Distributed hybrid energy storage configuration framework based on spatiotemporal mutual cross-attention

To address spatiotemporal uncertainties in power distribution networks with high renewable energy penetration, this paper proposes a distributed electricity–heat–hydrogen hybrid energy storage system (HESS) configuration framework based on a multi-agent deep reinforcement learning (DRL) model. The coordinated electricity–heat–hydrogen operation framework is established to characterize the complementary behavior of battery, thermal, and hydrogen storage across time periods and network locations. A multi-objective model is then formulated to describe the siting, sizing, dispatch, and operational constraints of the HESS while considering the issues of voltage security, lifecycle cost, and flexibility inadequacy. To enable adaptive decision-making under uncertainty, the constraints and objective terms are incorporated into a multi-agent DRL environment. Temporal attention is introduced to capture the dynamic dependencies of load and renewable outputs, while a graph attention network is used to model spatial coupling among distribution-network nodes. Furthermore, a spatiotemporal mutual cross-attention (STMCA) mechanism is designed to realize bidirectional fusion between temporal and spatial features and enhance policy learning stability. Case studies on the modified IEEE 33-bus system have shown satisfactory results on mitigating renewable energy fluctuations and enhancing distribution-network stability, in which the proposed method converges after approximately 600 episodes and achieves a final reward of 0.980 ± 0.006 p.u., which is approximately 5.6% and 1.9% higher than those of the DDPG and MADDPG baseline algorithms under the same test conditions, respectively. The optimized configuration maintains nodal voltages within 0.97–1.00 p.u.