Repurposing of retired electric vehicle lithium-ion batteries through state of charge estimation with deep learning techniques

The transition from Internal Combustion Engine (ICE) vehicles to Electric Vehicles (EVs) in the UK is driven by significant regulatory and financial incentives that are leading to an increase in EV registrations. As EV adoption rises, managing the end-of-life of EV batteries becomes necessary.
Repurposing retired EV batteries (REVBs) for second-life applications offers potential circular economy opportunities. However, the technical and economic viability of such repurposing is often challenged by the intensive nature of certain applications, which can shorten battery life and impact economic feasibility. Accurate State of Charge (SOC) estimation is critical in this context, as it directly influences battery performance, safety, and lifespan. Reliable SOC estimation prevents overcharging and over-discharging, which are key to extending battery life and ensuring the reliability of Battery Management Systems (BMS). Current SOC estimation methods, including Coulomb counting, OCV-SOC curves, and various model-based and data-driven approaches, have limitations, particularly under diverse operational conditions.

This thesis presents a novel framework to improve SOC estimation with focus on second life EV batteries as an enhancement over traditional standalone models. The Cluster-Based Learning Model (CBLM) framework, integrates K-Means clustering with Long Short-Term Memory (LSTM) networks with a centroid proximity mechanism which allows the estimation model to dynamically adapts to diverse operational conditions by segmenting battery data into meaningful clusters, enabling more precise and context aware SOC estimation. Comparative evaluations demonstrated that the CBLM achieved significant reductions in estimation errors, up to 62% in RMSE and 69% in MAE outperforming Standard LSTM (S. LSTM) model.

Additionally, the model's robustness was thoroughly evaluated under real-world conditions, including scenarios with varying ambient temperatures and noisy sensor measurements. Even under extreme sensor degradation due to wear, the CBLM maintained reliable performance, demonstrating resilience to noise and ensuring accurate SOC estimation. To address the computational complexity of CBLM with