Machine Learning-based Health Estimation and Optimal Allocation of Second-life Electric Vehicle Batteries for Grid Applications

Abstract

The rapid proliferation of electric vehicles (EVs) has generated a substantial and growing inventory of retired lithium-ion battery packs, which typically retain 70–80% of their original energy storage capacity upon automotive retirement. Repurposing these second-life batteries (SLBs) for stationary grid applications presents a compelling economic and environmental opportunity; however, their heterogeneous degradation histories introduce fundamental challenges for accurate health estimation and safe deployment. This study presents a comprehensive framework that integrates machine learning (ML) techniques with electrochemical characterization to precisely assess the state of health (SOH) of retired EV batteries and to formulate an optimal allocation strategy for diverse grid services, encompassing frequency regulation, peak shaving, and renewable energy integration. A hybrid model combining Gaussian process regression (GPR) and long short-term memory (LSTM) networks is proposed for SOH estimation, trained on a curated dataset of 1,240 retired battery modules spanning three distinct chemistries. The allocation optimization employs a multi-objective genetic algorithm that jointly minimizes degradation acceleration and system-level costs while maximizing grid service revenue. Simulation results demonstrate that the proposed framework achieves a mean absolute error (MAE) of 1.23% in SOH estimation and increases projected second-life revenue by 34.7% compared to heuristic allocation benchmarks. These findings underscore the practical viability of data-driven SLB management and provide actionable guidelines for grid operators and battery aggregators.