⚡ electrical engineering

Closed-Loop Active Sampling Method for Static Voltage Stability Region Characterization in Multi-Terminal HVDC Systems

This paper proposes a closed-loop active sampling method that combines a label-extension algorithm with an iterative geometric selection process to efficiently and accurately characterize the static voltage stability region of multi-terminal HVDC systems, overcoming the computational inefficiency of random sampling and the conservatism of approximation-based approaches.

Original authors: Ying Wang, Xiaowei Ma, Chao Huo, Penghan Li, Cong Wang

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ying Wang, Xiaowei Ma, Chao Huo, Penghan Li, Cong Wang

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

Power grids are vast, intricate networks that must balance supply and demand in real time. In the modern era, these systems are becoming increasingly complex as they integrate renewable energy sources like offshore wind farms, which are often far from the cities they power. To move this electricity efficiently, engineers use high-voltage direct current lines that can connect multiple points across a wide area, forming what are known as multi-terminal systems. These networks are flexible and powerful, but they are also delicate. Just as a bridge has a maximum weight it can hold before collapsing, an electrical grid has a limit to how much power it can carry while maintaining a stable voltage. If the demand at one end changes too quickly or too much, the entire system can lose its balance, leading to a sudden and widespread blackout. The challenge for engineers is to know exactly where that limit lies before it is reached.

For decades, the standard way to find these safety limits has been to run thousands of computer simulations, testing random combinations of power levels to see which ones work and which ones cause the system to fail. This approach is like trying to map the edge of a cliff by walking randomly across a field; eventually, you might find the drop-off, but it takes a long time and you might miss the most dangerous spots. Other methods try to guess the shape of the cliff with simple shapes, but these guesses are often too cautious, telling operators to stay far away from the edge when they could actually go closer safely. Both methods have flaws: one is too slow, and the other is too conservative.

A team of researchers from the State Grid Corporation of China has developed a smarter way to map these safety zones. Instead of guessing randomly or relying on rough approximations, they created a method that learns as it goes, much like a hiker who studies the terrain to find the most efficient path to the edge of a precipice. Their approach, described in a recent study, uses a "closed-loop" process. This means the system constantly checks its own map, identifies the areas where it is most unsure, and then focuses its computational energy on testing those specific, uncertain points. By doing so, it builds a precise picture of the safe operating region without needing to test every single possibility.

The researchers began by simplifying the complex physics of the grid into a steady-state model. They treated the power sources as fixed voltage points and the power demands as constant loads, allowing them to view the entire system as a network of connected nodes. In this simplified view, the question of stability becomes a question of whether a mathematical solution exists for the power flowing through the network. If a solution exists, the system is stable; if not, it is on the verge of collapse. The goal was to define the exact boundary of all the power combinations that keep the system stable.

To achieve this, the team introduced a two-step strategy. First, they used a technique called "label extension." Once they verified that a specific combination of power levels was safe, they used mathematical properties of the system to prove that a whole range of similar, slightly higher power levels would also be safe. This allowed them to expand their known safe zone instantly without running new simulations for every single point. Second, they built a geometric shape, specifically a convex hull, around the verified safe points. This shape acts as a guaranteed inner core of safety, ensuring that any point inside it is definitely stable.

However, a shape built from a few points is not the whole story. To refine the map, the researchers added an active learning component. They trained a computer classifier to predict where the boundary between safe and unsafe might be. This classifier does not just guess; it calculates a level of uncertainty for every point in the grid. The system then selects the points where it is most confused—the areas right near the edge of the known safe zone—and sends them for a rigorous, high-fidelity simulation. The results of these targeted tests are fed back into the system, updating the map and shrinking the area of uncertainty. This cycle repeats, with the system getting smarter and the map getting more detailed with each round.

The team tested this method on a simulated five-terminal high-voltage direct current system, a setup that mimics a real-world network with two power sources and three load points. The results were striking. In just 18 rounds of this iterative process, the method was able to map the stability region with an accuracy score of 0.98, a measure of how well the predicted boundary matched the true physical limits. More importantly, it achieved this level of accuracy in an average of 27 seconds. When compared to a standard machine learning approach that relied on random sampling, which took 51 seconds to reach the same level of accuracy, the new method was nearly twice as fast.

The study demonstrates that by combining mathematical guarantees with intelligent, targeted sampling, engineers can map the safety limits of complex power grids much more efficiently than before. The method does not just find a safe zone; it finds the true edge of that zone with high precision, avoiding the excessive caution that leaves valuable transmission capacity unused. While the current work was validated through computer simulations on a specific network topology, the results suggest a powerful new tool for managing the increasingly complex grids of the future. As renewable energy integration grows and power systems become more interconnected, having a fast and accurate way to understand where the limits lie will be essential for keeping the lights on.

