MapVC: Map-based deep learning for real-time current prediction in eco-driving electric vehicles

Abstract

The operating current prediction of power batteries is crucial for ensuring the working performance of Electric Vehicle (EV). However, complex real-world eco-driving scenarios—particularly the common engagement of regenerative braking systems (RBS) that produce negative current values—have introduced strong randomness into power system data. To overcome the limitations of conventional data-driven models in capturing such complexity, we propose the MapVC framework. First, a map-based encoder is introduced, which deduces the operation of the RBS via estimating the vehicle's motion state, greatly reinforcing prediction performance of data from complex real-world driving conditions. Additionally, a decoder leveraging multi-head self-attention is employed to extract multi-scale temporal features, enabling comprehensive modeling of intrinsic battery state changes. Moreover, a bidirectional gated recurrent network is integrated, which manages to address long-term dependency loss and exploit both past and future information for robust sequential modeling. To further mitigate overfitting problem caused by high-dimensional parameters, we introduce the Improved Hippopotamus Optimization (IHO) algorithm for efficient network tuning. Trained on real-world data from electric buses with RBS in Wuhan, China, our model achieves an MSE of 0.0709, MAE of 0.1859 and MAPE of 1.81 %, representing up to 93 % reduction in MSE and a 5.6-fold improvement in MAPE over prior work while maintaining outstanding computational efficiency. It outperforms its precursor in predicting key parameters of operating data and provides significant guidance for the application of geographic information to vehicle operating condition prediction.