Z-SpecNNet: A Real-Time Embedded NN-Based Parameters Estimation for WPT Systems

This article proposes a fully embedded real-time mutual inductance and load estimation technique (Z-SpecNNet) applicable to wireless power transfer (WPT) systems. The technique consists of a three-step process, starting with online noise injection from the supplying converter to excite the system over a wide bandwidth. During the noise injection, the voltage and current at the converter output are recorded, allowing the system impedance to be calculated by fast Fourier transform. Finally, a neural network computes an estimate of the desired parameters. In this work, the Z-SpecNNet is applied to a series--series compensated system as it is one of the most popular compensation topologies in the literature and because it is the topology for which the information of load and mutual coupling result most correlated and therefore more difficult to estimate. The proposed Z-SpecNNet offers significant advantages because impedance spectroscopy is a straightforward and model-free method for characterizing system behavior. Furthermore, the neural network can be rapidly trained on a known transfer function. The technique has been demonstrated to be effective on a low-cost microcontroller that integrates the control of the converter. Experimental results indicate a mean relative estimation error of 7.81% with a total estimation time of 85 ms.