Impact of System Initial Values on Multivariate LSTM Neural Network Performance: A Finding From a Control Process Perspective

Nowadays, artificial intelligence (AI) has made significant contributions to the identification and modeling of physical systems across a wide range of scientific disciplines. However, integrating domain knowledge into an AI model has always been imperative in order to ensure a more accurate representation and capture all the pertinent system information. In this context, the initial value of a physical system can have a significant impact on the dynamic behavior of that system, particularly from a control design perspective, given its effect on transient behavior and system stability. The present study examines how the system's initial values can affect the output forecasting of the system by employing a Long-Short- Term Memory Recurrent Neural Network (LSTM-RNN). In this regard, a novel architecture design and data acquisition framework have been presented with the aim of providing a precise AI model to address the challenges related to the control design. The inclusion of the initial value as an input to the LSTM architecture is intended to exhibit how it potentially impacts the network's ability to capture long-term dependencies and make accurate predictions. Based on the study findings, considering the initial values of the system to improve the effectiveness of LSTM-RNN, ultimately leads to superior accuracy and predictive capabilities network.