An equivalent circuit model (ECM) is a highly practical tool for simulating Li-ion battery behavior. There are many relevant studies which compare different ECM variants or suggest algorithms to extract model parameters from the experimental data. However, little attention has been given to the battery tests used for identification of the ECM parameters. Therefore, here the influence of experimental test pulse characteristics on the parameterized ECM accuracy was systematically studied. The test pulse duration was varied in a wide range from 9 s to about 2.5 min. The portion of the relaxation phase data used by the parameter optimization algorithm was also varied in an even wider range. Total 168 ECM parameter sets were obtained. Each parameter set was validated using nine diverse current profiles representing different battery operation conditions, including one based on Urban Dynamometer Driving Schedule (UDDS). The validation results prove that the impact of the test pulse choice on the parameterized ECM accuracy is great to the point that it can overshadow the use of a higher-order Thevenin model. By choosing the optimal parameter set, the simulated voltage root mean square error (RMSE) was reduced to as low as 3.0 mV and 1.2 mV for first- and second-order ECM, respectively, while the second-order model based on arbitrary chosen test pulse on average yields RMSE value above 5 mV.

An equivalent circuit model (ECM) is a highly practical tool for simulating Li-ion battery behavior. There are many relevant studies which compare different ECM variants or suggest algorithms to extract model parameters from the experimental data. However little attention has been given to the battery tests used for identification of the ECM parameters. Therefore, we systematically studied the influence of experimental test pulse characteristics on the parameterized ECM accuracy. Test pulse duration and amplitude were varied along with the portion of the relaxation phase data used by the parameters fitting algorithm. That resulted in 168 parameter sets, each validated using 9 diverse current profiles including one based on a realistic drive cycle. The validation results prove that the impact of the test pulse choice on the parameterized ECM accuracy is great to the point it can overshadow the use of a higher order Thevenin model. By choosing the optimal parameter set the simulated voltage root mean square error was reduced to as low as 1.2 mV.

State of charge (SoC) is the unobserved quantity that a battery management system must reconstruct from terminal voltage, current and temperature. Recent estimators adapt frozen large language models by prompt-tuning the key–value cache of a dense decoder, an interface that does not exist in the linear-attention majority of current hybrid models. We adapt a frozen four-billion-parameter hybrid decoder by matching the injection to each layer type: learned key–value prefixes in the full-attention minority, and a trainable initial recurrent state ($S_0$) in the gated-delta majority. Training only 22.3 million parameters (0.55%), this layer-type-aware adaptation reduces mean absolute error (MAE) by a factor of 2.3 relative to prefixes alone on a preliminary single-cell diagnostic. On the final four-billion-parameter configuration, the same comparison yields a factor of 1.86: pooled mean absolute error decreases from 2.18% for key–value prefixes alone to 1.17% for the full adapter. The nickel-based drive-cycle pool is 1.17% MAE on 18,712 windows. That pool does not include LFP. The LFP value on shared files is a fit of 1.22%. Held-out files of the same cell reach 2.53% in a separate run, and the nickel-based pool of that run is 1.24%. This represents parity with dedicated dense-LLM estimators without claiming superiority over compact conventional observers. On the same drive-cycle split, a gated recurrent unit with 0.2 million parameters reaches 1.18% pooled MAE, and the adapter reaches 1.17%. Instead, the primary contribution is mechanistic. We demonstrate that 98.4% of the residual error is a constant per-window offset on that diagnostic. On this configuration, a constant per-window offset accounts for 0.98 of the residual energy. The estimate therefore behaves as a near-exact Coulomb integration after an initial-condition inverse problem at $t=0$. Compact models trained on the same windows show the same offset-dominated residual, so this pattern is a property of the short-window task as well as of the backbone. This factorization is consistent with the observed failure regimes—concentration of error below 0 °C, and a zero-shot LFP error of 12% MAE on the reference range of about 0–47% state of charge, about ten times the nickel-based pool. Held-out files of an LFP cell partly seen in training reach 2.53% in a separate run. Consequently, we argue that the engineering target for LLM-based battery estimation should transition from expanding backbone capacity to explicitly conditioning the initial state.

Accurate state-of-charge (SoC) estimation from voltage, current, and temperature is required for lithium-ion pack energy management. Learning-based estimators need labeled field traces that are costly on a deployed sensing chain, and data augmentation is used to enlarge those sets. Published studies almost always train one synthesizer for one learner, so it remains unknown which operators help which estimator classes, at what intensity, and after how much real labeled volume. In this paper, a factorial of classical, physics-informed, and generative operators is evaluated on twelve SoC estimators (dose $L=1$–10, labeled volume $n=1$–5; $>5000$ configurations) using dual stand–logger unmanned aerial vehicle (UAV) pack flights as the labeled corpus, with root-mean-square error (RMSE) scored on a locked acceptance trajectory. RMSE change versus the matched unaugmented baseline is class-conditional: sequential and convolutional estimators improve by $+18\%$ to $+32\%$, LightGBM and XGBoost average $\approx -2.4\%$ and $-4.9\%$, and a physics-informed network degrades by $\approx 22\%$. Labeled volume accounts for $\approx 53\%$ of $\log \mathrm{RMSE}$ variance versus $\approx 2\%$ for synthesizer identity. Operators that preserve the empirical training distribution (thermo-voltage coupling and point shuffle) outrank mean-shift and TimeGAN. Dose L is a process parameter: thermo-voltage is optimal at $L=8$ for all eleven classical estimators, and selecting L on the acceptance set inflates reported gain by a factor of $\approx 2.5$ relative to averaging over L. Admissible operators therefore depend on estimator class and on labeled volume n.