⚡ electrical engineering

Deep Reinforcement Learning-Based Adaptive Coordination of DC Reactor Fault Current Limiters for Protection of Hybrid AC/DC Microgrids with High Renewable Penetration

This paper proposes a Deep Q-Network-based adaptive coordination framework for DC Reactor Fault Current Limiters that dynamically optimizes impedance and switching sequences in hybrid AC/DC microgrids, significantly outperforming conventional fixed strategies in fault current reduction, voltage restoration, and protection selectivity under high renewable energy penetration.

Original authors: B. Venkata Ramana¹, Sridhar Savarapu²

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: B. Venkata Ramana¹, Sridhar Savarapu²

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a hybrid AC/DC microgrid as a bustling, high-tech city with two types of power highways: one for alternating current (AC) and one for direct current (DC). This city is powered by a mix of traditional generators and a flood of renewable energy sources like wind turbines and solar panels. The problem? When a "traffic accident" (a short circuit or fault) happens, the old traffic cops (the protection relays) get confused. Because the renewable energy sources are so unpredictable and can send power in weird directions, the old rules for stopping the accident don't work. The traffic jams get worse, and the wrong streets get shut down.

For a long time, engineers tried to fix this with "smart barriers" called Fault Current Limiters (FCLs). Think of these as automatic speed bumps that pop up when a crash is detected to slow down the surge of electricity. However, the paper argues that the old way of using these barriers—setting them to a fixed height or using simple, pre-written rules—is like trying to drive a Formula 1 car with a map from 1990. It just doesn't work when the road conditions change, like when the city switches from being connected to the main power grid to running on its own (islanded mode), or when the sun and wind change their output.

The Paper's Big Idea: A Video Game AI for Power Grids

The authors propose a new, super-smart solution: a Deep Reinforcement Learning (DRL) agent. If the old protection system is a rulebook, this new system is a video game AI that learns by playing.

Here is how it works in the real world of the simulation:

The Training Ground: The researchers built a detailed digital twin of their hybrid microgrid in a computer program called MATLAB/Simulink. They didn't just test it once; they let their AI "player" crash into the grid 5,000 times in different scenarios. They threw everything at it: different types of electrical faults, different locations, and varying levels of renewable energy (from 20% up to 80% of the total power).
The Learning Process: The AI (specifically a Deep Q-Network, or DQN) learned to adjust the "speed bumps" (the DC Reactor FCLs) in real-time. It learned that sometimes it needs the barrier to be very high to stop a massive surge, and other times it needs to be lower to keep the voltage stable so the lights don't flicker out. It figured out the perfect balance by trying millions of combinations and getting "points" (rewards) for keeping the grid safe and "losing points" (penalties) when the grid got confused or the voltage dropped too low.
The Two-Stage Defense: The system has a two-step reaction plan.
- Stage 1 (The Reflex): The moment a fault is detected (within 0 to 10 milliseconds), the local hardware kicks in automatically to put up a maximum barrier. This is a fast, reflexive move that happens before the AI even finishes thinking.
- Stage 2 (The Brain): Just a split second later (after 10 milliseconds), the AI takes over. It looks at the whole city, calculates the exact best setting for every single barrier, and tells them to adjust. It does this every 5 milliseconds, constantly fine-tuning the protection as the situation changes.

What the Paper Says Works (and What It Doesn't)

The paper is very clear about what doesn't work. It explicitly rules out the idea that a single, fixed setting for the barriers can handle a modern grid with high renewable energy. It also argues against relying solely on pre-calculated rules that don't change when the grid switches modes (like going from grid-connected to islanded). The authors show that in their simulations, these older methods lead to "miscoordination," where the wrong breakers trip, leaving healthy parts of the city in the dark while the fault still burns.

In contrast, the AI-driven approach showed remarkable results in their simulations:

Zero Confusion: In 20 different test scenarios, the AI system had zero miscoordination events. The old fixed system had 3, and another adaptive method had 2.
Better Voltage: When a fault happened, the AI kept the voltage at 0.87 pu (per unit), whereas the best previous method only managed 0.50 pu. This means the lights stayed much brighter and the equipment safer.
Faster Response: The AI helped clear the faults faster, with a primary relay operating time of 312 ms, compared to 400 ms for the fixed system.
Current Reduction: It successfully reduced the peak fault current at the main connection point to 1.98 kA, a 47.9% reduction compared to having no limiter at all.

How Sure Are We?

It is important to note that these impressive numbers come from simulations. The authors ran their tests in a computer model, not on a physical power grid in the real world yet. They are very confident in the math and the logic, showing that the AI learned to handle everything from three-phase faults to line-to-ground faults, even when the renewable energy penetration was as high as 70%.

The paper concludes that while this is a major step forward, the next step is to prove it works on real hardware. The authors plan to test this on a "hardware-in-the-loop" system (where the AI controls a real-time simulator connected to physical equipment) and to see how it handles communication delays or cyber-attacks. But for now, in the digital world of their simulations, this AI "traffic cop" has learned to keep the hybrid city safe, fast, and stable in a way the old rulebooks never could.

