Enhancing Battery Health Prediction during Fast Charging Using Digital Twins Based on Physical Insights

Traditional machine learning methods for state-of-health (SOH) status prediction struggle with physical interpretability. Their performances fail when training and testing datasets are randomly split, making them less effective for unpredicted conditions and events. This work proposes a novel, model-free, physics-informed machine learning method that learns sensitive underlying physical and chemical parameters from randomly selected charging segments. Combined with battery aging indicators, it estimates SOH and predicts remaining useful life (RUL) using only routine maintenance labels. This approach offers a straight-forwardly implementable solution for battery management system (BMS) at charging ports, enabling smart charging protocol adjustments. This study uses fast-charging datasets from previous work 1,2 based on commercially available 18650 cylindrical NMC Samsung cells. Charging was performed using CCCV and multistep protocols at 0.6C, 1C, 1.3C, and 2C, with CV held until the current dropped to 0.01A (C/20). Discharging followed the World Harmonized Light Vehicles Test Cycle (WLTC) profile with voltage limits of 2.5V (~100% DOD) and 3.6V (~70% DOD). Reference performance tests (RPTs) were conducted every 100 equivalent full cycles (EFCs) until a 20% capacity fade was reached. Cell lifetimes varied from 187 to 926 EFCs. Figure 1 shows the validation results of the proposed method with and without the six physical parameters (P*). Figure 1(a) and 1(b) display the estimation errors for SOH/SOH* and prediction errors for RUL/RUL*, using features extracted from the [3.65V, 4.1V] range. A 0.7:0.3 random train-test split was applied to data from cells 1–10. As shown in Figure 1(a), incorporating physical parameters results in SOH errors mostly under 0.02 and RUL errors below 100 EFCs. In contrast, without these parameters (Figure 1(b)), SOH* errors increase to below 0.04, and RUL* predictions show larger deviations, mostly under 200 EFCs. To evaluate the robustness of P*-enhanced battery health prediction, Figures 1(c) and 1(d) present results from 100 Monte Carlo runs, each using a 70% training cell selection. The inclusion of P* consistently yields lower MAE and higher R² values, indicating improved accuracy and model fit. With P*, the mean absolute error (MAE) for SOH estimation is reduced to approximately 0.02 (Figure 1(c-1)), with R² values consistently near 0.8 (Figure 1(d-1)), indicating high predictive accuracy and low variance. For RUL prediction, the model achieves an MAE around 70 EFCs (Figure 1(c-2)) and R² exceeding 0.85 (Figure 1(d-2)). In contrast, excluding P* leads to higher MAE (up to 0.06 for SOH and >150 EFCs for RUL) and reduced R² scores, with some SOH predictions dropping below 0, suggesting poor generalization. These results confirm that the inclusion of physically meaningful parameters enhances both the precision and robustness of the model under diverse conditions. This physics-informed method can be readily integrated into existing BMS at charging stations, allowing for data exchange between EVs and the charging infrastructure. Figure 1. Performance evaluations with and without physical features (P*). (a, b) Parity plots from a random forest model using a 0.7:0.3 train-test split: (a) using P* and (b) exclude P*. (c, d) Results of 100 Monte Carlo evaluations using different random seeds to assess the performance of the proposed physics-informed method. Acknowledgement This study is funded by Nordic Energy Research (Best4Grid), the Swedish Electromobility Center and the Swedish Energy Agency project EV drivers in the driver’s seat . Reference [1]Guo, W.; Sun, Z.; Guo, J.; Li, Y.; Vilsen, S. B.; Stroe, D. I. Digital Twin-Assisted Degradation Diagnosis and Quantification of NMC Battery Aging Effects During Fast Charging. Adv Energy Mater 2024, 14 (36). https://doi.org/10.1002/aenm.202401644. [2]Li, Y.; Guo, W.; Stroe, D. I.; Zhao, H.; Kjær Kristensen, P.; Rosgaard Jensen, L.; Pedersen, K.; Gurevich, L. Evolution of Aging Mechanisms and Performance Degradation of Lithium-Ion Battery from Moderate to Severe Capacity Loss Scenarios. Chemical Engineering Journal 2024, 498. https://doi.org/10.1016/j.cej.2024.155588. Figure 1