Open Access

ARTICLE

Hybrid Temporal Convolutional Network-Transformer Model Optimized by Particle Swarm Optimization for State of Charge Estimation of Lithium-Ion Batteries

Xincheng Han¹, Hongyan Ma^1,2,3,*, Shuo Meng¹, Chengzhi Ren¹

1 School of Intelligence Science and Technology, Beijing University of Civil Engineering and Architecture, Beijing, 100044, China
2 Institute of Distributed Energy Storage Safety Big Data, Beijing, 100044, China
3 Beijing Key Laboratory of Super Intelligent Technology for Urban Architecture, Beijing University of Civil Engineering and Architecture, Beijing, 100044, China

* Corresponding Author: Hongyan Ma. Email:

Energy Engineering 2026, 123(4), 23 https://doi.org/10.32604/ee.2025.072906

Received 06 September 2025; Accepted 12 November 2025; Issue published 27 March 2026

Abstract

Lithium-ion (Li-ion) batteries stand as the dominant energy storage solution, despite their widespread adoption, precisely determining the state of charge (SOC) continues to pose significant difficulties, with direct implications for battery safety, operational reliability, and overall performance. Current SOC estimation techniques often demonstrate limited accuracy, particularly when confronted with complex operational scenarios and wide temperature variations, where their generalization capacity and dynamic adaptation prove insufficient. To address these shortcomings, this work presents a PSO-TCN-Transformer network model for SOC estimation. This research uses the Particle Swarm Optimization (PSO) method to automatically configure the architectural parameters of the Temporal Convolutional Network (TCN) and Transformer components. This automated optimization enhances the model’s ability to represent the dynamically evolving nature of SOC. Additionally, this integrated framework significantly increases the model’s capacity to capture SOC dynamics in complex operational scenarios. During training and evaluation using a comprehensive dataset that covers complex operating conditions and a broad temperature spanning from −20°C to 40°C, the proposed model achieves a root mean square error (RMSE) of less than 0.6%, a maximum absolute error (MAXE) below 4.0%, and a coefficient of determination (R²) of 99.99%. Additional comparative experiments on data from an energy storage company further verify the model’s superior performance, with an RMSE of 1.18% and an MAXE of 1.95%. The implications of this work extend to the development of optimization strategies and hybrid architectures, providing insights that can be adapted for state estimation across a range of complex dynamic systems.

---

Keywords

---