Real-Time Adaptive MPC via Data-Driven Controller Learning on FPGA

This paper presents a performance-driven adaptive Model Predictive Control (MPC) framework tailored for high-frequency power converters. By integrating offline optimization with supervised learning, our approach automatically tunes the MPC parameters to optimize closed-loop performance metrics—specifically, settling time and overshoot. A lightweight, quantization-aware multi-layer perceptron (MLP) is trained to map operating conditions and user-defined performance weights to the optimal weighting matrices in quadratic loss function, enabling rapid, on-the-fly parameter adaptation without incurring online training overhead. The entire framework is implemented on an FPGA using High-Level Synthesis, achieving sub-microsecond inference latency necessary for GaN-based converters operating at MHz -level switching frequencies. Experimental validation on a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$12 ~\mathrm{V}-3.3 ~\mathrm{V}, 1 \text{MHz}$</tex> Buck converter demonstrates reduction in voltage overshoot and improvement in settling time compared to fixed-parameter MPC, while maintaining robust performance under varying load conditions. This work effectively bridges model-based control and data-driven learning, offering a promising solution for real-time adaptive control in high-performance power electronics.