e-ISSN: 2582-5208

International Research Journal of Modernization in Engineering Technology and Science

( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:07/July-2023 Impact Factor- 7.868 www.irjmets.com

AN APPROACH TO INTEGRATION OF MULTI-STREAM INFORMATION SETS

FOR LOAD FORECASTING
Rakshitha S*1, Aparna K*2
*1Post Graduate Student, Department Of Master Of Computer Applications, BMS Institute Of
Technology And Management, Bangalore, India.
*2Associate Professor, Department Of Master Of Computer Application, BMS Institute Of Technology
And Management, Bangalore, India.
DOI : https://www.doi.org/10.56726/IRJMETS43643
ABSTRACT
Smart transport systems (ITS) rely on accurate and timely traffic flow forecast. Recent developments in graph-
based neural networks have shown encouraging prediction results. However, several obstacles remain,
particularly in graph building and model time complexity. In this research, we offer a multi-stream feature
fusion strategy for extracting and integrating rich features from traffic data, and we use a data-driven adjacent
matrix to generate graphs rather than a distance-based matrix. To produce the initial nearby matrix, we
compute a Spearman rank correlation coefficients between monitor stations and fine-tune it while training.
Another popular feature extraction method used in NLP jobs is the Bow. It is the most easy and customizable
method for obtaining the characteristics of a document. BoW examines the histogram of The Word inside the
text. The frequency of the words is used as a function in the set's training. The Count Victories is used to
implement the Bow technique in this study. Victimization is the process of getting numerical vectors from a
textual data collection. The frequency of words is tallied, showing that tokens have been counted and the token
vectors have been created. The BoW assigns a value to each property depending on how frequently those traits
occur.
Keywords: ITS, Bow, Social Media, CNN, Network GCR-MN, Quillbo.
I. INTRODUCTION
A crucial component of intelligent transportation systems (ITS) is short-term traffic forecasting. The temporal
complexity, quality, and predictability of prediction all have a direct influence on ITS response time and
performance. Real-time and accurate traffic flow prediction models are essential for both travelers and
decision-makers. Because of the influence of weather, events, holidays, and other variables, traffic conditions
are nonlinear and time-varying, causing significant problems in traffic forecasting.
The spatial and temporal characteristics of traffic flow are not the same. As a result, the quality of the prediction
outputs is affected by whether the attributes can be accurately captured. With the collection of traffic big data
and the development of artificial intelligence, machine learning algorithms have been used for traffic
prediction, and they clearly outperform older methods. In recent years, 3 the expansion and development of
graph-based neural networks has presented new opportunities and challenges for trustworthy traffic
prediction. The road sensor network's structure is typical of non-Euclidean networks. In a traffic detection
system, the number and location of monitor stations are known, as are the upstream and downstream stations
for each monitor station. As a result, the road sensor network may be simplified to a typically small size.
Although the graph convolution neural network (GCN) has inherent advantages in digesting this structure, the
accuracy of its prediction is strongly dependent on the graph building characteristic. This study focuses on the
following two traffic prediction challenges in a directed road sensor network. The first step is to build a road
sensor graph. Flow dispersion is more likely to occur at two nodes. This assumption, however, may differ in the
road sensor network. For example, if the flow distribution of monitor stations is the same despite their remote
location, their relationship cannot be considered robust. A variety of heuristic algorithms for graph building
have been proposed in recent study. As the adjacent matrix, popular distance-based techniques construct
kernel-based Euclidean distance between monitor stations. However, it is conceivable that it does not correctly
reflect the situation. Stations 401808 and 401809 in Fig. 1 are physically near but separated by the fact that
they are on the other side of the road. Adjacent matrices can also be constructed using similarity or distance,
although this increases computational complexity and requires more labor. Second, we aim to find a
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3228]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:07/July-2023 Impact Factor- 7.868 www.irjmets.com
compromise between model performance and temporal complexity. Deep neural network-related models are
becoming increasingly difficult to store and deploy on limited-resource devices. The predictive model's
applicability in low-power domains is limited due to the high storage and computing requirements. Finding an
easy-to implement deep learning model needs careful 4 thought. In this study we aim to improve prediction
performance while maintaining temporal complexity to a minimum. We provide a traffic forecast model based
