Dynamic Data-Driven Multi-Step-Ahead Prediction with Simulation Data and Sensor Measurements

Abstract

In this paper, the authors propose a dynamical data-driven prediction framework to estimate a system’s behavior multiple steps ahead. The main contribution of this work is in the construction of a new recursive multi-input/multi-output (REC-MIMO) strategy, which can attain real-time estimation through a combination of expensive simulation data (offline) and sensor measurements (online). This framework is composed of an offline phase and an online phase. In the offline phase, a multi-input/multi-output strategy is employed for constructing a predictive model for the system’s dynamical responses from simulation data. The offline phase is accomplished through a combination of proper orthogonal decomposition and Gaussian process. By using the measurement data collected from sensors, a recursive strategy is used to iteratively enhance multi-step-ahead predictions by using a new reduced-order particle filter method during the online phase. An illustrative example involving aeroelastic responses of a joined-wing SensorCraft is used to demonstrate the effectiveness of the proposed framework. It has been shown that state predictions obtained through the REC-MIMO strategy have accuracy comparable to those obtained from high-fidelity simulations with significant reduction in computational expense. It is envisioned that the proposed methodology can be used to support system decision making in an efficient manner.