Large Language Model-Assisted Zoning for Heterogeneous Automatic Voltage Control in Renewable-Rich Regional Power Grids

The modern electrical grid is no longer just a one-way street where power flows steadily from massive coal or gas plants to homes and factories. Today, it is a dynamic, two-way network increasingly fed by wind turbines and solar panels. These renewable sources are clean and essential, but they are also fickle; the wind stops blowing and clouds pass over solar fields, causing the amount of electricity they generate to fluctuate wildly. This unpredictability creates a specific challenge for grid operators: keeping the voltage stable. Voltage is the electrical pressure that pushes current through wires; if it gets too high, it can damage equipment, and if it drops too low, lights dim and motors stall. In a traditional grid, operators could rely on large, steady power plants to adjust their output and smooth out these bumps. But with thousands of scattered wind and solar farms, the grid becomes a complex web where too much or too little power in one area can cause voltage to swing dangerously in another. Managing this requires a system that can instantly decide which power sources should adjust their output and by how much, a task that has become too difficult for standard, uniform rules to handle alone.

In a recent study, researchers from State Grid Shanxi Electric Power Company in China proposed a new way to manage this complexity by treating different parts of the grid differently, guided by an artificial intelligence tool known as a large language model. Instead of forcing every solar farm and wind turbine to follow the exact same set of instructions, the team developed a system that first looks at the local conditions of the grid to decide which strategy works best for each specific area. They divided the grid into zones based on how the power sources interact with the local voltage. In some areas, where the voltage is naturally stable, the system simply asks the power plants to maintain a steady, preset amount of reactive power—a type of electrical support that helps maintain voltage without actually delivering energy to homes. In other areas where voltage swings are more noticeable, the system switches to a mode where the plants actively track and correct the voltage, much like a thermostat adjusting a heater to keep a room at a set temperature.

The most complex part of the grid, however, involves areas where many different power plants are tightly connected and influence each other strongly. Here, if every plant tried to fix the voltage on its own, they might end up working against one another, creating a chaotic tug-of-war that wastes energy and stresses the equipment. To solve this, the researchers introduced a sophisticated coordination method using a type of artificial intelligence called multi-agent deep deterministic policy gradient. In this setup, a central controller determines the total amount of voltage support needed for the whole group, and the AI then acts as a fair referee, deciding exactly how much work each individual plant should do to meet that goal. This ensures that the burden is shared evenly and that no single plant is overworked while others sit idle.

To make these decisions, the researchers used a large language model not to generate the final commands, but to act as a high-level planner. This AI tool analyzed the grid's current state, including the output of the wind and solar farms, the layout of the power lines, and the history of voltage fluctuations. Based on this data, it recommended how to group the power plants, which control strategy to use for each group, and what initial settings to apply. If the system's initial plan failed a safety check—meaning the voltage still drifted too far from the safe range—the AI would review the results, learn from the mistake, and suggest adjustments to the grouping or the settings. This cycle of planning, testing, and refining allowed the system to adapt to changing conditions without human intervention.

The team tested this approach on a simulated version of a real power grid, modeled after a standard test system with thirty-nine connection points, populated with eight wind farms and eight solar plants. They compared their new, flexible method against older strategies that applied a single rule to the entire grid. The results showed that the new method was significantly more effective at keeping voltage stable. It reduced the number of times the voltage fell outside the safe range by nearly 88 percent compared to having no special control at all, and it cut those violations by more than 40 percent compared to the best existing uniform method. Furthermore, the new system balanced the workload much better; in the older methods, some plants were often forced to make large, abrupt changes while others did nothing, leading to uneven wear and tear. The new approach smoothed out these adjustments, reducing the size of the changes each plant had to make by roughly 30 percent and eliminating the instances where plants would accidentally cancel each other's efforts.

A key finding of the study was that the large language model's ability to refine its own suggestions made a measurable difference. When the researchers let the AI adjust the importance it placed on different goals—such as prioritizing voltage stability versus balancing the workload among plants—the system performed even better. In the most difficult, tightly connected zones, the AI-optimized system reduced voltage errors by nearly 19 percent and the imbalance in workload by over 70 percent compared to a standard control method. This suggests that the future of grid management may not lie in a single, rigid set of rules, but in intelligent systems that can look at the unique characteristics of a local neighborhood of power plants and decide on the best way to keep the lights on. The study confirms that by combining human-designed control logic with the adaptive reasoning of large language models, it is possible to create a more resilient and efficient grid capable of handling the high levels of renewable energy required for a low-carbon future.