Enabling Delayed-Full Charging Through Transformer-Based Real-Time-to-Departure Modeling for EV Battery Longevity

TL;DR

The paper tackles delaying full EV charging to protect battery longevity by accurately predicting daily departure times in real time. It reframes departure prediction as a time-to-event problem and introduces a Transformer-based TTD model that updates survival probabilities across discretized time grids using streaming contextual cues from smartphone sensing. The method, aided by Gaussian-smoothed supervision and carefully designed regularization, outperforms historical and context-aware baselines on a real-world dataset of 93 participants, achieving a MAE of 2.20 hours and benefiting from personalization. The work provides a practical benchmark for real-time departure prediction in smart EV charging, with implications for battery health, user experience, and scalable deployment.

Abstract

Electric vehicles (EVs) are key to sustainable mobility, yet their lithium-ion batteries (LIBs) degrade more rapidly under prolonged high states of charge (SOC). This can be mitigated by delaying full charging \ours until just before departure, which requires accurate prediction of user departure times. In this work, we propose Transformer-based real-time-to-event (TTE) model for accurate EV departure prediction. Our approach represents each day as a TTE sequence by discretizing time into grid-based tokens. Unlike previous methods primarily dependent on temporal dependency from historical patterns, our method leverages streaming contextual information to predict departures. Evaluation on a real-world study involving 93 users and passive smartphone data demonstrates that our method effectively captures irregular departure patterns within individual routines, outperforming baseline models. These results highlight the potential for practical deployment of the \ours algorithm and its contribution to sustainable transportation systems.

Figures (12)

• Figure 1: Virtual operation of DFC. Early predictions cause prolonged high SOC, reducing battery health benefits. Late predictions risk insufficient charge and range anxiety. Accurate prediction of departure at least 30 minutes in advance ensures full charging for effective DFC operation, extending battery life while maintaining user confidence.
• Figure 2: Real-time survival modeling with streaming contextual inputs. During training, the model estimates survival probabilities up to the observed departure event for uncensored sequences. At inference, survival estimates are updated incrementally as new tokens arrive, triggering a departure once the probability falls below a threshold $p$.
• Figure 3: A diagram of the TTD model architecture. Contextual features and day-of-week embeddings are fused with absolute time features through the alpha-fusion mechanism, with absolute time scaled by a learnable parameter to control its contribution relative to contextual signal. The fused embeddings, combined with positional encoding, are processed by a multi-layer Transformer encoder to capture long-range temporal dependencies. The output layer predicts per-interval survival probabilities $\hat{S}(t \mid \mathbf{X})$ across all intervals in parallel during training and supports real-time inference by updating predictions as new tokens arrive.
• Figure 4: Kernel Density Estimation (KDE) plots of departure times across models. The proposed model shows the closest alignment with the ground-truth distribution compared to the best regressor and classifier baselines.
• Figure 5: The impact of hyperparameters: event weight $\omega_e$, weekend weight $\omega_w$, and survival probability threshold $p$ on the departure time prediction accuracy.