Machine learning enables stable wireless power transfer at 86.7%

A research team from Chiba University in Japan has developed a machine learning-based design method for wireless power transfer (WPT) systems that remains efficient and stable regardless of changing load conditions.

WPT systems transmit electrical energy without the need for physical wires or connectors, instead relying on electromagnetic fields. The concept, dating back to the 1890s and famously experimented with by Nikola Tesla, is now central to a variety of modern consumer devices, such as smartphones, electric toothbrushes, and Internet of Things sensors.

One of the major engineering challenges for WPT systems has been voltage instability and decreased efficiency when the load varies, an issue that has limited practical implementation and reliability. Historically, achieving load-independent (LI) operation required precise tuning of circuit components, such as inductors and capacitors, calculated by complex analytical equations. These equations were typically based on idealised conditions and did not fully reflect actual device behaviour or diverse operating environments.

To address these shortcomings, Professor Hiroo Sekiya and colleagues from the Graduate School of Informatics at Chiba University have introduced a fully numerical, machine learning–based optimisation method for the design of WPT circuits. The study, conducted with Mr. Naoki Fukuda and Dr. Yutaro Komiyama from Chiba University, Dr. Wenqi Zhu from Tokyo University of Science, and Dr. Akihiro Konishi from Sojo University, was published in IEEE Transactions on Circuits and Systems I: Regular Papers.

In this approach, the team's method involves describing the WPT circuit using differential equations that account for the actual characteristics and behaviour of each component, including non-idealities. These equations are solved numerically across time until the system reaches a stable operating point. An objective function evaluates the performance based on voltage stability, efficiency, and harmonic distortion. A genetic algorithm then iteratively updates the system parameters, seeking to optimise the overall score and achieve the target of load-independent operation.

"We established a novel design procedure for a LI-WPT system that achieves a constant output voltage without control against load variations. We believe that load independence is a key technology for the social implementation of WPT systems. Additionally, this is the first success of a fully numerical design based on machine learning in the field of power electronics research," said Prof. Sekiya.

The researchers demonstrated their method on a class-EF WPT topology, which comprises a class-EF inverter and class-D rectifier. Traditionally, a class-EF inverter without load independence maintains zero-voltage switching (ZVS) only at its rated load. If the load deviates from this point, ZVS is disrupted and system efficiency decreases. The team's design, by contrast, preserved ZVS and stable output voltage even as loads were varied.

Experimental results highlighted that, while conventional LI inverters could experience output voltage variations as high as 18% under changing loads, the new fully numerical method stabilised this variation to less than 5%. The team also observed improved performance at lighter loads, attributed to more accurate modelling of diode parasitic capacitance. Power-loss analysis showed nearly consistent energy dissipation in the transmission coil across different loads, indicating effective current regulation by the load-independent design.

At its rated operating point, the LI class-EF WPT system achieved a high power-delivery efficiency of 86.7% at 6.78 MHz, delivering more than 23 watts of output power. This suggests the method stands to improve both the reliability and applicability of WPT for a range of devices and use cases.

Looking to future applications, the researchers expect their findings to have implications beyond just WPT systems. Prof. Sekiya commented, "We are confident that the results of this research are a significant step toward a fully wireless society. Moreover, due to LI operation, the WPT system can be constructed in a simple manner, thereby reducing the cost and size. Our goal is to make WPT commonplace within the next 5 to 10 years."

He also highlighted the broader significance of the work, noting that the successful application of machine learning and artificial intelligence to power electronics design could herald a move toward automated circuit design processes in the field.