OIRL: Real-time Optimization of Peak Shaving and Energy Arbitrage in Industrial Microgrids with Offline Inverse Reinforcement Learning Framework

View at publisher Summarise with VIRAT
Authors: Ying Hu, Sen Li
Cited by: 0
References: 9

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Industrial microgrids offer a promising pathway toward sustainable manufacturing, but their real-time operation presents significant challenges in concurrently optimizing peak shaving and energy arbitrage. These challenges stem from renewable energy volatility, complex temporal dependencies in peak demand charges, and conflicting operational objectives that existing model-based approaches and standard reinforcement learning methods fail to adequately address. To overcome these limitations, we present a novel Offline Inverse Reinforcement Learning (OIRL) framework that formulates the optimization problem as a soft Markov Decision Process (MDP). The proposed solution features a dual reward-policy iterative optimization mechanism that uniquely integrates contrastive reward learning from expert demonstrations with conservative soft Q-learning. This innovative combination enables precise temporal reward decomposition for monthly operational costs while guaranteeing stable and sample-efficient policy improvement through regularized soft value function updates. Comprehensive validation using real-world data from a Chinese industrial microgrid shows our framework achieves remarkable performance improvements of 117.9% in economic efficiency and 226.4% in peak shaving capability compared to state-of-the-art baselines, while demonstrating consistent robustness across varying operational conditions.

Fields of Study & Subjects

Microgrid Control and Optimization Smart Grid Energy Management Electric Power System Optimization

Publication details

Journal / source
IEEE PES International Meeting (PES IM)
Pages
1–5
Publication date
18 January 2026
Language
EN