Semantic-mask-enhanced deep reinforcement learning for energy storage scheduling in virtual building-cluster microgrids with vehicle-to-building flexibility

Jiang, Ben, Zhang, Chengyu, Li, Yu, Rezgui, Yacine ORCID: https://orcid.org/0000-0002-5711-8400, Sun, Chang, Ghoroghi, Ali, Wang, Peng, Gu, Xianliang, Xie, Xiaolan, Yang, Yuee, Yu, Qiang and Zhao, Tianyi 2026. Semantic-mask-enhanced deep reinforcement learning for energy storage scheduling in virtual building-cluster microgrids with vehicle-to-building flexibility. Energy , 142835. 10.1016/j.energy.2026.142835

Full text not available from this repository.

Abstract

To address the optimal operation of building-cluster microgrids integrating multiple distributed energy resources and vehicle-to-building (V2B) flexibility, this study constructs a 30-day virtual environment using adapted multi-source building load data and public electric vehicle (EV) charging records. Vehicle arrival, charging, discharge eligibility, and departure are modeled under time-of-use prices. A state-dependent semantic action masking mechanism is developed for deep reinforcement learning (DRL) based stationary energy storage system (ESS) scheduling. The mechanism integrates ESS physical constraints and operational information to dynamically construct a state-dependent candidate action set. Deep Q-Network (DQN), Dueling Double Deep Q-Network (D3QN), and Proximal Policy Optimization (PPO) select the final ESS action. Under the V2B-enabled condition, EVs supply 141.5 MWh to buildings, and the average net electricity cost per unit charging energy decreases by 24.2%. Across six paired comparisons, semantic action masking reduces operating cost and total grid purchase by averages of 1.27% and 3.77%, with backbone-dependent effects on peak-period grid purchase. The action modification rate decreases by 56.18 percentage points, and the variability of total episode returns during late-stage training decreases by 51.05%. Among the evaluated configurations, PPO with semantic action masking and V2B achieves the lowest operating cost, reducing operating cost, total grid purchase, and peak-period grid purchase by 18.5%, 13.2%, and 44.7%, respectively, compared with the baseline configuration without ESS or V2B. These findings provide a methodological reference for the intelligent operation and coordinated utilization of flexible energy resources in building-cluster microgrids.

Item Type:	Article
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Engineering
Publisher:	Elsevier BV
ISSN:	0360-5442
Date of Acceptance:	18 October 2026
Last Modified:	26 Oct 2026 11:45
URI:	https://orca.cardiff.ac.uk/id/eprint/190417

Actions (repository staff only)

Edit Item