Attentive primal-dual policy optimization for grid-constrained charging guidance of heterogeneous electric vehicle fleets

Effective electric vehicle (EV) charging guidance is essential for urban electrification, necessitating satisfactory energy replenishment to alleviate range anxiety while respecting grid operational security limits. This task is fundamentally a constrained spatiotemporal optimization problem in the context of integrated distribution-transportation systems (IDTS), where the superposition of heterogeneous user archetypes and voltage-responsive power modulations induces significant observational ambiguity, rendering instantaneous observations insufficient for optimal decision-making. To address these challenges, this paper introduces the novel attentive primal-dual policy optimization (AttPDO) approach. AttPDO integrates an attentive encoder to perform latent state reconstruction from historical observation trajectories, effectively unveiling the hidden turnover patterns to facilitate foresighted guidance decisions. Building upon these distilled representations, a decoupled primal-dual optimizer employs dual-critic networks to independently evaluate user satisfaction and grid security, enabling adaptive constraint management via a Lagrangian mechanism. Finally, validated across two coupled testbeds of varying scales, AttPDO demonstrates superior robustness and scalability against state-of-the-art baselines. Notably, in a large-scale regional network in Hong Kong involving 10,000 EVs, the proposed framework achieves a 36.5% improvement in grid operational stability while maintaining high service reliability. Evaluation results underscore AttPDO’s potential as a potent solution for grid-constrained guidance within stochastic IDTS.