Reinforcement Learning-Based Real-Time Control of a Photovoltaic-Integrated Welding Power Supply: Experimental Validation and Performance Benchmarking

Imagine a factory floor where heavy metal is joined together by a searing, electric arc. This process, known as welding, is a cornerstone of modern manufacturing, but it is also a voracious consumer of electricity. For decades, the machines that create these arcs have relied on steady, predictable power from the electrical grid. However, as industries seek to reduce their carbon footprint, there is a growing desire to power these machines directly from the sun. The challenge lies in the nature of sunlight itself; it is fickle, shifting with the clouds and the time of day. At the same time, the welding arc is a chaotic, unpredictable force that changes its behavior the moment the distance between the tool and the metal shifts even slightly. Trying to feed a machine that demands instant, precise power with a source that is constantly fluctuating is like trying to fill a bucket with a hose that changes its pressure every second while the bucket itself is moving.

Researchers have long relied on standard electronic controllers to manage this balance, but these traditional systems often struggle when faced with such complex, changing conditions. They are like a driver who keeps their foot on the gas at a fixed rate, unable to react quickly enough when the road suddenly turns or the engine sputters. To solve this, a team of engineers has developed a new approach that combines the reliability of established electronics with the adaptability of artificial intelligence. Their work focuses on a specific type of machine: a high-frequency welding power supply that runs directly on solar energy. By teaching the machine's control system to learn from its own environment, they have created a device that can adjust itself in real-time, maintaining a perfect weld even as the sun hides behind a cloud or the welder's hand moves.

The researchers built a physical prototype of this system to test if their ideas would work in the real world, not just in computer simulations. The heart of their design is a two-layered control system. The first layer is a fast-acting electronic regulator that handles the immediate, split-second demands of the welding arc, ensuring the current remains steady. This layer operates with the speed and predictability of a traditional machine. The second layer is where the innovation lies: an artificial intelligence agent that watches the bigger picture. This agent monitors the available power from the solar panels and the changing length of the welding arc. Instead of following a rigid set of rules, it learns how to adjust the settings of the fast regulator to match the current conditions. It is a system that adapts its own sensitivity, tightening or loosening its grip on the power flow depending on what is happening at that exact moment.

To prove this concept, the team constructed a working model using advanced silicon carbide components, which are known for their ability to handle high speeds and heat efficiently. They tested this machine in two ways: first on a smaller laboratory bench with simulated conditions, and then on a full-scale, five-kilowatt industrial test bed. During these tests, they subjected the machine to variable arc lengths, sudden bursts of power demand, and fluctuating solar input that mimicked the passing of clouds. The results were measured against three other common methods: a system that uses a battery to smooth out the power, a standard controller with fixed settings, and a system that follows a basic set of pre-programmed rules.

The findings showed that the new artificial intelligence-driven system outperformed the others in efficiency, dynamic response time, and energy consumption per weld cycle. It achieved an energy efficiency of approximately 90 percent, meaning very little power was wasted as heat. In contrast, the battery-buffered system was less efficient due to the energy lost in charging and discharging the battery. The new system also responded to changes in the welding process about 33 percent faster than the standard controller, allowing for a smoother, more consistent weld. Regarding power quality, the system maintained a total harmonic distortion (THD) of approximately 4 percent, which meets industrial safety standards. However, the researchers noted a nuanced trade-off: while the system achieved excellent overall performance, the adaptive nature of the learning policy introduced a minor increase in switching-related harmonic content compared to the fixed-gain PI and battery-buffered baselines in simulation. This suggests that the system prioritizes rapid tracking and stability, accepting a slight compromise in harmonic purity to achieve its superior dynamic response.

Perhaps most importantly, the system proved to be robust and safe. The researchers ran the machine continuously for eight hours under heavy load, and the components remained within safe temperature limits, showing no signs of stress or failure. The artificial intelligence did not just react; it adjusted its behavior based on real-time state feedback to minimize energy consumption while maintaining the integrity of the weld. While the system did introduce a very minor trade-off in the form of slightly more electrical switching activity compared to the simplest controllers, the overall gains in speed, efficiency, and stability were substantial.

This work demonstrates that it is possible to run heavy industrial machinery directly from renewable sources without sacrificing performance or safety. By combining the speed of traditional electronics with the learning capability of artificial intelligence, the researchers have shown a viable path toward sustainable manufacturing. The machine they built does not just survive the unpredictability of solar power and the chaos of the welding arc; it thrives in it, adjusting itself to find the perfect balance every time. This suggests that in the future, factories could rely more heavily on the sun, using smart systems to turn the variable nature of renewable energy into a steady, reliable force for industry.