PAR-LLM: Peak-Aware Residual LLM Over Anchor for EV Charging Station Demand

Urban electric vehicle (EV) charging data display strong station-level heterogeneity, nonstationary patterns, and sharp peaks with long-tail surges. These characteristics undermine conventional forecasting: zone-specific models face data scarcity and unstable scales; global models dilute cross-zone differences; and graph or sequence architectures handle abrupt surges and drift under distribution shifts poorly. Therefore, we propose Peak-Aware Residual LLM (PAR-LLM) framework decoupling dense numerical learning from context-conditioned adjustments. A global long short-term memory (LSTM) with zone embeddings provides a stable baseline by learning short-term dynamics from normalized station data. A risk localizer identifies critical periods including sudden surges or daily peaks by analyzing historical volatility. For flagged pairs, a selective large language model (LLM) adjusts residuals based on recent statistics and calendar cues. These adjustments are merged with the numeric anchor through a discrepancy-aware weighting scheme, while guardrails constrain temporal variation and apply a soft reversion toward hour-of-day baselines to maintain physical plausibility. Evaluated on the Urban EV hourly dataset (120 zones) across three forecast horizons, the hybrid method outperforms graph-, sequence-, and tree-based models while keeping interventions sparse and latency controlled. The design addresses scale mismatch, cross-zone heterogeneity, and peak sensitivity by limiting language-driven corrections to high-risk hours and relying on the numeric anchor for routine behavior.