2025) 115346

    2025

  17. [17]

    L. Wu, Z. Lyu, Z. Huang, C. Zhang, C. Wei, J. Energy Chem. 89 (2024) 27–40

    2024

  18. [18]

    Fuller, M

    T.F. Fuller, M. Doyle, J. Newman, J. Electrochem. Soc. 141 (1994) 1–10

    1994

  19. [19]

    Guo, L.D

    F. Guo, L.D. Couto, J. Power Sources 650 (2025) 237365

    2025

  20. [20]

    Guo, L.D

    F. Guo, L.D. Couto, J. Power Sources 649 (2025) 237309

    2025

  21. [21]

    Plett, J

    G.L. Plett, J. Power Sources 134 (2004) 277–292

    2004

  22. [22]

    Z. Chen, H. Sun, G. Dong, J. Wei, J. Wu, J. Power Sources 414 (2019) 158–166

    2019

  23. [23]

    Choobar, H

    B.G. Choobar, H. Hamed, M. Safari, J. Energy Storage 96 (2024) 112628

    2024

  24. [24]

    W. Qian, W. Li, X. Guo, H. Wang, Energy 292 (2024) 130585

    2024

  25. [25]

    Q. Zhu, L. Li, X. Hu, N. Xiong, G.-D. Hu, IEEE Trans. Veh. Technol. 66 (2017) 10853–10865

    2017

  26. [26]

    Xiong, D

    Y . Xiong, D. Zhang, X. Ruan, S. Jiang, X. Zou, W. Yuan, X. Liu, Y . Zhang, Z. Nie, D. Wei, Y . Zeng, P. Cao, G. Zhang, Energy Storage Mater. 73 (2024) 103860

    2024

  27. [27]

    X. Ren, S. Liu, X. Yu, X. Dong, Energy 234 (2021) 121236

    2021

  28. [28]

    Hannan, D.N

    M.A. Hannan, D.N. How, M.H. Lipu, M. Mansor, P.J. Ker, Z. Dong, K. Sahari, S.K. Tiong, K.M. Muttaqi, T.I. Mahlia, F. Blaabjerg, Sci. Rep. 11 (2021) 19541

    2021

  29. [29]

    W. Qi, W. Qin, Z. Yun, Energy 307 (2024) 132805

    2024

  30. [30]

    Kim, K.-H

    K.-H. Kim, K.-H. Oh, H.-S. Ahn, H.-D. Choi, IEEE Trans. Power Electron. 39 (2023) 125–134

    2023

  31. [31]

    Korkmaz, Energy 294 (2024) 130913

    M. Korkmaz, Energy 294 (2024) 130913

    2024

  32. [32]

    J. Tian, C. Chen, W. Shen, F. Sun, R. Xiong, Energy Storage Mater. 61 (2023) 102883

    2023

  33. [33]

    W. Ma, Y . Lei, X. Wang, B. Chen, J. Energy Chem. 80 (2023) 768–784

    2023

  34. [34]

    Sorouri, A

    H. Sorouri, A. Oshnoei, Y . Che, R. Teodorescu, J. Energy Storage 100 (2024) 113604

    2024

  35. [35]

    Raabe, J.R

    D. Raabe, J.R. Mianroodi, J. Neugebauer, Nat. Comput. Sci. 3 (2023) 198–209

    2023

  36. [36]

    J. Chen, T. Hannan, Y . Yao, G. Song, Energy Storage Mater. 72 (2024) 103730

    2024

  37. [37]

    H. Zhao, Q. Li, J. Hu, J. Energy Storage 124 (2025) 116841

    2025

  38. [38]

    Rezaei, A

    O. Rezaei, A. Rahdan, S. Sardari, M. Dahmardeh, Z. Wang, J. Energy Storage 68 (2023) 107883

    2023

  39. [39]

    F. Feng, S. Teng, K. Liu, J. Xie, Y . Xie, B. Liu, K. Li, J. Power Sources 455 (2020) 227935

    2020

  40. [40]

    Zheng, X

    Q. Zheng, X. Yin, D. Zhang, J. Energy Storage 73 (2023) 109244

    2023

  41. [41]

    J. Tian, R. Xiong, J. Lu, C. Chen, W. Shen, Energy Storage Mater. 50 (2022) 718–729

    2022

  42. [42]

    Y . Zeng, Y . Li, T. Yang, Batteries 9 (2023)

    2023

  43. [43]

    J. Tian, R. Xiong, W. Shen, J. Lu, Appl. Energy 291 (2021) 116812

    2021

  44. [44]

    J. Hou, J. Xu, C. Lin, D. Jiang, X. Mei, Energy 290 (2024) 130056

    2024

  45. [45]

    Zhang, M

    S. Zhang, M. Liu, R. Guo, J. Tian, Z. Man, W. Shen, IEEE Trans. Transp. Electrific. 12 (2026) 1223–1234

    2026

  46. [46]

    Guo, L.D

    F. Guo, L.D. Couto, Comput. Phys. Commun. 322 (2026) 110075

    2026

  47. [47]

    Biography of the main authors Feng Guo is currently an FWO Senior Postdoctoral Fellow at Hasselt University and VITO (Flemish Institute for Technological Research), Belgium

    Graphical Abstract Electrochemical state propagation and gated recurrent unit (GRU) residual learning are integrated to correct state of charge (SOC) errors of the extended Kalman filter (EKF), improving lithium iron phosphate (LFP) battery estimation under dynamic cycles, temperature variation, and initialization uncertainty. Biography of the main author...

    2015