New transfer-learning model could improve real-world EV charging duration prediction

Researchers have proposed a SENet-CNN-Transformer model for predicting electric vehicle charging duration, aiming to improve estimates of how long it will take a vehicle to charge from its current state of charge to a target state of charge. The method combines data enhancement, channel attention, convolutional neural networks, Transformer modeling, and transfer learning to address real-world data scarcity and nonlinear battery behavior.

Charging duration is a practical issue for electric vehicle users and charging-network operators. Drivers want reliable estimates of how long a charging session will take, while operators need accurate predictions to manage charger availability, station scheduling, and service quality. If charging time is underestimated, users may face delays and poor planning. If it is overestimated, charging resources may be used less efficiently.

The challenge is that real-world charging duration is not controlled by a single simple factor. Battery state, charging behavior, system configuration, and operating conditions can interact in nonlinear ways. According to the article, two major technical barriers are the scarcity of real-world charging data and the deeply nested nonlinear features inherent in battery systems. These issues make it difficult for simple models or single-network approaches to generalize reliably.

The new study addresses these barriers by proposing a prediction method based on a Squeeze-and-Excitation Network, or SENet, combined with a convolutional neural network and Transformer architecture. The SENet module uses a channel attention mechanism to dynamically adjust feature weights. In practical terms, this helps the model focus on the more informative parts of the input data instead of treating all feature channels as equally important.

The CNN-Transformer component is designed to combine complementary modeling strengths. The convolutional neural network, or CNN, supports local feature extraction, helping capture short-range patterns in charging-related data. The Transformer contributes global modeling capacity, helping capture longer-range dependencies and interactions that may be missed by a CNN alone. The article frames this combined architecture as a way to overcome the limitations of single networks in feature extraction and dependency modeling.

Data enhancement is another part of the framework. The researchers use data-enhancement techniques to improve the richness of the dataset and reduce the impact of overfitting. This is important because real-vehicle charging data can be limited, uneven, or difficult to collect at scale. A model trained on too little data may appear accurate in a narrow setting but fail when exposed to broader real-world charging behavior.

The study also adopts transfer learning to connect laboratory data and real-vehicle data. According to the article, the proposed model is first pre-trained on laboratory data and then fine-tuned on real-world data. This approach allows the model to learn useful battery-related patterns from more controlled laboratory conditions before adapting to practical vehicle data. It also significantly reduces training time, which can matter when models need to be updated or deployed across different charging contexts.

Experiments on real-world battery data suggest that the SENet-CNN-Transformer model outperforms several comparison models. The authors report mean absolute error reductions of 56%, 65%, and 75% compared with CNN-Transformer, Transformer, and Long Short-Term Memory models, respectively. With the transfer learning technique, training time was reduced by 4.5 times without sacrificing prediction accuracy. These results indicate that the proposed architecture improves accuracy while also making model training more efficient.

Further validation will still be needed across different vehicles, battery chemistries, charging infrastructures, climates, and user behaviors. Even so, the study offers a strong indication that combining attention-based feature weighting, hybrid neural-network modeling, data enhancement, and transfer learning could make EV charging duration prediction more accurate and practical. As charging networks expand, better prediction tools may help improve user experience, station management, and the integration of electric vehicles into intelligent transportation systems.

Reference
Author:
Yuan Chen ^a, Siyuan Zhang ^a, Xiaopeng Tang ^b, Heng Li ^c

Title of original paper:
Real-world vehicle charging duration prediction based on transfer learning with SENet-CNN-Transformer Model

Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725001227

Journal:
Green Energy and Intelligent Transportation

DOI:
10.1016/j.geits.2025.100372

Affiliations:

^a School of Artificial Intelligence, Anhui University, Hefei 230039, China

^b Science Unit, Lingnan University, 8 Castle Peak Road, Tuen Mun 999077, Hong Kong

^c School of Electronic Information, Central South University, Changsha 410083, China