Input-output selection for LSTM-based reduced-order state estimator design

Abstract

In this paper, we present a sensitivity-based approach to construct reduced-order state estimators based on recurrent neural networks (RNNs). Our approach assumes that a mechanistic model is available but too computationally complex for estimator design and that only a subset of the states needs to be estimated. To address this problem, a reduced-order estimator that can estimate the target outputs is sufficient. We propose an approach based on sensitivity analysis to determine the appropriate inputs and outputs for data collection and data-driven model training so that the target outputs can be estimated accurately. In particular, we consider RNN networks as the tool to train the data-driven model and extended Kalman filter (EKF) as the estimator. Simulations are carried out to illustrate the effectiveness of the proposed approach.