DC-Screening-Oriented Solver-Referral Auditing for Grid Foundation Model Predictions under Topology and Operating Stress

The electric grid is a vast, invisible machine that must remain in perfect balance at every second. Generators push power into the system, and homes and factories pull it out, all while the current flows through a web of wires that can be damaged by storms, overloaded by demand, or altered by maintenance crews. For decades, engineers have relied on complex mathematical models to predict whether this delicate balance will hold or if the system will collapse. These models are incredibly accurate but also incredibly slow; running them for every possible scenario, from a single broken wire to a heatwave that spikes demand, would take too long to be useful for real-time decisions. In recent years, scientists have begun training artificial intelligence to act as a fast, rough draft of these models, offering instant predictions about grid stability. However, these AI tools are not perfect. They can make silent mistakes, especially when the grid faces unusual stress or changes in its structure that the AI has never seen before. The critical question for engineers is not just whether the AI is usually right, but how to know exactly when to trust it and when to stop and ask a slower, more rigorous computer program to double-check the work.

A team of researchers at the Dongguan University of Technology has developed a new method to answer this question, creating a safety net that decides when to let an AI prediction pass and when to send a case for a full, traditional review. They call this system a "solver-referral audit." Imagine the AI as a fast, experienced foreman who can spot most problems instantly, but who might miss a subtle crack in a specific type of beam. The new system acts as a second pair of eyes that checks the foreman's work before a final decision is made. It does not try to replace the slow, perfect math entirely; instead, it creates a transparent process where the AI's guess, a quick physical check of the numbers, and a specific type of safety test are kept separate. This separation ensures that the fast AI is never mistaken for the slow, perfect math, and that every time the system decides to call in the experts, there is a clear record of why.

The researchers tested this idea on three different electrical grid models, ranging from 500 to nearly 800 connection points, simulating hundreds of stressful situations. They created scenarios where the grid was pushed to its limits: loads were increased, wires were removed to simulate outages, and safety limits were tightened. In total, they generated 843 distinct situations to see how the AI would perform under pressure. The core of their new method is a "credibility score," a single number calculated before any heavy-duty computer program is run. This score is built from six different clues. It looks at how much the AI's prediction differs from the normal state, checks if the basic laws of physics hold up in the prediction, measures how much the grid's structure has changed, and evaluates how tight the safety constraints are. By combining these six clues into one score, the system can determine if a specific situation is too risky to trust the AI alone.

When the researchers applied this method, they found that it worked remarkably well at catching the cases where the AI would have failed. In their tests, the system successfully identified every single scenario where the slower, rigorous math program said the grid was in trouble. This is a crucial result because missing a dangerous situation is far worse than checking a safe one unnecessarily. However, the system also learned to be efficient. While it flagged many safe scenarios for extra checking, it did so in a way that significantly reduced the number of dangerous situations that were missed compared to older methods that only looked at how severe a stress was. For example, on one of the test grids, the new method reduced the number of missed dangerous cases by nearly 80 percent compared to a simpler approach that only looked at the severity of the stress. This means that engineers could rely on the fast AI for the vast majority of routine checks, knowing that the new system would reliably catch the rare, tricky cases that need human or high-powered computer attention.

The study also revealed important limits to what the AI and the new system can claim. The researchers were careful to define exactly what their "safety test" was measuring. They used a simplified version of the math that only looks at the flow of active power, ignoring the more complex interactions of voltage and reactive power that occur in the real world. Because of this, the system does not promise that the grid is perfectly safe in every physical sense; it only promises that the grid is safe according to the specific, simplified rules it was tested against. This distinction is vital. The system is not a magic shield that guarantees the lights will never go out; it is a disciplined protocol that keeps the fast AI's predictions separate from the rigorous safety checks. It ensures that when a decision is made to trust the AI, it is a conscious choice based on a clear set of evidence, and when a decision is made to call in the experts, it is done with a specific, reproducible reason.

Ultimately, the value of this work lies in its transparency and its practicality. The researchers did not just build a better AI; they built a better way to use the AI. By keeping the different layers of evidence—the AI's guess, the physical check, and the safety test—separate and accountable, they created a workflow that engineers can trust. The system allows for fast decision-making without sacrificing safety, provided that the operators understand the boundaries of the test. The results suggest that this approach of combining multiple simple checks into a single credibility score is a powerful tool for managing the complex, high-stakes environment of modern power grids. It offers a path forward where artificial intelligence can be deployed at scale, not by pretending to be perfect, but by knowing exactly when to step aside and let the proven methods take over.