Technical Summary: Closed-Loop Active Sampling Method for Static Voltage Stability Region Characterization in Multi-Terminal HVDC Systems

Problem Statement
Multi-terminal High-Voltage Direct Current (MT-HVDC) systems are critical for enhancing transmission flexibility and integrating offshore renewables. However, the nonlinear power flow equations and complex inter-terminal couplings make characterizing the Static Voltage Stability Region (SVSR)—the set of terminal power profiles admitting a feasible steady-state equilibrium—extremely difficult. Existing numerical methods rely on extensive random sampling and repeated power-flow calculations, which become computationally prohibitive as the number of terminals increases. Conversely, approximation-based methods improve efficiency but often yield conservative stability boundaries that lack fidelity near critical operating conditions, potentially underestimating admissible operating ranges. Furthermore, while dynamic stability in MT-HVDC systems is well-studied, SVSR characterization has received limited attention in the literature.

Methodology
The paper proposes a closed-loop active sampling method to efficiently characterize the SVSR in the terminal power-profile space. The approach is built upon a steady-state equivalent DC network model where sending-end converters are treated as controlled voltage sources and receiving-end converters as constant power loads (CPLs). The methodology consists of two primary modules:

Rapid Labeling via Guaranteed Subsets:
Instead of performing numerical solvability checks for every candidate, the method constructs a guaranteed inner approximation of the SVSR using verified solvable profiles. This involves two complementary geometric constructions:
- Interval-based Subsets: Leveraging the monotone solvability property of the MT-HVDC model, any power profile $p$ that dominates a verified solvable profile $p^{(i)}$ (i.e., $p \succeq p^{(i)}$ ) is automatically labeled as solvable.
- Convex Subset: The convex hull of the set of verified solvable profiles is constructed to capture correlations across dimensions.
  The union of these subsets forms a guaranteed inner region ( $\mathcal{M}_g$ ). Any candidate profile falling within $\mathcal{M}_g$ is labeled as "solvable" without further numerical simulation, significantly reducing the need for dense random sampling.
Geometry-Guided Closed-Loop Active Sampling:
For profiles outside the guaranteed region (labeled "undecided"), the method employs an iterative active learning loop:
- A classifier (implemented as a Multi-Layer Perceptron) estimates the posterior probability of a candidate profile being solvable.
- An uncertainty metric quantifies the informativeness of unlabeled samples, prioritizing those near the current decision boundary.
- The most informative candidates are selected for rigorous numerical verification (using an IPOPT-based solver).
- Newly verified solvable samples are added to the training set, expanding the interval-based and convex subsets, and the process repeats. This closed-loop procedure progressively tightens the SVSR characterization while concentrating computational effort on the most impactful boundary regions.

Key Contributions
The paper outlines three main contributions:

Steady-State Equivalent Modeling: The establishment of a steady-state equivalent model for MT-HVDC systems in the terminal power-profile parameter space, casting SVSR characterization as a region-identification problem based on DC network solvability.
Network-Informed Framework: The development of a framework that combines steady-state solvability testing with label extrapolation (via interval and convex subsets). This allows for the construction of a guaranteed inner SVSR subset from a limited number of verified profiles, reducing reliance on large-scale random sampling.
Closed-Loop Active Sampling Strategy: The design of an iterative refinement scheme that actively selects terminal power profiles for verification based on the geometry of the current SVSR approximation. This strategy focuses computational resources on the most informative candidates, improving characterization efficiency.

Results
The proposed method was validated using a five-terminal HVDC test system (two source terminals, three load terminals) with a three-dimensional SVSR.

Accuracy: The method successfully captured the main geometric features of the SVSR boundary, achieving an F1 score of 0.98 after 18 iterations (epochs).
Efficiency: Compared to a conventional machine learning baseline that relies on random sampling, the proposed closed-loop active sampling approach reached the target F1 score of 0.98 significantly faster. The average runtime was 27 seconds for the proposed method versus 51 seconds for the baseline, demonstrating a substantial reduction in computational effort.
Visualization: The uncertainty heatmaps showed that the sampling strategy effectively concentrated queries on the transition regions between stable and unstable operating points, rather than exploring the space uniformly.

Significance
The paper claims that the proposed framework provides an accurate and computationally efficient tool for SVSR characterization in MT-HVDC systems. By integrating network physics (monotonicity and convexity) with active learning, the method overcomes the trade-off between the high computational cost of exhaustive numerical methods and the low fidelity of approximation-based methods. The results suggest that accurate SVSR boundaries can be recovered with reduced conservativeness and significantly fewer power-flow calculations, offering a practical solution for secure scheduling and risk-aware early warning in complex MT-HVDC networks. Future work is noted to extend this approach to higher-dimensional spaces and more complex topologies.

More like this