New Algorithm Boosts Battery Parameter Accuracy for Electric Vehicles
In the rapidly evolving world of electric mobility, precision in battery management has become a cornerstone for performance, safety, and longevity. As automakers push the boundaries of range and charging speed, the need for highly accurate real-time monitoring of lithium-ion batteries has never been more critical. At the heart of this challenge lies the battery management system (BMS), which relies heavily on accurate state estimation to ensure optimal operation. A recent breakthrough from researchers at Northeast Electric Power University introduces a novel method that significantly improves the accuracy and adaptability of online battery parameter identification—key to unlocking smarter, more efficient electric vehicles.
The study, led by Duan Shuangming and Zhang Shengli from the Key Laboratory of Modern Power System Simulation and Control & Renewable Energy Technology, Ministry of Education, presents an innovative approach called Adaptive Multi-Layer Recursive Least Squares (AMLRLS). Published in Energy Storage Science and Technology, the research addresses a long-standing limitation in existing parameter identification techniques: their inability to maintain high accuracy when battery parameters change rapidly due to dynamic driving conditions, temperature fluctuations, or varying states of charge (SOC).
Traditional methods such as Recursive Least Squares (RLS) and its variant, Forgetting Factor Recursive Least Squares (FFRLS), have served as foundational tools in system identification. These algorithms are designed to estimate model parameters by minimizing the error between predicted and actual output. However, they face significant challenges when applied to lithium-ion batteries, whose internal characteristics—such as polarization capacitance and internal resistance—are inherently time-varying. Under frequent current changes typical of urban driving cycles, these models often lag behind real parameter shifts, leading to accumulated errors that degrade the overall performance of the BMS.
The core innovation of AMLRLS lies in its layered recursive structure. Instead of relying solely on adjusting a forgetting factor to weigh recent data more heavily, the new algorithm introduces a multi-layer framework where each layer processes the residual voltage error left unexplained by the previous layer. This cascading correction mechanism allows the algorithm to recursively extract and refine parameter estimates, effectively isolating the true parameter dynamics from noise and transient effects.
Imagine a scenario where a sudden acceleration causes a spike in current draw. Conventional RLS might take several seconds to converge to the new polarization capacitance value, during which the model’s voltage prediction drifts from reality. In contrast, AMLRLS immediately captures this discrepancy in the first layer, then uses it as a target signal for the second layer to identify the underlying parameter shift. Subsequent layers further refine the estimate, resulting in faster convergence and higher accuracy.
What sets AMLRLS apart is not just its layered architecture but also its intelligent layer selection mechanism. Rather than computing all layers for every data point—a computationally expensive process—the algorithm dynamically adjusts the number of active layers based on the magnitude of the voltage error. When the error is small, indicating stable operating conditions, only one or two layers are used. But when abrupt changes occur—such as during regenerative braking or rapid charging—the system automatically increases the number of layers to enhance tracking precision.
This adaptive layering strategy strikes a crucial balance between computational efficiency and estimation accuracy. In practical terms, it means that the algorithm can run efficiently on embedded systems within a vehicle’s BMS without overburdening the processor. The researchers validated this claim through extensive simulations and experiments using real-world drive cycles, including the Dynamic Stress Test (DST) and Federal Urban Driving Schedule (FUDS), both known for their aggressive current profiles.
Results from the simulation phase were striking. When tracking a polarization capacitance varying over a 2000-second cycle, AMLRLS reduced parameter error by up to 69% compared to standard RLS and by 46.5% compared to Adaptive Forgetting Factor RLS (AFFRLS). Even under more challenging conditions with a 1000-second variation cycle—simulating faster-changing driving behaviors—the algorithm maintained superior performance, demonstrating its robustness in high-dynamics environments.
But simulations alone are not enough to prove real-world applicability. The team conducted experimental validation using lithium-ion cells (INR 18650-20R) sourced from the University of Maryland’s battery dataset, a widely recognized benchmark in the field. Tests were performed under various temperatures (0 °C, 25 °C, and 45 °C), initial SOC levels (50% and 80%), and different current profiles. Across all scenarios, AMLRLS consistently outperformed both RLS and AFFRLS in terms of voltage prediction accuracy.
Under the DST cycle at 25 °C, the root mean square error (RMSE) of the terminal voltage was reduced by 43.9% compared to other methods, while the average absolute error (MAE) dropped by 32.1%. These improvements are not merely statistical—they translate directly into better state-of-charge estimation, enhanced fault detection, and improved thermal management. For automakers striving to meet stringent safety and warranty standards, such gains represent a tangible step forward in battery intelligence.
