Physically consistent battery charging trajectory generation under mixed protocols using flow-matched diffusion transformers

Li, Xiaotian

doi:10.1016/j.eswa.2026.133813

Physically consistent battery charging trajectory generation under mixed protocols using flow-matched diffusion transformers

2026-08-01

0

PRE

AI

DOI：10.1016/j.eswa.2026.133813

Original

Original request for help

Save

Abstract

En 中文

Accurate full-lifecycle management and optimal charging control in battery management systems require largescale time-series datasets covering diverse charging protocols. However, it is infeasible to realize next-generation battery intelligence entirely through costly and time-consuming physical testing experiments. This study proposes a simulation-free conditional generative modeling framework for synthesizing lithium-ion battery charging time-series under mixed charging protocols. Based on charging protocols and battery aging states, the proposed approach integrates flow matching with a diffusion transformer architecture formulated as an ordinary differential equation (ODE)-based generative process. It introduces a dual-stream decoupled self-attention mechanism to separately capture localized transient responses and long-range degradation-dependent dependencies. By continuously incorporating the state of health (SOH) as a conditioning prompt to capture degradation-induced variations in constant-voltage transition behavior, the model enables the generation of protocol-consistent trajectories. The model is validated on a dataset containing 70 batteries across 29 mixed charging protocols, achieving average root-mean-square errors of 0.184 C, 0.010 V, and 0.002 Ah for current, voltage, and capacity, respectively, while maintaining stable generation performance even for severely degraded batteries (SOH < 85%). Furthermore, incremental capacity peak analysis of the generated charging curves highlights the strong physical consistency of the synthesized data. Overall, this framework delivers a scalable battery data generation solution aimed at overcoming the data acquisition bottlenecks inherent in physical experiments, thereby accelerating the implementation of next-generation battery management systems.

Keywords:

Lithium-ion batteries

Mixed charging protocols

Time-series data generation

State of health

Flow matching

Physically consistent battery charging trajectory generation under mixed protocols using flow-matched diffusion transformers

Abstract

Journal

Organization

Cited Papers