Abstract

Traditionally, the real-time implementation of Model Predictive Control (MPC) has been limited by processing and storage requirements. Recently, the idea of using Artificial Neural Networks (ANN) to approximate MPC control laws, including implementations on Field Programmable Gate Array (FPGA), has been explored. This work presents a complete design flow from software controller to hardware implementation, utilizing Keras and QKeras for ANN design and quantization and HLS4ML with the AMD-Xilinx Design Suite for FPGA implementation. The evaluation and analysis conducted provides insights into the trade-offs involved in the proposed workflow. Experimental results are validated on a PYNQ-Z1 board, achieving latencies of less than one microsecond in the case study, demonstrating a hardware precision comparable to traditional MPC methods.