Citation: Isah, A.; Shin, H.; Oh, S.;
Oh, S.; Aliyu, I.; Um, T.-w.; Kim, J.
Digital Twins Temporal
Dependencies-Based on Time Series
Using Multivariate Long Short-Term
Memory. Electronics 2023, 12, 4187.
https://doi.org/10.3390/
electronics12194187
Academic Editors: Martin Reisslein
and Franco Cicirelli
Received: 31 July 2023
Revised: 5 October 2023
Accepted: 7 October 2023
Published: 9 October 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
electronics
Article
Digital Twins TemporalDependencies-Based on Time Series
Using Multivariate Long Short-Term Memory
Abubakar Isah
1
, Hyeju Shin
1
, Seungmin Oh
1
, Sangwon Oh
1
, Ibrahim Aliyu
1
, Tai-won Um
2,
*
and Jinsul Kim
1,
*
1
Department of ICT Convergence System Engineering, Chonnam National University,
Gwangju 61186, Republic of Korea; abubakarisah@jnu.ac.kr (A.I.); sinhye102@jnu.ac.kr (H.S.);
216655@jnu.ac.kr (S.O.); osw0782@naver.com (S.O.); aliyu@jnu.ac.kr (I.A.)
2
Graduate School of Data Science, Chonnam National University, Gwangju 61186, Republic of Korea
* Correspondence: stwum@jnu.ac.kr (T.-w.U.); jsworld@jnu.ac.kr (J.K.)
Abstract: Digital Twins, which are virtual representations of physical systems mirroring their be-
havior, enable real-time monitoring, analysis, and optimization. Understanding and identifying the
temporal dependencies included in the multivariate time series data that characterize the behavior of
the system are crucial for improving the effectiveness of Digital Twins. Long Short-Term Memory
(LSTM) networks have been used to represent complex temporal dependencies and identify long-
term links in the Industrial Internet of Things (IIoT). This paper proposed a Digital Twin temporal
dependency technique using LSTM to capture the long-term dependencies in IIoT time series data,
estimate the lag between the input and intended output, and handle missing data. Autocorrelation
analysis showed the lagged links between variables, aiding in the discovery of temporal dependen-
cies. The system evaluated the LSTM model by providing it with a set of previous observations and
asking it to forecast the value at future time steps. We conducted a comparison between our model
and six baseline models, utilizing both the Smart Water Treatment (SWaT) and Building Automation
Transaction (BATADAL) datasets. Our model’s effectiveness in capturing temporal dependencies
was assessed through the analysis of the Autocorrelation Function (ACF) and Partial Autocorrelation
Function (PACF). The results of our experiments demonstrate that our enhanced model achieved a
better long-term prediction performance.
Keywords: temporal dependency; Digital Twins; LSTM; multivariate time series
1. Introduction
Digital Twins connect the real and virtual worlds [1,2], offering simulations, projec-
tions, and insights that can be applied to decision-making, optimization, and maintenance
tasks [3]. The Digital Twin can learn and capture the underlying patterns and dependencies
of the dynamic system by examining the historical and real-time data of the multivariate
time series and training an LSTM network with this data [4]. This enables it to generate
precise simulations and predictions. Digital Twin is a technology that is still under devel-
opment, but it has the potential to revolutionize the way we manage assets. Digital Twins
are virtual representations of physical assets that can be used to simulate, monitor, and
optimize the performance of those assets [5].
Temporal dependencies are the patterns and relationships that develop over time
between the variables in multivariate time series data. These dependencies can include
seasonality, trends, lagged relationships, sequential patterns, and other temporal structures.
Temporal dependencies and multivariate time series analysis are crucial in many areas,
including weather forecasting, industrial operations, and others [6]. To make effective
decisions and maximize system performance, it is essential to have the ability to accurately
analyze and forecast [7] the behavior of complex dynamics. Time series data can be utilized
Electronics 2023, 12, 4187. https://doi.org/10.3390/electronics12194187 https://www.mdpi.com/journal/electronics