A Novel Gain Modeling Technique for LLC Resonant Converters based on The Hybrid Deep-Learning/GMDH Neural Network
Parham Mohammadi
TL;DR
This work tackles LLC resonant converter gain estimation across wide operating conditions by integrating deep learning with GMDH. A deep neural network is first trained on FPGA-based real-time simulator data to capture nonlinear behavior, and its predictions are used to train a GMDH network that delivers a compact algebraic VKG polynomial, aided by an auxiliary feature to simplify terms. The approach achieves high accuracy and outperforms traditional FHA and some time-domain methods while providing a design-friendly, easily implementable model for practitioners. The combination enables accurate, data-driven gain estimation with a simple algebraic form suitable for design and control applications in power electronics.
Abstract
This paper presents a novel hybrid approach for modeling the voltage gain of LLC resonant converters by combining deep-learning neural networks with the polynomial based Group Method of Data Handling (GMDH). While deep learning offers high accuracy in predicting nonlinear converter behavior, it produces complex network models. GMDH neural networks, in contrast, yield simpler algebraic equations that can be more convenient in converter design. By training a deep network on data from an FPGA based real time simulator and then using the network s predictions to train a GMDH model, the proposed hybrid method achieves both high accuracy and design friendly simplicity. Experimental results show significant improvements over traditional methods such as First Harmonic Approximation (FHA) and frequency domain corrections, particularly for wide operating ranges.