Enhanced passive balancing approach for battery management systems using machine learning

Abstract

Passive cell-balancing remains attractive for cost-sensitive battery management systems (BMS), but its fixed-threshold control wastes energy and accelerates component ageing. Many cost-sensitive applications rely on passive shunt-resistor balancing for lithium-ion battery packs because it is simple, reliable and low-cost, which typically uses open circuit voltage (OCV) as a proxy of state of charge (SOC), and this does not provide enough accuracy, mainly in lithium-ion cells where the voltage remains nearly the same for a wide range of SOC levels. In these aspects, despite large differences in SOC, the voltage variations may be too small for the BMS to measure accurately. For those reasons, the cells may remain persistently imbalanced during the balancing process, which can result in them not being fully used, a higher risk of overcharging and faster wear from heat produced. This study aims to investigate recent advancements in passive balancing and machine learning for SOC estimation, develop an enhanced passive balancing BMS architecture that integrates machine learning, and evaluate the proposed system against conventional OCV-based approaches in terms of accuracy, efficiency, and performance. An enhanced passive balancing architecture has been developed that couples a low-cost switched-shunt network with a long-short-term-memory (LSTM) SOC estimator trained on multi-temperature public datasets. The data-driven predictor supplies real-time SOC values to a hysteresis controller, enabling balancing decisions that are informed by cell dynamics rather than voltage alone. A three-cell lithium-ion in 3S1P arrangement was modelled in MATLAB/Simulink and validated against a hardware prototype configured for conventional OCV-based estimation that was used for the balancing process. Performance metrics such as balancing time, switching frequency, power dissipation and thermal rise were recorded for both strategies, and such findings were discovered in this study. The ML-assisted approach reduced balancing time to 80% SOC by 29% (2.88 hours vs. 4.05 hours), decreased average switching frequency by 97% (10.2 mHz vs. 333.8 mHz), and lowered total power dissipation by 81% (0.422 W vs. 2.22 W). The LSTM-based SOC estimator maintained high predictive accuracy across a temperature range of −10°C to 25°C with a root mean square error (RMSE) of 0.025, ensuring reliable SOC guidance under typical operating conditions. Thermal analysis revealed minimal temperature rise, with cell temperatures remaining within 0.3–0.5°C of ambient during the final charging phase. These results show that machine learning enables passive balancing circuits to match the efficiency of much more expensive active balancers, while maintaining simplicity, low cost, and safety of the battery cell. The proposed framework extends battery life, reduces thermal management costs, lowers total power loss, and increases efficiency.