Abstract:
Traffic congestion is a significant issue for all social classes in society. The nation's
economy is impacted by traffic congestion either directly or indirectly. The convenience of
road users is necessary to guarantee the nation’s economic progress. Traffic prediction is
necessary to address this issue since it allows us to estimate or predict future traffic. These
days, one of the most significant and well-known growing fields is machine learning (ML),
which is a subfield of artificial intelligence (AI). Machine learning has emerged as a crucial
and promising study field in transportation engineering, particularly in the area of traffic
forecasting but the traditional models use shallow networks for prediction.
Now, deep architecture models such as Deep learning, a kind of machine learning technique,
has recently captured both academic and commercial attention but it has limitations for an
accurate traffic flow prediction as the model performance can be impaired by external factors
like weather conditions, which made the researchers to switch to hybrid models for better
prediction. A decomposition method for traffic flow can lessen the effects of noise and
increase the accuracy of predictions. In this study, Multivariate Empirical Mode
Decomposition (MEMD) was integrated along with different Machine Learning and Deep
Learning techniques to predict traffic flow. In order to test its applicability in real-field
scenarios, a real-time field data was also used. When compared to standalone models, the
hybrid models, notably MEMD-LSTM and MEMD-RF, performed better. When used with
the PeMS dataset, these hybrid models showed lower errors and higher determination
coefficients of 0.925 and 0.820, respectively. Similar to this, MEMD-LSTM and MEMD ANN demonstrated improved determination coefficients of 0.891 and 0.878, respectively,
when tested on the real-field dataset. This study addresses a gap in the literature by
examining how well the proposed hybrid decomposition model predicts traffic flow utilising
hybrid models
