Advanced Multi-Task Learning Model Boosts EV Battery Lifespan Prediction Accuracy
In a significant stride toward safer and more reliable electric vehicles (EVs), researchers have unveiled a novel artificial intelligence-driven framework that dramatically improves the accuracy of lithium-ion battery remaining useful life (RUL) prediction under real-world driving conditions. This breakthrough, detailed in a recent study published in Electric Power Construction, leverages multi-task ensemble learning to address longstanding challenges in battery health monitoring—particularly under the complex mechanical and electrochemical stresses experienced during vehicle operation.
The research team, led by Weiliang Wang from State Grid Jiangsu Electric Power Co., Ltd., in collaboration with scholars from Tianjin University, Tianjin University of Technology, State Grid Tianjin Electric Power Company, Pinggao Group Energy Storage Technology Co., and State Grid Jilin Electric Power Co., has developed a predictive model that achieves unprecedented precision: mean absolute error (MAE) below 1.4%, mean absolute percentage error (MAPE) under 0.06%, and root mean square error (RMSE) less than 1.20%. These metrics represent a meaningful improvement over existing data-driven approaches such as support vector regression (SVR), extreme learning machines (ELM), gradient boosting decision trees (GBDT), XGBoost, and even standard LightGBM models.
At the heart of this innovation lies a sophisticated integration of electrochemical diagnostics and machine learning. Unlike conventional methods that rely solely on capacity fade or voltage curves—metrics that are either difficult to measure in real time or highly sensitive to noise—the new model incorporates deep physical insights into battery degradation mechanisms. Specifically, the team quantified three primary aging modes using incremental capacity–differential voltage (IC-DV) analysis: loss of conductivity (LC), loss of active material (LAM), and loss of lithium ions (LLI). Simultaneously, they extracted four key impedance parameters from electrochemical impedance spectroscopy (EIS): ohmic resistance, charge transfer resistance, solid electrolyte interphase (SEI) resistance, and Warburg diffusion impedance.
These features provide a multidimensional view of internal battery health, capturing both electronic and ionic degradation pathways. However, integrating such rich data into a predictive model introduces complexity and the risk of overfitting—especially when training data is limited, as is often the case in battery aging studies. To overcome this, the researchers employed multi-task learning (MTL), a paradigm in which multiple related prediction tasks are trained jointly, allowing the model to share representations and improve generalization.
In this context, each “task” corresponds to RUL prediction under a specific vibration condition—namely, static (control), X-axis, Y-axis, and Z-axis vibration—simulating the multidirectional mechanical stresses experienced by EV batteries during actual driving. By analyzing feature correlations across these tasks, the MTL framework identifies shared patterns while preserving task-specific nuances. This not only enhances prediction accuracy but also reduces the need for extensive, condition-specific experimental data, thereby lowering testing costs and accelerating model deployment.
The core predictive engine is built upon LightGBM, a highly efficient gradient boosting framework known for its speed and scalability. However, the team significantly enhanced its robustness by replacing the standard loss function with an adaptive robust loss (AR loss). This modified loss function dynamically adjusts its sensitivity to prediction errors based on residual magnitude, effectively down-weighting the influence of outliers and noisy measurements—common issues in real-world battery data. As a result, the model maintains stability and accuracy even when faced with imperfect or incomplete sensor inputs.
To validate their approach, the researchers conducted extensive aging experiments on commercial lithium nickel cobalt manganese oxide (NCM)/graphite pouch cells with a nominal capacity of 2,400 mAh. Batteries were subjected to a realistic driving profile that included idle (0C), constant-speed (1C), acceleration (2C), and deceleration (0.5C) phases, while simultaneously exposed to controlled vibrations along three orthogonal axes using a six-degree-of-freedom shaker table. This setup closely mimics the combined electrochemical and mechanical stresses encountered in real-world EV operation.
The results were compelling. All vibration conditions accelerated battery degradation compared to the static control, with Z-axis vibration causing the most severe aging—evidenced by LAM of 36.94% and LLI of 35.12% at end-of-life. Crucially, the proposed model consistently outperformed all benchmark algorithms across all four conditions, demonstrating not only superior accuracy but also remarkable robustness to varying stress environments.
This advancement carries profound implications for the future of EV battery management systems (BMS). Accurate RUL prediction enables proactive maintenance scheduling, prevents unexpected failures, extends usable battery life, and enhances overall vehicle safety. For fleet operators and automakers, it translates into reduced downtime, lower warranty costs, and improved customer satisfaction. Moreover, by enabling more precise state-of-health estimation, the technology supports second-life applications for retired EV batteries in stationary energy storage—a critical component of the circular economy.
From a scientific standpoint, the study bridges the gap between fundamental electrochemistry and applied machine learning. Rather than treating the battery as a “black box,” the model leverages domain-specific knowledge to guide feature engineering and model architecture. This physics-informed data-driven approach exemplifies the next generation of intelligent diagnostic systems—one that respects the underlying science while harnessing the power of modern AI.
The research also underscores the importance of experimental design in battery AI development. By systematically varying mechanical stress conditions and correlating them with electrochemical signatures, the team created a rich, high-fidelity dataset that captures the true complexity of in-use battery aging. Such datasets are invaluable for training models that generalize beyond laboratory conditions to real-world deployments.
Looking ahead, the authors acknowledge that further work is needed to extend the model to diverse battery chemistries, temperature ranges, and charging protocols. They also emphasize the need for validation under even more dynamic driving scenarios, including urban stop-and-go traffic and high-speed highway cruising. Nevertheless, the current framework provides a robust foundation for next-generation BMS software.
Industry experts note that as EV adoption accelerates globally, the demand for intelligent battery analytics will only intensify. Regulatory bodies in the European Union and United States are already considering mandates for battery health reporting and lifespan transparency. Technologies like the one described in this study could become essential compliance tools, offering both manufacturers and consumers reliable, interpretable insights into battery longevity.
Furthermore, the multi-task learning strategy employed here has broader applicability beyond batteries. It could be adapted to predict the remaining life of other critical components in EVs—such as motors, inverters, and power electronics—under varying operational stresses. This holistic approach to vehicle health management represents a paradigm shift from reactive repairs to predictive resilience.
In conclusion, this research marks a pivotal step toward smarter, safer, and more sustainable electric mobility. By fusing deep electrochemical understanding with cutting-edge machine learning, the team has delivered a practical, high-performance solution to one of the most persistent challenges in battery technology. As EVs continue to transform the transportation landscape, innovations like this will be instrumental in ensuring that their power sources remain reliable, efficient, and trustworthy throughout their operational lives.
Authors: Weiliang Wang¹, Huiqiao Liu²,³, Tianyu Zhang⁴, Peng Ruan⁵, Jin Xu⁶, Qian Xiao³
Affiliations:
¹ State Grid Jiangsu Electric Power Co., Ltd., Nanjing 210024, China
² Zhonghuan Information College, Tianjin University of Technology, Tianjin 300380, China
³ Key Laboratory of Smart Grid of Ministry of Education (Tianjin University), Tianjin 300072, China
⁴ State Grid Tianjin Electric Power Company Economic and Technological Research Institute, Tianjin 300171, China
⁵ Pinggao Group Energy Storage Technology Co., Ltd., Tianjin 300300, China
⁶ Changchun Power Supply Company of State Grid Jilin Electric Power Co., Ltd., Changchun 130021, China
Published in: Electric Power Construction, Vol. 45, No. 11, November 2024
DOI: 10.12204/j.issn.1000-7229.2024.11.003