A research team at Chiba University in Japan has developed a machine learning-based design method for efficient and stable wireless power transfer (WPT) systems regardless of changes in load conditions.
Instead of relying on electromagnetic fields, WPT systems transmit electrical energy without the need for physical wires or connectors. Back in the 1890s, Nikola Tesla's famous concept is now at the heart of a wide range of modern consumer devices, including smartphones, electric toothbrushes and Internet of Things sensors.
One of the major engineering challenges of WPT systems is voltage instability and reduced efficiency in various loads, and problems with limited practical implementation and reliability. Historically, achieving load-independent (LI) operations required accurate adjustment of circuit components such as inductors and capacitors calculated by complex analytical equations. These equations were usually based on idealized conditions and did not fully reflect the actual device's operation or the diverse operating environment.
To address these drawbacks, Professor Hirao Sekiya, Graduate School of Information Studies at Chiba University, and colleagues, introduced a completely numerical machine learning-based optimization method for the design of WPT circuits. This study was conducted with Fukuda and Dr. Komiyamo from Wayama University, Dr. Wenki Chu from Tokyo University, and Dr. Kojima from Soho University, and was published in IEEE Transaction I on Circuits and Systems: Official Paper.
In this approach, the team's method involves writing a WPT circuit using differential equations that explain the actual properties and behavior of each component, including non-ideals. These equations are resolved numerically over time until the system reaches a stable operating point. The objective function evaluates performance based on voltage stability, efficiency, and harmonic distortion. The genetic algorithm then repeatedly updates the system parameters, optimizing the overall score, and attempts to achieve a target for load-independent operations.
“We have established a new design procedure for LI-WPT systems that achieve constant output voltage without control over load variation. We consider load independence to be an important technique for social implementation of WPT systems. Furthermore, this is the first success of a complete numerical design based on machine learning in the field of power electronics research.
The researchers have demonstrated a method for a Class-EF WPT topology consisting of a Class-EF inverter and a Class-D rectifier. Traditionally, non-load independence class-EF inverters maintain zero voltage switching (ZVS) only at rated loads. If the load deviates from this point, ZVS will be destroyed and the system will be reduced. In contrast, the team's design preserved ZV and stable output voltage even when the load was changed.
Experimental results highlight that traditional LI inverters can experience 18% output voltage fluctuations under load changes, but the new complete numerical method stabilizes this fluctuation below 5%. The team also observed improved performance at light loads due to more accurate modeling of the diode parasitic capacitance. Power loss analysis showed almost consistent energy dissipation of the transmission coil across a variety of loads, and effective current regulation with load-independent designs.
At a rated operating point, the LI Class-EF WPT system achieved high output efficiency of 86.7% at 6.78 MHz, providing output power of over 23 watts. This suggests that this approach improves both the reliability and applicability of WPTs for a variety of devices and use cases.
Aiming for future applications, researchers hope that their findings will have an impact beyond just WPT systems. Professor Sekiya said, “We believe that the results of this study are an important step into a fully wireless society. Furthermore, with the operation of LI, WPT systems can be built in a simple way, thereby making WPT common within the next five to ten years.”
He also highlighted the broader importance of the work, and said that machine learning and artificial intelligence successfully designed power electronics can signal the move towards the automated circuit design process in this field.
