Forecasting container throughput with long short-term memory networks
| Date | 04 December 2019 |
| Published date | 04 December 2019 |
| Pages | 425-441 |
| DOI | https://doi.org/10.1108/IMDS-07-2019-0370 |
| Author | Sonali Shankar,P. Vigneswara Ilavarasan,Sushil Punia,Surya Prakash Singh |
| Subject Matter | Information & knowledge management,Information systems,Data management systems,Knowledge management,Knowledge sharing,Management science & operations,Supply chain management,Supply chain information systems,Logistics,Quality management/systems |
Forecasting container
throughput with long short-term
memory networks
Sonali Shankar
Bharti School of Telecom Technology & Management, New Delhi, India, and
P. Vigneswara Ilavarasan, Sushil Punia and Surya Prakash Singh
Department of Management Studies,
Indian Institute of Technology, New Delhi, India
Abstract
Purpose –Better forecasting always leads to better management and planning of the operations. The
container throughput data are complex and often have multiple seasonality. This makes it difficult toforecast
accurately. The purpose of this paper is to forecast container throughput using deep learning methods and
benchmark its performance over other traditional time-series methods.
Design/methodology/approach –In this study, long short-term memory (LSTM) networks are
implemented to forecast container throughput. The container throughput data of the Port of Singapore are
used for empirical analysis. The forecasting performance of the LSTM model is compared with seven
different time-series forecasting methods, namely, autoregressive integrated moving average (ARIMA),
simple exponential smoothing, Holt–Winter’s, error-trend-seasonality, trigonometric regressors (TBATS),
neural network (NN) and ARIMA +NN. The relative error matrix is used to analyze the performance of the
different models with respect to bias, accuracy and uncertainty.
Findings –The results showed that LSTM outperformed all other benchmark methods. From a statistical
perspective, the Diebold–Mariano test is also conducted to further substantiate better forecasting
performance of LSTM over other counterpart methods.
Originality/value –The proposed study is a contribution to the literature on the container throughput
forecasting and adds value to the supply chain theory of forecasting. Second, this study explained the
architecture of the deep-learning-based LSTM method and discussed in detail the steps to implement it.
Keywords Logistics, Time series, Forecasting, Deep learning, LSTM, Maritime supply chain
Paper type Research paper
1. Introduction
The container logistics is a billion-dollar industry and a small improvement at strategical,
tactical or operational decision levels will render huge cost benefits. The research on
container throughput forecasting has evolved with time in terms of using more
sophisticated techniques and methods to reduce forecasting errors. The efficient use of port
data and its management directly influence the economic development of the port-hinterland
region. The number of managerial decisions such as infrastructure-related investments at
the port, container inventory management, equipment planning, port layout design and
schedules are based on a forecast. Therefore, the forecaster is always looking for better and
more accurate forecasting methods. The container throughput forecasting models can be
broadly classified as econometric models; hybrid models and machine-learning-based
models. The econometric models include autoregressive integrated moving average
(ARIMA), simple exponential smoothing (SES), Holt–Winter’s (HW) and so on. Hybrid
models, on the other hand, essentially include either different econometric models or an
econometric model with a machine-learning-based method such as ARIMA +NN. The
machine-learning-based models include an artificial neural network (ANN) and deep
learning methods. Although machine learning methods have gained a lot of popularity in
the past decade, they are still underexplored in the area of the maritime supply chain.
Furthermore, very little literature is available on the potential of deep-learning-based
Industrial Management & Data
Systems
Vol. 120 No. 3, 2020
pp. 425-441
© Emerald PublishingLimited
0263-5577
DOI 10.1108/IMDS-07-2019-0370
Received 10 July 2019
Revised 21 October 2019
Accepted 10 November 2019
The current issue and full text archive of this journal is available on Emerald Insight at:
www.emeraldinsight.com/0263-5577.htm
425
Forecasting
container
throughput
methods to predict container throughput. This study aims to bridge this gap and improve
the forecasting accuracy of the container throughput with the implementation of the
deep-learning-based method, long short-term memory (LSTM).
For empirical analysis, the time-series data of monthly container throughput at the Port
of Singapore Authority (PSA) are selected. The monthly data are divided such that the
planning horizon represents short-, medium- and long-term forecasts. The performance of
LSTM is compared with seven other widely used time-series benchmark methods with the
help of an error matrix. To further substantiate the results, the performance of the models is
tested through a statistical test called Diebold and Mariano (1995) (DM) test. Because the
LSTM network is a bit difficult to comprehend initially, for a better understanding of
the readers, the implementation of the LSTM method is provided, as well as based on its
performance, the important findings are discussed.
The remainder of the paper constitutes five parts where Section 2 summarizes the
literature on various models used for container throughput forecasting. Section 3 briefly
describes the different univariate time-series forecasting methods. The data statistics and
the LSTM method are explained in detail in Section 4. The results and performance
analysis of LSTM over benchmark methods through error matrix are described in Section
5, followed by discussion. Finally, Section 6 concludes the study and proposes future
research avenues.
2. Review of related literature
2.1 Container throughput forecasting models
As early as pre-2003, the container throughput was forecasted using logit models (Veldman
and Báckmann, 2003), error correction models (Fung, 2002), vector error correction model
(Fung, 2001) and by derivingthe formula using diversion distanceand transshipment volume
(Zohil and Prijon, 1999). Later, the moving average (MA) and decomposition models became
popular (Schulzea and Prinzb, 2009). The container throughput elasticity is measured for
the Hamburg–Le Havre ports with the help of an autoregressive distributed lag model
(Rashed et al., 2018). The spatiotemporal analysis of the port of Ningbo-Zhoushan , China, is
carried out and the container throughput is predicted for 2026 by means of the ARIMA model
(Feng et al., 2019). The portfolio of different time-series forecasting methods is provided by
Chan et al. (2018). The different model selection and model averaging methods are compared
by Gao et al. (2016)and it is concluded that model averagingmethods performed better among
both. The averaging method like seasonal ARIMA (SARIMA) when compared with
Holt–Winter’s exponential smoothing (ES)produces similar results (Ee et al., 2014). However,
when both the approaches were applied to predict the transshipment container volume in
Germany, SARIMA performed better than Holt–Winter’s ES (Schulzea and Prinzb, 2009).
On the other hand, Diazet al. (2011) observed Winter’s methodto be more accurate. Therefore,
for the sake of comparison, both SARIMA and Holt–Winter’s methods are considered in this
paper. Although the SARIMA model is widely used in the singleand hybrid model forecasts,
Chen and Chen (2010) observed that genetic programming models predict 32–36 percent
better than SARIMA. Furthermore, even meta-heuristics like discrete particle swarm
optimization are also used to forecast the container throughput at port (Xiao et al., 2014).
Because this study focuses on the state-of-the-art sequential-learning-based time-series
models, the meta-heuristics are not discussed here.
Several econometric and machine learning methods are hybridized to improve the
accuracy of the container throughput forecast. The forecasting accuracy and stability is
found to be improved by hybrid decomposition ensemble where the low-frequency
components of the series are predicted by ARIMA model, whereas support vector regression
is used to predict the high-frequency components (Niu et al., 2018). Similarly, SARIMA and
least square support vector regression are used (Xie et al., 2013), but ARIMA and ANN are
426
IMDS
120,3
To continue reading
Request your trial