Technical Summary: Deep Reinforcement Learning-Based Adaptive Coordination of DC Reactor Fault Current Limiters for Hybrid AC/DC Microgrids

Problem Statement
The increasing integration of renewable energy sources (RES) and distributed generation (DG) into hybrid AC/DC microgrids has fundamentally altered fault current characteristics, introducing bidirectional power flows and dynamic impedance profiles. These changes severely disrupt conventional protection coordination schemes, particularly for Directional Overcurrent Relays (DOCRs) and distance relays. The "infeed effect" from DG units often causes relays to misoperate (overreach or underreach), leading to coordination failures. While Fault Current Limiters (FCLs), specifically DC Reactor FCLs (DCR-FCLs), offer a hardware solution by limiting fault currents, existing coordination strategies rely on fixed impedance thresholds or offline optimization techniques (e.g., Genetic Algorithms, Particle Swarm Optimization). These static approaches fail to adapt to the rapid, stochastic changes in microgrid topology, islanding transitions, and varying RES penetration levels, rendering pre-computed settings suboptimal or invalid in real-time scenarios.

Methodology
The paper proposes a Deep Reinforcement Learning (DRL) framework utilizing a Deep Q-Network (DQN) agent to achieve adaptive coordination of DCR-FCLs. The methodology is structured as follows:

System Modeling: A hybrid AC/DC microgrid test system is modeled in MATLAB/Simulink, featuring an 11 kV AC sub-network and an 800 V DC sub-network interconnected by a bidirectional interlinking converter. The system includes wind turbines, PV units, battery storage, and diesel generators. DCR-FCL units are strategically placed at the Point of Common Coupling (PCC), DG terminals, and critical line sections.
DRL Formulation (MDP): The coordination problem is formulated as a Markov Decision Process (MDP):
- State Space ( $S$ ): Includes bus voltages, line currents, DG currents, estimated fault location, fault type, total RES generation, and operating mode (grid-connected vs. islanded).
- Action Space ( $A$ ): Discrete duty cycle settings ( $D \in \{0, 0.1, \dots, 1.0\}$ ) for the IGBT switches within each DCR-FCL unit, which directly control the effective limiting impedance ( $Z_{FCL}$ ).
- Reward Function ( $R$ ): A weighted sum of rewards for maintaining line currents within bounds, keeping bus voltages stable (0.85–1.05 pu), ensuring correct primary/backup relay coordination, and maintaining selectivity (isolating only the faulted zone).
Two-Stage Coordination Scheme:
- Stage I (Local Autonomous Response): Upon fault detection (0–10 ms), local controllers immediately engage maximum limiting impedance to suppress initial fault current peaks, independent of communication latency.
- Stage II (DRL Global Coordination): After the initial window, a central Decision-Making Center (DMC) hosting the trained DQN agent receives the full system state via an IoT layer. The agent computes optimal duty cycles for all FCLs and simultaneously updates DOCR settings to ensure optimal protection coordination (OPC). This process repeats every 5 ms.
Training: The DQN agent is trained over 5,000 episodes in a stochastic environment covering various fault types (LG, LL, LLG, 3Φ), locations, resistances, and RES penetration levels (20–80%).

Key Contributions
The paper claims the following specific contributions:

First Application of DRL to DCR-FCL: It presents the first application of Deep Q-Networks to the adaptive coordination of DC Reactor FCLs in hybrid AC/DC microgrids, moving beyond static optimization.
Unified Two-Stage Framework: It integrates FCL impedance control directly with DOCR setting optimization, creating a unified framework that dynamically balances fault current suppression, voltage support, and relay coordination.
Extension to Hybrid Topologies: The architecture extends previous multi-FCL coordination concepts (originally for radial AC systems) to complex hybrid AC/DC microgrids, incorporating interlinking converter dynamics and DC-side fault characteristics.
Comprehensive Benchmarking: The study provides a unified performance benchmark by comparing the proposed DRL scheme against fixed DCR-FCL strategies and adaptive Virtual Impedance (VI-FCL) schemes across diverse fault scenarios and RES penetration levels.

Results
Simulation results demonstrate that the proposed DRL-based approach significantly outperforms conventional and adaptive rule-based strategies:

Relay Coordination: The DRL scheme achieved zero miscoordination events across 20 test cases, whereas fixed FCL and adaptive VI-FCL schemes experienced 3 and 2 miscoordination events, respectively.
Fault Current Reduction: The DRL scheme reduced peak fault current at the PCC by 47.9% (to 1.98 kA), slightly outperforming the fixed (44.7%) and adaptive VI (47.4%) schemes.
Voltage Stability: During faults, the DRL scheme maintained a minimum bus voltage of 0.87 pu, a marked improvement over the 0.50 pu achieved by the best prior approach. This indicates superior ability to balance current limiting with voltage support.
Speed: The primary relay operating time was reduced to 312 ms under the DRL scheme, compared to 400 ms for fixed FCLs, leading to faster fault clearance.
Adaptability: The agent successfully adapted to islanded mode operation and high RES penetration (70%) without retraining, whereas static schemes failed or performed poorly in these dynamic conditions.
Real-Time Feasibility: The computational time for DQN inference was measured at 0.8 ms, well within the 5 ms decision interval, confirming real-time applicability.

Significance and Claims
The authors position this work as the culmination of a research program that has systematically advanced FCL technology from single-device applications to intelligent, AI-driven real-time coordination. The paper claims that the proposed framework addresses the fundamental limitation of existing protection strategies: their inability to adapt to the dynamic and uncertain operating conditions of modern microgrids. By leveraging DRL, the system can perceive the full network state and learn optimal policies that simultaneously optimize fault current suppression, voltage restoration, and relay coordination integrity. The authors assert that this approach provides a comprehensive solution for protecting hybrid AC/DC microgrids with high renewable penetration, offering a significant step forward in smart grid protection reliability.

Future work identified by the authors includes hardware-in-the-loop (HIL) validation, extension to multi-agent DRL for larger networks, robustness analysis against communication delays and cyber threats, and the investigation of model-based RL to reduce training data requirements.

More like this