High-Fidelity Digital Twin Dataset Generation for Inverter-Based Microgrids Under Multi-Scenario Disturbances

This paper introduces a high-fidelity, labeled digital twin dataset generated from a MATLAB/Simulink electromagnetic transient model of a ten-inverter microgrid, capturing synchronized three-phase waveforms and power metrics across eleven distinct operating and disturbance scenarios to serve as a robust benchmark for training surrogate models and analyzing cyber-physical resilience.

The Gist

Imagine you are trying to teach a robot how to drive a car. You could give it a textbook full of rules about traffic laws and engine mechanics. But if you want the robot to actually drive in a storm, handle a sudden flat tire, or react to a child running into the street, you need something much better: real-world experience.

This paper is about creating a massive, ultra-detailed "driving simulator" for the electrical grid, specifically for modern neighborhoods powered by solar panels and batteries (called microgrids).

Here is the breakdown of what the authors did, using simple analogies:

1. The Problem: The "Blurry" Old Maps

For a long time, scientists studying the power grid used old, low-resolution maps. These maps showed the "average" speed of electricity and the general flow of power. They were great for planning where to build a new power plant, but they were useless for understanding what happens in a split-second when a storm hits or a solar panel suddenly shuts off.

• The Analogy: It's like trying to learn how to surf by looking at a photo of the ocean. You can see the waves, but you can't feel the water, the speed, or the sudden drop. The old data was too slow and blurry to teach computers how to handle the fast, chaotic "storms" of modern electricity.

2. The Solution: The "High-Speed Camera" (Digital Twin)

The authors built a Digital Twin. Think of this as a perfect, virtual clone of a real neighborhood's power grid, running inside a supercomputer.

• The Speed: They didn't just take a photo; they filmed the action with a camera that takes 500,000 pictures every second (a time step of 2 microseconds).
• The Detail: They recorded everything: the voltage (pressure), the current (flow), the frequency (speed), and exactly what every single solar panel and battery was doing.
• The Result: They created a dataset (a giant library of data) that captures the "electromagnetic transient" (EMT) behavior. This is the split-second "shock" and "recovery" of the grid, which happens so fast that normal tools miss it completely.

3. The "Training Drills" (The 11 Scenarios)

To make sure their AI models learn well, they didn't just let the grid sit there. They put the virtual grid through 11 different "stress tests" or drills, like a pilot training for emergencies.

1. Normal Day: Just cruising along.
2. Sudden Load: Turning on a massive air conditioner instantly (like a sudden traffic jam).
3. Voltage Sag: A temporary power dip (like a car hitting a pothole).
4. Slow Ramp: Gradually turning up the heat (a slow, steady change).
5. Frequency Ramp: The grid speed slowly speeding up or slowing down.
6. Generator Trip: One solar panel suddenly dies (like a car engine cutting out).
7. Island Mode: The neighborhood gets cut off from the main city grid (like a boat detaching from a tugboat).
8. Reactive Power: A sudden change in how much "push" the electricity needs.
9. Single Fault: One wire touches the ground (a specific type of short circuit).
10. Noise: Adding static to the radio signal (simulating bad sensors).
11. Communication Delay: The time it takes for the solar panels to talk to the battery (simulating a slow internet connection).

4. The "Safety Check" (Validation)

You might think, "Okay, you simulated these events, but how do you know the simulation is real?"

The authors didn't just trust the computer. They acted like detectives. After every simulation, they checked the "evidence."

• The Analogy: If you simulate a car crash, you don't just say "crash happened." You check the crumpled metal, the skid marks, and the airbags.
• In the paper: They checked if the frequency dropped when a generator died, if the voltage dipped during a fault, and if the power shifted correctly. They proved that the data wasn't just random numbers; it followed the actual laws of physics.

5. Why Does This Matter? (The "Surrogate Model")

The ultimate goal is to train Surrogate Models.

• The Analogy: Imagine a super-fast AI that can predict what will happen to the grid in the next millisecond.
• The Problem: Real simulations (like the one the authors built) are incredibly accurate but take a long time to run. You can't run them in real-time during a crisis.
• The Solution: You feed this high-fidelity dataset into a Machine Learning model. The model learns the patterns. Once trained, the AI can predict the grid's behavior instantly, almost like a reflex, allowing for real-time protection and control.

Summary

This paper is about building the ultimate training manual for the future of electricity.

• Old way: Slow, blurry, missing the fast stuff.
• New way: A high-speed, ultra-detailed video of a virtual grid going through 11 different disasters and recoveries.
• Goal: To teach AI how to keep our lights on, even when the grid gets hit by a storm, a glitch, or a cyber-attack.

The authors are essentially saying: "We built the perfect simulator, ran the drills, checked the physics, and now we are giving this data to the world so we can build smarter, faster, and safer power grids."

Technical Summary

Here is a detailed technical summary of the paper "High-Fidelity Digital Twin Dataset Generation for Inverter-Based Microgrids Under Multi-Scenario Disturbances."

