Spatial-temporal feature extraction and residual correction for robust battery state-of-charge estimation under real-world operating conditions

Accurate state-of-charge (SOC) estimation remains critical for electric vehicles but is severely challenged by highly nonlinear battery dynamics, environmental noise, and cumulative biases in onboard battery management systems (BMS). This study proposes a robust SOC estimation framework driven by physics-constrained offline reconstruction and heterogeneous feature decoupling. To overcome the absence of absolute ground truth in real-world data, a tailored label reconstruction mechanism is developed. By extracting physical anchors from rest voltages and full-charge boundaries, onboard raw data are physically refined to construct a high-confidence reference trajectory. Building upon this, a heterogeneous Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Light Gradient Boosting Machine (LightGBM) prediction framework is introduced. A deep CNN-LSTM network captures spatial-temporal evolution trends within a continuous parameter space, whereas a LightGBM decision tree executes high-frequency nonlinear residual compensation within a discrete decision space. This cross-optimization architectural design alleviates the gradient conflicts and fitting degradation commonly observed in conventional single-model networks. Additionally, a complementary hierarchical feature reduction strategy is implemented to minimize computational overhead without sacrificing critical physical characteristics. Validation on real-world datasets shows robust performance, with an MAE of 0.62 and an RMSE of 0.89. Through the “offline refinement to online mapping” design, the proposed framework approximates the offline reconstructed SOC trajectory using current and historical measurements under the evaluated real-world operating conditions.