Assessment of limited-observability state-of-charge prediction for commercial electric vehicle charging infrastructure using a hybrid LSTM-GRU framework

Abstract

Accurate state-of-charge (SOC) prediction supports intelligent electric vehicle charging, but many high-accuracy methods rely on laboratory measurements or vehicle-side data that are unavailable to commercial charging infrastructure. This study assesses whether protocol-restricted charging-pile records can support multi-step prediction of the BMS-reported SOC trajectory under limited observability. The framework combines charging-session segmentation, data cleaning, secondary feature construction, first-order hold interpolation, correlation-based input selection and group-aware sliding-window training. A hybrid long short-term memory-gated recurrent unit model is compared with LSTM, GRU, LSTM-CNN and Transformer baselines. Group-aware training reduces mean absolute error from 0.36% to 0.30% and mean relative error from 0.64% to 0.58%. Under the same evaluation protocol, LSTM-GRU achieves the lowest errors with 16.00% fewer parameters than LSTM. Across the four seasonal subsets, MAE remains below 0.95% and MRE below 1.40%; under Gaussian noise, MAE and MRE are 0.92% and 1.17%, respectively. These results show that charging-pile records can support SOC trajectory prediction for charging management and infrastructure-side monitoring.