State of charge estimation for LiFePO4 power battery based on multi-dimensional operating features

LiFePO 4 battery exhibit flat regions in their OCV-SOC curves, posing challenges for accurate SOC estimation using traditional methods. Existing data-driven approaches typically rely on cell-level current, voltage, and ambient temperature, which are insufficient to capture complex operating conditions. Moreover, models trained on experimental data often suffer from distribution drift in real-world deployment. To address these limitations, this study develops a complete SOC estimation pipeline. SOC label is derived from full-charge reverse coulomb counting. At the pack level, ten-dimensional features are introduced, including total voltage, current, ambient temperature, maximum/minimum/average cell voltage and temperature, average current, driving distance, and time interval. A Multi-Scale Deep Neural Network based on Multi-dimensional Operating Conditions(MS-MDOC) is proposed based on these features. It employs hierarchical modeling: a local difference enhancement module highlights subtle variations in flat regions; a GRU-based temporal compression module captures long-term SOC trends; and a delayed fusion module incorporates global features to improve global dependency learning. Experiments on the Changan real vehicle dataset show that MS-MDOC reduces MAE, ME, and RMSE by 29.9%, 21.9%, and 33.8%, respectively, and improves R 2 by 1.5%, compared to existing methods. The model also demonstrates strong generalization under temperature variation on the public A123 dataset. After Int8 quantization, the model size is reduced to 1.4 MB. It supports asynchronous inference and enables SOC correction when the vehicle is stationary. MS-MDOC achieves low latency and resource efficiency, making it suitable for embedded system deployment. • A hierarchical-input neural network is proposed for SOC estimation of LiFePO 4 batteries. • Multi-dimensional operating features are introduced to improve SOC estimation accuracy. • Trend features in SOC-OCV plateau regions are adaptively enhanced and preserved. • A complete data processing pipeline is designed for real-world electric vehicle datasets.