Configuration and Scheduling of Hybrid Energy Storage Systems via Variational Mode Decomposition and Deep Reinforcement Learning

To address the volatility challenges associated with high penetrations of renewable energy, this paper proposes a feedback‐coupled bilevel optimization architecture for hybrid energy storage systems (HESS) that integrates deep reinforcement learning (DRL). In the upper layer, variational mode decomposition (VMD) is employed to perform multiscale decomposition of fluctuating power profiles. Subsequently, the optimal allocation of power and capacity for the HESS is determined based on the specific response characteristics of each storage technology. The lower layer utilizes the proximal policy optimization (PPO) algorithm to optimize scheduling against stochastic fluctuations in renewable energy output and load demand. By establishing an agent–environment interaction mechanism and designing a specific reward function, the agent is guided to autonomously learn optimal scheduling strategies. Simulation results demonstrate that the proposed method effectively mitigates mode mixing. Furthermore, the VMD‐based allocation scheme reduces costs by 12.1% and 4.1% compared to unoptimized benchmarks and empirical mode decomposition (EMD) schemes, respectively. Moreover, the storage utilization and power supply self‐sufficiency rates under the PPO strategy significantly outperform traditional methods, validating the efficacy of the proposed architecture in enhancing both the economic viability and operational performance of the HESS.