1. Problem Statement

Publicly available power-system datasets often lack the granularity required for modern inverter-based microgrid research. Specifically, they suffer from three major limitations:

• Lack of Electromagnetic Transient (EMT) Data: Most datasets provide slow phasor or steady-state measurements (seconds to minutes) rather than microsecond-level waveforms, failing to capture fast inverter control loops and transient dynamics.
• Insufficient Disturbance Diversity: Existing data rarely covers a wide range of realistic operating conditions, including cyber-physical stressors like communication delays and measurement noise.
• Poor Suitability for Surrogate Modeling: Data-driven surrogate models (used for real-time prediction and control) require high-quality, labeled training data that accurately reflects physical system responses. Current datasets often lack the synchronized, multi-channel, high-resolution data needed to train these models effectively.

2. Methodology

The authors developed a high-fidelity digital twin dataset generation pipeline using MATLAB/Simulink with the following core components:

• System Model:

  • A low-voltage AC microgrid topology featuring 10 inverter-based Distributed Generation (DG) units.
  • Control Architecture: Each inverter utilizes a detailed grid-following control structure, including inner current/voltage loops, outer droop control (P-f and Q-V), and Phase-Locked Loops (PLL).
  • Simulation Engine: EMT-style simulation with a fixed time step of $\Delta t = 2 \mu s$ to capture fast transients without down-sampling.

• Measurement Architecture:

  • 38 Synchronized Channels: Data is recorded for every time step, including:
    • Point of Common Coupling (PCC): 3-phase voltages and currents.
    • Per-DG (10 units): Active power, reactive power, and frequency.
    • System-wide: Time stamps and scenario labels.
  • Sampling: $N = 500,001$ samples per scenario over a duration of $T = 1$ second.

• Scenario Design:

  • 11 Distinct Scenarios: The dataset covers normal operation, electrical disturbances (load step, load ramp, voltage sag/3-phase fault, single-line-to-ground fault, DG trip, tie-line trip, reactive power step), and cyber-physical effects (measurement noise injection, communication delay).
  • Deterministic Labeling: Scenario labels are generated internally within the digital twin and synchronized with physical events to ensure precise temporal alignment.

• Data Processing & Validation:

  • Cleaning: A repair-based strategy is used to handle numerical instabilities (NaN, Inf, outliers) via linear interpolation, preserving the full signal length and temporal alignment.
  • System-Level Validation: Each scenario is validated against physical evidence (e.g., frequency dips, voltage magnitude drops, zero-sequence current rise) to confirm that the labeled disturbance corresponds to a real, observable system response.

3. Key Contributions

• High-Fidelity EMT Dataset: The creation of a labeled, multi-scenario dataset with microsecond resolution ($2 \mu s$), specifically designed for inverter-dominated microgrids.
• Synchronized Multi-Channel Structure: A consistent schema of 38 channels across all scenarios, enabling direct use in supervised machine learning without complex alignment.
• Comprehensive Scenario Coverage: Inclusion of both traditional electrical faults and emerging cyber-physical stressors (noise, delay), which are critical for testing robustness.
• Physics-Based Validation: Unlike many datasets that rely solely on labels, this work validates every scenario using system-level metrics (frequency, power, voltage unbalance) to ensure physical realism.
• Open Access: The dataset and processing scripts are made available to the community to serve as a benchmark for surrogate modeling.

4. Results and Validation

The paper provides extensive validation evidence for each of the 11 scenarios, demonstrating that the dataset accurately captures expected physical behaviors:

• Load Step: Shows immediate active power increase, frequency dip, and subsequent recovery via droop control.
• Voltage Sag: Captures voltage reduction, inverter ride-through behavior, and power redistribution.
• Generator/Tie-Line Trips: Demonstrates sharp power drops, frequency transients, and load redistribution among remaining units.
• Faults (SLG): Clearly shows voltage unbalance and the emergence of zero-sequence currents.
• Cyber-Physical Effects:
  • Noise Injection: Successfully adds high-frequency variance to raw signals while preserving the underlying trend.
  • Communication Delay: Produces subtle deviations in frequency and power-sharing dynamics without sharp electrical transients, mimicking real control loop latency.

Statistical analysis confirms that the cleaned data maintains numerical stability while preserving the transient features essential for training dynamic models.

5. Significance

This work bridges a critical gap between traditional power system simulation and modern data-driven engineering:

• Enables Advanced Surrogate Modeling: Provides the high-resolution, transient-rich data necessary to train AI models for real-time prediction, disturbance classification, and control optimization in inverter-heavy grids.
• Cyber-Physical Resilience: Offers a unique benchmark for studying how communication delays and sensor noise impact microgrid stability, a topic often overlooked in standard power datasets.
• Reproducibility: By standardizing the topology, controller parameters, and event timing, the dataset allows for fair, apples-to-apples comparison of different machine learning architectures (e.g., CNNs, RNNs, Transformers).
• Future-Proofing: As power systems transition to 100% inverter-based resources, this dataset serves as a foundational resource for developing the next generation of resilient, data-driven grid management tools.