ANN-Assisted Intelligent Control of a High-Gain Non-Isolated DC–DC Converter for PEM Fuel Cell Electric Vehicles

Electric vehicles are often celebrated for their quiet hum and zero tailpipe emissions, but the heart of many next-generation designs is a chemical battery that does not store electricity at all. Instead, it generates power on demand by combining hydrogen with oxygen, a process that produces only water as a byproduct. This technology, known as a proton exchange membrane fuel cell, offers the promise of quick refueling and long driving ranges, yet it comes with a fundamental electrical quirk. The voltage it produces is low and wobbly, fluctuating like a stream that rises and falls with the weather. To power the heavy motors that drive a car, this weak, unstable current must be boosted into a strong, steady stream of electricity. Engineers solve this mismatch using a device called a DC–DC converter, which acts as a voltage elevator, stepping up the low fuel cell voltage to the high levels needed by the vehicle's motor. However, making this elevator work smoothly is difficult because the fuel cell's output changes constantly, and the car's demand for power shifts instantly as the driver accelerates or climbs a hill.

In a new study, researchers at Vellore Institute of Technology in India have developed a smarter way to manage this voltage elevator for fuel cell vehicles. They focused on a specific type of converter that can raise the voltage significantly without needing bulky, heavy transformers. While the physical design of this converter was already known, the challenge lay in controlling it. Traditional controllers act like a steady hand on a thermostat, constantly checking the temperature and making small adjustments to keep it stable. This works well when conditions are predictable, but it can be slow to react when the fuel cell voltage drops suddenly or the load changes sharply. The researchers proposed a hybrid approach that combines this traditional steady hand with a digital brain capable of predicting what to do before a problem even happens. They trained an artificial neural network, a type of computer program modeled after the human brain, to learn the relationship between the fuel cell's input, the desired output, and the load. Instead of waiting for an error to occur and then correcting it, this digital brain calculates the exact setting needed to maintain the correct voltage based on the current conditions, while the traditional controller remains in the background to catch any small mistakes the prediction might miss.

The team did not build a physical prototype for this specific test; instead, they created a highly detailed mathematical model of the converter and used it to generate thousands of data points representing how the system behaves under different conditions. They fed this data into the artificial neural network, teaching it to recognize patterns and predict the precise control signal required for any given situation. The result was a system that learned the converter's behavior with remarkable precision, achieving a level of accuracy that suggests it could almost perfectly mimic the ideal control signal. When the researchers tested this new hybrid system in a computer simulation, the results were striking. The system reached the target voltage of 400 volts in just 5.2 milliseconds, which is less than half the time it took for the traditional controller to do the same. It also settled into a stable state much faster, with almost no overshoot, meaning the voltage did not spike dangerously high before settling down. The traditional controller, by comparison, took more than twice as long to stabilize and allowed the voltage to fluctuate more widely before finding its balance.

This improvement matters because electric vehicles need power electronics that can react instantly to the driver's commands. A slow response can lead to sluggish acceleration or instability in the power system, while a system that overshoots can damage sensitive components. The study shows that by letting the artificial neural network handle the heavy lifting of prediction and leaving the traditional controller to handle fine-tuning, the system becomes both faster and more stable. The researchers noted that the artificial neural network was trained using data generated directly from the mathematical model of the converter, a method they describe as physics-informed. This means the computer learned the laws of physics governing the device rather than just memorizing random numbers, which helps it make reliable predictions even in situations it has not seen before. The final system maintained the output voltage with an error of less than one-tenth of a percent, a level of precision that is critical for the reliable operation of a vehicle.

While the results are impressive, the authors are careful to note that these findings come from computer simulations. The next step, which they plan to undertake, is to build a physical version of this controller and test it on real hardware. They envision installing the system on digital chips that can run the calculations fast enough to keep up with the rapid changes in a moving car. If successful, this approach could make fuel cell vehicles more responsive and efficient, helping to overcome one of the remaining hurdles in making hydrogen-powered transport a practical reality. The work demonstrates that combining old-school engineering control with modern artificial intelligence can solve problems that neither approach could tackle as effectively on its own, offering a path toward more robust and intelligent power systems for the vehicles of the future.