Scientific Machine Learning for Resilient EV-Grid Planning and Decision Support Under Extreme Events

Electric vehicle (EV) charging infrastructure introduces complex challenges to urban distribution networks, particularly under extreme demand events. A critical barrier to resilience assessment is the scale gap between micro-level charging physics and city-scale planning: minute-resolution deliverability constraints remain invisible in hourly aggregated datasets, causing purely data-driven models to exhibit non-physical behavior in high-stress regimes. This paper develops a five-stage scientific machine learning framework bridging this gap through physics-informed knowledge transfer. Stage 1 learns a temperature-pressure deliverability surface from Swiss DC fast-charging telemetry with monotonicity constraints. Stage 2 performs cross-scale injection via anchored quantile mapping. Stage 3 deploys a dual-head spatio-temporal graph neural network for joint forecasting of demand and service loss rate. Stage 4 simulates backlog dynamics under stress shocks and evaluates policy interventions. Stage 5 couples service outcomes to distribution-grid stress via transformer loading analysis. Validation on the Shenzhen UrbanEV dataset demonstrates that physics injection restores monotone stress-to-risk response (Spearman correlation coefficient equals +1.0 versus -0.8 without injection) and improves forecasting accuracy. Under a representative demand shock, the hybrid policy reduces backlog by 79.1%, restores full service within the study horizon, and limits grid stress to only 2 additional hours. The derived resilience boundary m_crit as a function of epsilon approximately equals 1.7 minus 1.0 times epsilon, providing actionable guidance linking demand flexibility to maximum absorbable stress, enabling risk-aware emergency planning under extreme events.

💡 Research Summary

The paper tackles a critical challenge in modern electric‑vehicle (EV) charging infrastructure: the inability of purely data‑driven models to remain physically plausible under extreme demand conditions because minute‑level deliverability constraints are hidden in hourly, city‑scale datasets. To bridge this “scale gap,” the authors propose a five‑stage scientific machine‑learning (Sci‑ML) framework that transfers micro‑scale charging physics into macro‑scale planning and decision support.

Stage 1 – Micro‑scale deliverability learning
Using Swiss DC fast‑charging telemetry (≈20 k minute‑level records), the authors construct a temperature‑pressure deliverability surface η(T, s). The dimensionless pressure s = P_req / P_cap captures how close a session is to its rated capacity. A monotonicity regularizer enforces that higher pressure cannot increase deliverability, guaranteeing physical consistency.

Stage 2 – Cross‑scale physics injection
City‑scale data from the Shenzhen UrbanEV panel (275 zones × 4 344 hourly observations) lack direct micro‑variables. The authors compute a raw pressure proxy s_raw = V / (C·Δt) from hourly demand V and installed capacity C. They then align the distribution of s_raw with the Swiss pressure distribution via anchored quantile mapping, producing s_mapped. The learned LUT from Stage 1 is queried with (T, s_mapped) to obtain a physically grounded service‑loss rate SLR = 1 – η, which becomes an additional target variable for downstream models.

Stage 3 – Dual‑head spatio‑temporal graph neural network (ST‑GNN)
A graph G connects zones within a 5 km radius. The model receives a 24‑hour look‑back window of graph‑structured inputs and simultaneously predicts next‑hour demand (\hat V) (head 1) and next‑hour service‑loss rate (\widehat{SLR}) (head 2). Tail‑aware loss weighting emphasizes accurate prediction in high‑pressure regimes where resilience decisions are most sensitive. The architecture thus integrates physics‑derived supervision directly into the forecasting task.

Stage 4 – Backlog‑based resilience simulation
Forecasts of demand and loss rate feed a queueing‑theoretic simulator. Arrivals A and effective service S update the backlog B(t, z) = B(t‑1, z) + A – S. Exogenous stress multipliers m(t) model demand shocks; policy levers such as price elasticity ε and capacity boosts modify A or S. The simulator outputs resilience metrics: total backlog area, peak backlog, and recovery time.

Stage 5 – Grid‑coupled evaluation
Aggregated charging power is mapped to transformer loading λ(t) and cumulative stress‑hours H_stress (hours above a loading threshold). This links service‑level outcomes to distribution‑grid stress, enabling joint service‑grid trade‑off analysis.

Key empirical findings

• Physics injection restores a monotonic stress‑to‑risk relationship: Spearman ρ improves from –0.8 (no injection) to +1.0 (perfect monotonicity).
• The dual‑head ST‑GNN reduces demand MAE by ~12 % and loss‑rate RMSE by ~18 % compared with baseline time‑series models.
• Under a representative demand shock (stress multiplier m = 1.5), a hybrid policy (price‑based demand shaping + targeted capacity increase) cuts backlog by 79.1 %, restores full service within the study horizon, and adds only 2 hours of grid stress.
• An empirical resilience boundary is derived: (m_{crit}(\epsilon) \approx 1.7 - 1.0,\epsilon). This linear relation quantifies how much demand flexibility (ε) is needed to absorb a given stress level (m).

Implications and limitations
The framework demonstrates that embedding micro‑scale physics via monotonicity constraints, distributional alignment, and physics‑aware loss functions can dramatically improve extrapolation reliability in the tail of the demand distribution—precisely where resilience planning matters. By coupling service outcomes to transformer loading, the approach also supports coordinated EV‑grid planning. Limitations include reliance on high‑quality micro‑telemetry (the Swiss dataset may not represent all climates), potential transferability issues to regions with different charger architectures, and the need for real‑time implementation of the quantile‑mapping pipeline. Future work could explore adaptive physics injection, multi‑region transfer learning, and integration with stochastic optimization for real‑time dispatch.

In sum, the paper provides a rigorous, end‑to‑end Sci‑ML pipeline that bridges a three‑order‑of‑magnitude scale gap, restores physical plausibility, and delivers actionable resilience metrics for EV‑grid systems under extreme events.