Research
Enabling Delayed-Full Charging Through Transformer-Based Real-Time-to-Departure Modeling for EV Battery Longevity
Enabling Delayed-Full Charging Through Transformer-Based Real-Time-to-Departure Modeling for EV Battery Longevity Overview Research area: Machine learning for electric vehicle (EV) smart charging, spe
- arXiv
- 2512.07723
- Published
- 2025-12-08
- Authors
- Yonggeon Lee, Jibin Hwang, Alfred Malengo Kondoro, Juhyun Song, Youngtae Noh
AI summary
Enabling Delayed-Full Charging Through Transformer-Based Real-Time-to-Departure Modeling for EV Battery LongevityOverview
- Research area: Machine learning for electric vehicle (EV) smart charging, specifically time-to-event (survival) modeling of driver departure times from smartphone passive sensing data.
- Technical level: Intermediate. The paper assumes familiarity with Transformer encoders, discrete-time survival analysis, and standard regression/classification baselines, though the problem framing is explained concretely.
- Scope: The paper formulates EV departure-time prediction as a time-to-departure (TTD) survival problem, proposes a Transformer-based model that updates survival probabilities in real time using contextual sensing features, and evaluates it on a field study of 93 users against historical-statistics and context-aware baselines.
What This Paper Is About
EV lithium-ion batteries degrade faster when they sit at a high state of charge (SOC) for long periods, which happens when a car is charged to full hours before it is actually driven. The Delayed-Full Charging (DFC) strategy addresses this by waiting until about 30 minutes before departure to fast-charge to 100% SOC, but it only works if the departure time can be predicted accurately in advance. This paper builds a model that predicts when a user will depart, using passively collected smartphone signals rather than only historical departure statistics.
Key Contributions
- Problem reformulation. The authors introduce EV departure-time prediction as a time-to-event (TTE) prediction task rather than a classification task, arguing this reduces label imbalance and supports probabilistic, real-time inference for DFC.
- Transformer-based TTD framework. They design a model that performs token-wise streaming inference over discretized 5-minute intervals, combining an ordinal Gaussian-smoothed loss with regularization strategies (dropout-time, time-scale, and alpha-fusion) for robustness under dynamic contexts.
- Real-world validation. They evaluate on a field study dataset of 93 users and 3,906 daily sequences, reporting the lowest mean absolute error (MAE) among all baselines tested.
- Dataset and code release. They state that the anonymized dataset and processing pipeline will be released upon publication, and the code is available at https://github.com/LYGLeo/3TD-AISI-26.
Main Findings
- Lowest departure-prediction error. The proposed model achieves 2.20 hours MAE across all days (2.26 on weekdays, 2.07 on weekends), versus 2.57 hours for the best historical-statistics baseline (SVR) and 2.59 hours for the best context-aware classifier (iTransformer).
- Reported relative improvements. The paper states gains of 14.4%/15.1% overall, 10.3%/11.7% on weekdays, and 21.3%/21.9% on weekends relative to the best historical baseline and the best context-aware classifier, respectively.
- Contextual features matter most. In the ablation study, removing contextual features (w/o context) causes the largest degradation, raising MAE to 4.47 ± 0.18 hours. Removing positional encoding (w/o PE) raises MAE to 4.25 ± 1.74 hours, and removing absolute time features (w/o time) to 3.01 ± 0.67 hours.
- Smaller but real contributions from other components. Removing day-of-week features gives 2.64 ± 0.60 hours; removing alpha-fusion gives 2.55 ± 0.42 hours; removing time-scale and Gaussian-smoothed supervision together gives 2.36 ± 0.04 hours, compared with 2.20 ± 0.13 for the full model.
- Distributional tracking. Kernel density estimation plots show the model's predicted departure distribution closest to the ground-truth distribution, while historical baselines concentrate around global averages. A slight bias toward earlier predictions is attributed to the decision threshold p = 0.1, which the authors frame as a beneficial buffer for DFC.
- Surprising weekend result. Several configurations show lower MAE on weekends than weekdays, contradicting the expectation of greater weekend irregularity; the authors attribute this plausibly to the 2021–2022 collection period coinciding with COVID-19 and post-pandemic transitions affecting weekday routines.
- Hyperparameter behavior. Medium event weights (1.0–1.5) give the best MAE; the best results come from event weight 1.5 combined with weekend weight 1.5 and a small survival threshold p. Small weekend weights consistently harm weekend predictions.
- Personalization helps modestly. Fine-tuning only the last Transformer layer and the output layer per user improves overall MAE from 2.20 to 2.13 hours, with larger gains on weekends (2.07 to 1.85) than weekdays (2.26 to 2.23).
