A Two-Stage Dispatch Disaggregation Strategy for Virtual Power Plants with Multi-Objective Scheduling

Abstract

Virtual power plants (VPPs) must translate system-level net power instructions into executable responses of heterogeneous flexible resources while accounting for renewable uncertainty, economic performance, and low-carbon operation. This paper proposes a two-stage coordinated disaggregation–scheduling framework for VPPs. In Stage 1, heterogeneous resource characteristics and residual instruction features are represented in a unified multidimensional space, and a Transformer-enhanced deep Q-network is developed to learn sequential resource activation policies over multiple matching rounds under dynamically updated residual instructions and regulation margins. The resulting activation matrix is transferred to Stage 2 as fixed activation information to constrain the subsequent continuous dispatch. In Stage 2, a day-ahead tri-objective scheduling model jointly minimizes net peak shaving cost, wind–PV curtailment cost, and carbon trading cost. Renewable uncertainty is characterized using Wasserstein-distance-based dynamic confidence intervals, while a tiered carbon trading mechanism captures nonlinear emission costs. The resulting constrained multi-objective problem is solved using DP-NSGA-III. Case studies show that the proposed DQN reduces MAE and RMSE by 20.38% and 22.18%, respectively, while DP-NSGA-III reduces total cost by up to 14.56% compared with benchmark algorithms, demonstrating improved coordination between executable instruction disaggregation and low-carbon VPP scheduling.