Learn to Coordinate City-Scale Virtual Power Plants: A Reasoning-Guided Evolutionary Framework for Hierarchical Heterogeneous Device Scheduling

Abstract

As intermittent renewables increasingly penetrate power systems, virtual power plants (VPPs) have emerged as a critical component of power systems, aggregating geographically distributed devices to respond to price signals and mitigate supply-demand imbalances. Consequently, coordinating heterogeneous resources across hierarchical multi-region pricing to fulfill committed bids while maximizing arbitrage has become a critical challenge. Existing operations-research and reinforcement-learning approaches rely on handcrafted formulations or learned policies that struggle to scale or lack interpretability, limiting trustworthiness in safety-critical energy systems. In contrast, large language models enable evolving interpretable and effective optimization by reasoning over structured decisions and generating executable programs. Building on this paradigm, we propose VPPEvolve, a reasoning-guided evolutionary framework for hierarchical VPP scheduling. VPPEvolve tackles three key challenges through three complementary designs: (i) an evolvable chained program representation that formalizes hierarchical VPP dynamics through executable structures; (ii) a spatio-temporal profiling and reflection module that bridges the semantic gap between volatile numerical signals and structured reasoning space; and (iii) a device attribution-aware inspiration module that enables LLM-informed evolution by explicitly attributing device-level contributions to improve heterogeneous coordination. Extensive experiments on two city-scale datasets confirm consistent gains in economic profit and bid-tracking stability, with post-hoc analyses and deployment confirming a white-box paradigm that reduces grid-side stress. Codes and data are available at: https://github.com/JinweiZzz/VPPEvolve.