Energizing sustainable transport: a data-driven approach to estimating state-of-charge in electric vehicle batteries

Abstract

In recent years, the increasing global demand for clean energy and sustainable transportation has highlighted the significant potential of electric vehicles (EVs) in mitigating carbon emissions and reducing reliance on fossil fuels. Lithium-ion batteries, characterized by their high energy density, extended cycle life, and relatively low self-discharge rates, have thus emerged as the primary power source for EVs. Nonetheless, realizing the full potential of EVs in terms of driving range, overall efficiency, and operational safety requires meticulous management of their core energy storage systems. In particular, the accurate estimation of the remaining battery capacity, known as the State of Charge (SOC), is of paramount importance for improving vehicle performance, extending battery lifespan, and ensuring safe operation. Accurate SOC estimation for lithium-ion batteries is a critical factor in optimizing energy management, preventing over-discharge or overcharge situations, and ultimately enhancing the reliability of EVs. Data-driven methods have garnered considerable attention in the research community owing to their strong predictive capabilities and robustness to varying operational conditions. However, many existing data-driven approaches rely heavily on time-domain input signals and may not fully exploit other informative representations, such as the frequency-domain characteristics of battery behavior. Furthermore, the real-world data collected from EVs in daily operation often suffers from issues such as limited sample sizes, high noise levels, and variability in operational conditions (e.g., temperature fluctuations, diverse driving cycles), which can substantially degrade estimation accuracy. These constraints underscore the necessity for advanced techniques that can handle data sparsity and noise while maintaining reliable performance across different usage scenarios. Another key challenge arises from the limited physical interpretability of most data- driven models, which often treat the battery system as a “black box.” While high accuracy is crucial, the absence of physical insights can restrict the applicability of such models beyond the specific conditions under which they are trained. Without incorporating domain knowledge—such as electrochemical dynamics, thermal effects, or aging mechanisms—these purely data-driven methods may fail to generalize to new operating conditions or battery chemistries. To address these limitations, this paper introduces a novel data-driven method based on a transfer dual-stream physical informed network (TDSPIN). In this framework, a dual-stream network architecture integrates both long short-term memory (LSTM) and convolutional neural network (CNN) modules to extract time- and frequency-domain features, thus improving SOC estimation accuracy by leveraging complementary information. In parallel, a physics informed neural network (PINN) is employed to embed prior knowledge of battery dynamics into the model structure, thereby enabling greater interpretability and facilitating transferability across different operating regimes. By combining data-driven feature extraction with physics-based constraints, the proposed TDSPIN offers a robust, transparent, and versatile solution suitable for real-world EV applications. Extensive experiments on three distinct lithium-ion battery datasets commonly used for EVs demonstrate the effectiveness and high precision of the TDSPIN approach, especially when compared with traditional data-driven techniques. The results indicate that incorporating both time- and frequency-domain information, along with physically informed modeling, significantly enhances the reliability and transferability of SOC estimation under various operational conditions. Consequently, the TDSPIN framework represents a promising step toward more accurate and interpretable battery management systems, paving the way for safer and more efficient EV deployments in the future.