Equally important is the algorithm’s computational efficiency. Without optimization, multi-layer structures can become prohibitively slow, especially when processing thousands of data points per second. However, the integration of the layer selector proved transformative. In DST tests, the computation time was reduced by 37.4% compared to a fixed-maximum-layer implementation. Under the even more demanding FUDS cycle, the savings reached 28.6%. This level of optimization ensures that AMLRLS remains viable for deployment in production-grade BMS platforms, where processing power and energy consumption are tightly constrained.
One of the most compelling aspects of the research is its validation across diverse operating conditions. Battery behavior is notoriously sensitive to temperature, with low temperatures increasing internal resistance and reducing available capacity. At 0 °C, for instance, the voltage errors for RLS and AFFRLS grew significantly, reflecting their struggle to adapt to cold-weather dynamics. AMLRLS, however, maintained tight error bounds, showcasing its resilience under thermal stress. Similarly, at elevated temperatures (45 °C), the algorithm continued to deliver precise parameter estimates, underscoring its versatility across the full operational envelope of modern EVs.
Initial SOC also plays a critical role in battery modeling accuracy. The study found that higher initial SOC levels generally resulted in lower voltage errors, likely due to the more linear region of the SOC-OCV (open-circuit voltage) curve. Nevertheless, regardless of starting conditions, AMLRLS consistently provided the most accurate results, reinforcing its reliability in real-world use cases where drivers begin trips with varying charge levels.
The implications of this work extend beyond immediate performance gains. Accurate parameter identification forms the foundation for advanced battery diagnostics, including state-of-health (SOH) estimation, internal short-circuit detection, and degradation modeling. By providing a more faithful representation of the battery’s internal state, AMLRLS enables predictive maintenance strategies, reduces the risk of unexpected failures, and supports longer battery warranties—key selling points in the competitive EV market.
Moreover, the algorithm’s design philosophy reflects a broader trend in automotive engineering: moving from static, rule-based systems to adaptive, learning-capable architectures. While machine learning models like LSTM networks have shown promise in battery modeling, they often require extensive training data and lack interpretability. AMLRLS, in contrast, combines the transparency of physics-based models with the adaptability of recursive estimation, offering a balanced solution that is both explainable and effective.
From a systems integration perspective, the modular nature of AMLRLS makes it highly compatible with existing BMS software stacks. It can be implemented as a standalone parameter estimator feeding into a Kalman filter or other state observers, enhancing their input quality without requiring a complete overhaul of the control architecture. This plug-and-play capability lowers the barrier to adoption for OEMs and Tier 1 suppliers alike.
Looking ahead, the researchers acknowledge that there is still room for improvement. One area of future work involves incorporating historical identification data into the current estimation process, potentially enabling long-term trend analysis and drift compensation. Additionally, extending the method to multi-cell battery packs—where cell-to-cell variations add another layer of complexity—could unlock even greater benefits for pack-level management.
The automotive industry is undergoing a fundamental transformation, driven by electrification, autonomy, and connectivity. In this new era, the battery is no longer just an energy source—it is a smart, data-rich component that must be continuously monitored and optimized. Algorithms like AMLRLS represent the next generation of battery intelligence, turning raw sensor data into actionable insights that improve vehicle performance, safety, and user experience.
As governments around the world accelerate their transition to zero-emission transportation, the demand for smarter, more reliable batteries will only grow. Innovations like those developed by Duan Shuangming and Zhang Shengli are not just academic achievements—they are essential building blocks for the sustainable mobility systems of tomorrow.
The success of AMLRLS also highlights the importance of interdisciplinary collaboration in advancing battery technology. Drawing from signal processing, control theory, and electrochemistry, the research exemplifies how cross-domain expertise can yield practical solutions to complex engineering problems. It serves as a model for how academic institutions can contribute meaningfully to industrial innovation, particularly in strategic sectors like clean energy and transportation.
For consumers, the impact may be subtle but profound. They may never see the algorithm running inside their car’s computer, but they will feel the difference in smoother acceleration, more accurate range predictions, and longer-lasting batteries. In an age where trust in new technologies is paramount, such invisible improvements build confidence in electric vehicles and pave the way for wider adoption.
In conclusion, the development of the Adaptive Multi-Layer Recursive Least Squares method marks a significant milestone in battery management science. By rethinking how parameters are updated in real time, the researchers have delivered a solution that is not only more accurate but also more efficient and adaptable than existing approaches. As the automotive world continues its electrified journey, innovations like AMLRLS will play a vital role in ensuring that the power beneath the hood is managed with unprecedented precision.
Duan Shuangming, Zhang Shengli. Lithium-ion battery parameter identification based on adaptive multilayer RLS. Energy Storage Science and Technology. doi: 10.19799/j.cnki.2095-4239.2023.0605