Personalized Federated Lithium-Ion Battery Capacity Prediction via Cluster and Fusion Modules

Federated learning (FL) is a promising solution for addressing information security sharing challenges in the Internet of Vehicles (IoV). It enables individual-level capacity prediction of lithium-ion batteries in electric vehicles (EVs). However, existing FL algorithms primarily focus on training a single global sharing model, neglecting the predictive capabilities of individual participants, which poses challenges for batteries with distinct capacity degradation trends. To address this limitation, we introduce battery-personalized FL (BT-PFL), a novel FL framework that provides personalized capacity prediction models based on the local data distribution of each battery. Our approach involves constructing cluster and fusion modules, creating a personalized learning space, and leveraging valuable domain knowledge. On the one hand, we proactively identify the capacity degradation distribution of different batteries, ensuring data privacy during FL within clusters. Each cluster retains domain-specific style information, enabling collaborative training and shared model parameters among participants. On the other hand, we introduce a knowledge distillation (KD) algorithm that facilitates knowledge transfer between clusters by constructing a teacher model incorporating multidomain knowledge. The experimental results on both single-domain and multidomain show that our approach not only significantly enhances individual battery prediction accuracy but also outperforms other methods across various datasets, network structures, and noise levels.

I. Introduction

As the global energy crisis continues to escalate, the widespread adoption of clean energy has become an inevitable trend. Lithium-ion batteries have emerged as indispensable sources of energy in various applications, including electric vehicles (EVs), due to their long lifespan, high energy density, and environmentally friendly characteristics [1]. However, being the core energy source in EVs, lithium-ion batteries inevitably experience performance degradation during usage, and this degradation is often challenging to directly measure. While accurately predicting battery capacity poses challenges, it is of paramount importance for the reliable utilization of lithium-ion batteries. Precise capacity forecasts can accurately reflect the driving range of interest to users and enable the accurate calculation of a vehicle’s maximum energy storage capacity [2]. Furthermore, an accurate prediction of battery capacity aids in detecting battery failure conditions and proactively prevents failures through timely maintenance measures [3]. Therefore, the accurate prediction of lithium-ion battery capacity is an essential research endeavor.