Methodology in Plain English
The researchers reformulate departure prediction as a survival problem: instead of asking "will the user leave in this 5-minute window?" (a heavily imbalanced classification question), they ask "what is the probability the user has not left yet at time t?" Each day is split into 5-minute intervals, and the model outputs a survival probability for every interval.
Each interval becomes a token containing three kinds of information: contextual signals from smartphone sensing (such as activity transitions, step counts, screen state, app usage, ambient light, and ambient sound), the absolute time of day, and day-of-week indicators. Contextual and day-of-week features are embedded together, absolute time is passed through a small neural network and scaled by a learnable parameter, and the two are combined using a learnable weight called alpha-fusion. Positional encodings are added, and a multi-layer Transformer encoder processes the whole sequence in parallel, with a sigmoid output layer producing the survival probability per interval.
Training uses an ordinal regression loss applied only up to the observed departure time in each sequence, encouraging high survival probability before the event and a sharp drop at the event. A Gaussian-smoothed weighting kernel softens supervision near the event boundary, and additional event weight and weekend weight factors control how strongly the event point and weekend sequences contribute. Departure is detected during inference when the predicted survival probability falls below a threshold p.
Data came from an IRB-approved field study: 506 Android users aged 18–69 were recruited in the Seoul metropolitan area between May 2021 and July 2022, using a custom app (EV Analyzer) that collected nine passive sensing streams. After filtering for data quality, temporal consistency, and realistic value ranges, the final dataset contained 93 participants with 42 days each, totaling 3,906 daily sequences. Ground-truth departure was defined as the first "in-vehicle" activity transition after 4:00 AM, and the event time was set 30 minutes before actual departure to leave charging time. Participants were split 4:1 into training and testing, further split 4:1 into train and validation, giving 60 training, 15 validation, and 18 test users. The model uses a 3-layer Transformer encoder with d_model = 32 and a single attention head, trained with Adam, early stopping, a maximum of 100 epochs, in PyTorch on an NVIDIA RTX 4090, with three random seeds (42, 43, 44). The threshold p was varied from 0.05 to 0.20 in increments of 0.05, with 0.10 selected as optimal.
Why This Matters
Accurate departure prediction is the control parameter that determines whether DFC actually reduces time spent at 100% SOC. Early predictions erase the battery-health benefit; late predictions risk an undercharged car and range anxiety. The paper positions this as a first step toward benchmark models and datasets for smart EV charging, aligning with regulatory pressure such as the European Union's 2027 mandate on battery health.
Real-world applications:
- Smart home and EV charger scheduling: Chargers could delay full charging until a predicted departure window, minimizing high-SOC dwell time.
- Battery health and warranty management: Fleets and manufacturers could use predicted departure to reduce degradation, extending pack lifetime and lowering replacement costs.
- Consumer mobile apps: A smartphone-based predictor (using the same passive sensing signals) could advise drivers on charging timing without requiring vehicle telemetry.
- Grid and demand-response programs: Knowing when vehicles will actually leave supports charge scheduling that shifts load and supports sustainable transportation systems.
Industry relevance spans EV manufacturers, charging network operators, battery management system developers, and utilities running managed charging programs, since the method uses unobtrusive smartphone sensing rather than dedicated vehicle hardware.
Future Directions
- Stronger personalization. The authors suggest meta-learning or adaptive regularization to better balance global patterns with individual behavior, and note that richer contextual signals from large language models may further improve personalization.
- Cold-start behavior. The paper states that a cold-start strategy and analysis of intra-user variability are provided in the supplementary material, leaving room for more systematic treatment of new users.
- Threshold and weight tuning in deployment. The interaction between event weight, weekend weight, and the survival threshold p determines whether errors occur early or late, so transferring these settings to new populations is an open question.
- Dataset breadth. The study recruited general smartphone users rather than EV owners under the assumption that departure routines are behaviorally consistent across populations; validating on actual EV owners and across regions remains open. The paper also notes the 2021–2022 collection period overlapped with pandemic-related schedule shifts.
Target Audience
This paper is most useful to machine learning researchers working on survival analysis and time-to-event modeling, mobility and smart-charging researchers, and EV or charging-infrastructure engineers interested in translating departure prediction into battery-health benefits. It also suits practitioners in battery management and demand-response systems who need a concrete benchmark for real-time departure prediction, and researchers interested in passive smartphone sensing and digital phenotyping who want a field-study example with real-world data.
Authors’ 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.