Safety-Certified Hierarchical Control of DC Microgrids via Continuous Deep Reinforcement Learning and Covariance-Weighted Sliding Mode Synthesis
- 1Laboratory of Advanced Systems, Polytechnic School of Tunisia (EPT), B.P. 743, Marsa 2078, Tunisia
- 2Department of Physics, College of Sciences, Umm Al-Qura University, Makkah 21955, Saudi Arabia
Abstract
This paper presents a hierarchical safe control approach for DC microgrids, combining learning-based energy management with rigorous constraint enforcement. The main novelty is a covariance-aware coupling mechanism that simultaneously informs CBF constraint tightening, sliding-mode gain scheduling, and DRL state augmentation. A continuous Soft Actor-Critic agent generates smooth current references and a flexible load-shedding ratio. A predictive CBF quadratic program acts as a safety filter, using online confidence intervals from a recursive least-squares predictor. The filter ensures a sampled-data safety property with an inter-sample margin, applying to the actual uncertain microgrid via a robust-CBF term, provided the QP is feasible with zero slack and the true voltage stays within the confidence interval. The same covariance information serves as an online uncertainty metric for adaptive gain and weight scheduling. A composite Lyapunov–CBF theorem proves uniform ultimate boundedness and constraint satisfaction. Simulations under severe conditions show high-probability constraint satisfaction with negligible slack, while improving transient recovery and reducing battery stress. Comparisons with fuzzy logic, discrete DQL, and plain SAC isolate the CBF contribution, and ablation studies assess individual components. While robust MPC can also enforce constraints, the proposed method achieves comparable satisfaction with lower complexity. A clear distinction is maintained between theoretical guarantees and simulation observations.
Related
- Reinforcement Learning for Hierarchical Reasoning Rewards: Minimax-Optimal Rates with Transformers
- Average-Reward Reinforcement Learning for Multichain MDPs: A Hierarchical Decomposition Approach
- HALO: Hierarchical Evidence Selection and Reinforcement Learning for Allocating Budgeted Label Optimization from Longitudinal Clinical Records
- Continuous Delayed-Memory Stochastic Gradient Descent and Continuous-Time Reinforcement Learning from History of Astrophysical Time Series Studies
- Continuous-Time Reinforcement Learning for N-Player Stochastic Differential Games with Exploratory Policies
BibTeX
@article{jouili2026safetycertified,
title = {{Safety-Certified Hierarchical Control of DC Microgrids via Continuous Deep Reinforcement Learning and Covariance-Weighted Sliding Mode Synthesis}},
author = {Khalil Jouili and Monia Charfeddine and Mongi Ben Moussa},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14203766},
url = {https://www.mdpi.com/2227-7390/14/20/3766},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Safety-Certified Hierarchical Control of DC Microgrids via Continuous Deep Reinforcement Learning and Covariance-Weighted Sliding Mode Synthesis
AU - Khalil Jouili
AU - Monia Charfeddine
AU - Mongi Ben Moussa
PY - 2026
DA - 2026/10/16
JO - Mathematics
DO - 10.3390/math14203766
UR - https://www.mdpi.com/2227-7390/14/20/3766
AB - This paper presents a hierarchical safe control approach for DC microgrids, combining learning-based energy management with rigorous constraint enforcement. The main novelty is a covariance-aware coupling mechanism that simultaneously informs CBF constraint tightening, sliding-mode gain scheduling, and DRL state augmentation. A continuous Soft Actor-Critic agent generates smooth current references and a flexible load-shedding ratio. A predictive CBF quadratic program acts as a safety filter, using online confidence intervals from a recursive least-squares predictor. The filter ensures a sampled-data safety property with an inter-sample margin, applying to the actual uncertain microgrid via a robust-CBF term, provided the QP is feasible with zero slack and the true voltage stays within the confidence interval. The same covariance information serves as an online uncertainty metric for adaptive gain and weight scheduling. A composite Lyapunov–CBF theorem proves uniform ultimate boundedness and constraint satisfaction. Simulations under severe conditions show high-probability constraint satisfaction with negligible slack, while improving transient recovery and reducing battery stress. Comparisons with fuzzy logic, discrete DQL, and plain SAC isolate the CBF contribution, and ablation studies assess individual components. While robust MPC can also enforce constraints, the proposed method achieves comparable satisfaction with lower complexity. A clear distinction is maintained between theoretical guarantees and simulation observations.
SN - 2227-7390
ER -