Ultrafast embedded explicit model predictive control for nonlinear systems

Abstract

The design of energy-efficient and ultrafast nonlinear model predictive controllers (NMPCs) is critical for decision-making in modern engineering systems. To this end, an embedded systems approach is proposed for hardware acceleration of a stabilizing explicit NMPC (ENMPC). Tools from approximate computing are employed to simplify the ENMPC control law and design an ultra-fast, low-power, miniaturized ASIC (application specific integrated circuit) deploying the control mechanism. Approximation bounds on the embedded controller and stability guarantees of the closed-loop system are provided. The efficacy and energy-savings of the embedded ENMPC is verified in an ASIC-in-the-loop simulation experiment. Whereas the exact ENMPC law requires 79K gates for implementation, consumes 13.66 mW of power, and operates at 0.3 GHz on 45 nm Nangate technology, the approximating ASIC requires only 3.6K gates (resulting in a 25× area reduction), consumes a meager 0.47 mW of power (29× power reduction), and runs at 0.5 GHz (more than 105× faster than cutting-edge embedded NMPCs).