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

TL;DR

This work tackles resilience planning for urban EV charging by bridging minute-scale deliverability physics with city-scale forecasting through a five-stage cross-scale Sci-ML framework. A monotonic deliverability surface learned from micro telemetry is anchored to city-scale data via anchored quantile mapping and injected into a dual-head spatio-temporal GNN, enabling tail-aware predictions and physically consistent stress responses. A backlog-based simulator evaluates policy interventions (price shaping, capacity boosts, and hybrids) under shocks, with grid coupling to transformer loading to quantify service–grid trade-offs. The key finding is that physics injection restores monotone stress sensitivity and reveals a resilience boundary $m_{crit}(\varepsilon) \approx 1.7 - 1.0\varepsilon$, providing actionable guidance for emergency planning and risk-aware decision making in extreme events. Practically, the framework supports coordinated EV service planning and grid management, delivering quantitative benchmarks for policy design and resilience assessment in urban power systems.

Abstract

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.

Figures (12)

• Figure 1: Data sources and cross-scale complementarity. Swiss minute-level telemetry provides micro deliverability physics with high overload exposure (50.4% of observations exceeding $s=1$); Shenzhen zone-hour panel provides city-scale demand dynamics with aggregated pressure proxies ($<$0.1% exceeding $s=1$). This three-order-of-magnitude difference in overload exposure motivates distributional alignment.
• Figure 2: [TO BE ADDED] Overview of the five-stage cross-scale physics-injected scientific machine learning framework.
• Figure 3: Micro-scale deliverability surface $\eta(T,s)$ learned from Swiss minute-level data (Stage 1). The surface exhibits monotonically decreasing efficiency as pressure increases, with temperature-dependent degradation patterns consistent with thermal derating physics.
• Figure 4: Pressure distribution comparison motivating cross-domain alignment (Stage 2). Swiss micro-scale data shows 50.4% of observations exceeding $s=1$; Shenzhen macro-scale shows $<$0.1%. Exceedance curves illustrate the three-order-of-magnitude difference requiring distributional alignment before physics injection.
• Figure 5: Anchored quantile mapping diagnostics (Stage 2). The mapping aligns Shenzhen raw pressure proxies to the Swiss pressure domain while enforcing $s_{mapped}(s_{raw}=1)=1$, ensuring the capacity boundary remains physically interpretable after alignment. Panels show: (a) empirical CDFs before and after mapping; (b) Q--Q diagnostics; (c) inverse-CDF mapping function with anchor point.