Multi Step Neural Network-Based System Identification and NMPC for an Unmanned Surface Vehicle

This paper investigates the application of Neural Network (NN)-based system identification for Nonlinear Model Predictive Control (NMPC) of an underactuated Unmanned Surface Vehicle (USV). We focus on practical NMPC implementation by employing neural networks to approximate system dynamics, specifically using a multi-step ahead prediction in a single inference pass. We compare a multi-step MLP predictor and a multi-step LSTM predictor against a conventional recursive single-step prediction approach. Simulations, using data from a Clearpath Robotics Heron USV, demonstrate that multi-step predictors achieve comparable trajectory tracking accuracy to the single-step approach. Critically, the multi-step predictors reduce the computational load of calculating the gradients of the optimal cost function with respect to the control input trajectory by up to an order of magnitude. This significant reduction in computational burden facilitates real-time NMPC implementation and enhances its deployment in dynamic and challenging environments.