Combination of Elman Neural Network and Kalman Network for Modeling of Batch Distillation Process
Adi Novitarini Putri, Carmadi Machbub, Egi Hidayat
Abstract
Batch distillation is one of the nonlinear systems in a real-world application. Before designing and implementing a control system, a model is needed to capture its dynamics. The batch distillation process could be represented in a state-space model. Nonetheless, the sensor is limited so it could not access its states directly from the sensor. Then, the state estimation method is needed to estimate its state trajectories. On the other hand, state estimation is one of the fundamental problems in signal processing. Kalman Filter (KF) is an optimal filtering algorithm and the most widely used in signal processing. Meanwhile, deep learning is the most active and present topic nowadays on modeling and estimation. We use a recurrent neural network scheme to develop the Nonlinear State Space (NLSS) model. In this research, we use KalmanNet to overcome the state estimation problem if the noise statistic matrices are unknown. This idea has been done using a Long Short-Term Memory (LSTM) scheme which produces the Kalman Gain value. In this study, we compare conventional Kalman gain value and KalmanNet to find which one is better to solve the state estimation problem. From the experiment, the combination of the Elman Neural Network (ENN) model and KalmanNet which have the best performance from its Mean-Squared Error (MSE) value, 1.30E03.