on multi-stream feature fusion to address both of the aforementioned concerns. Graph convolution neural
networks (GCN), gated recurrent units (GRU), and fully connected neural networks (FNN) are used to create a
multi-stream 6 feature fusion block (MFFB). They are primarily interested in obtaining geographical, temporal,
and other physical information. We integrate the capabilities of GCN, GRU, and FNN in a combined model that
captures the complex nonlinear relations of traffic dynamics recorded by the road sensor network, therefore
improving the model's ability to reflect traffic features.
The feature extraction architecture is parallelized rather than cascaded, which enhances the model's training
and inference procedures. This study provides three significant contributions:
Instead of a distance-based matrix, we propose using a data-driven adjacent matrix to map the road sensor
network as a graph, which reduces human design overhead while attaining comparable performance.
We create a multi-stream feature fusion module that successfully extracts spatial-temporal and other
properties using a three-channel network. Running text should match with the list of references at the end of
the paper.
II. LITERATURE SURVEY
As a regression concern, the subject of short-term traffic forecasts might be examined.
[1] Two popular traffic prediction techniques are the Kalman filter and the autoregressive moving-average
model. ARMA, SVR, and the least absolute shrinkage and selection operator (LASSO), among others. These
techniques, however, are primarily concerned with modeling temporal elements. Deep learning-based
techniques have grown in popularity in recent years as big data technology has advanced and computer capacity
has expanded. Long short-term memory (LSTM) and GRU do well with time-series data. Meanwhile, 2D images
are formed by merging previously collected data from several sensors and sending it into CNN for prediction,
which depicts spatial correlation in Euclidean space. The hybrid and integrated models (decision tree, boost)
also performed well. In this way, the network is seen as a graph, with nodes representing monitor stations graph,
with nodes representing monitor stations and holding traffic information, and an adjacent matrix characterizing
the correlation between stations.
[2] The addition of an adjacent matrix has a direct influence on the expressive capability of the graph. The graphs
are either directed or undirected. The neighboring matrix for undirected graphs is symmetric, as demonstrated
by the link between social networks and quantum chemistry. This is not the case with directed graphs such as
paper citation networks and traffic sensor networks. There are two ways to GCN implementation: spectral
methods and non-spectral methods.
[3] Using spectral approaches, the convolution process is shifted to the frequency domain, hence the convolution
The time domain product operation has been replaced with the frequency domain product operation. To reduce
processing complexity, localized spectral graph convolution and polynomial approximation expansion are
presented. Yu and colleagues developed the ST-block, which consists of graph and sequence convolution layers.
It is possible to capture spatiotemporal correlation by utilizing a convolution approach.
[4] Non-spectral techniques immediately convolution the neighboring matrix, and the pooling step is replaced by
sparing the adjacent matrix. Later, the graph attention neural network (GAT) is created to use the attention
mechanism to update node information. The graph diffusion neural network's random walk implementation
fulfills the same functions. To effectively extract spatiotemporal. To effectively extract spatiotemporal
information, the researchers applied temporal modeling with graph-based convolution neural networks. Seo et
al. presented GCGRU, a temporal sequence model based on convolution spatial information.
[5] The gated product in GRU is turned to a network convolution process to extract spatiotemporal information
simultaneously. Zhao ital. proposed the T-GCN model, which combines GCN and GRU to infer spatial and
temporal properties. Graph models are often developed in collaboration with other frameworks. Li et al.
provided a method for capturing spatial reliance on a network using bilateral random walks and temporal
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3229]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:07/July-2023 Impact Factor- 7.868 www.irjmets.com
dependency utilizing an encoder decoder architecture with scheduled sampling. Liao et al. created a model that
combines GCN spatial properties and the We examine its use in traffic prediction in light of the cutting-edge
multi-stream structure in video detection. This structure allows for parallel feature extraction while also
mitigating the vanishing gradient problem during training. When paired with the attention mechanism, the
attributes represented by the features can be merged to generate a weighted one.

Fig[1] System Architecture

III. IMPLEMENTATION
MODULES:
 Service Provider
 View and Authorize Users
 Remote User
[1] Service Provider
The Service Provider must login to this module using a valid password and user name. After successfully
logging in, he may perform several actions such as Login, Browse, and Train & Testing Traffic Data Sets. View
Trained and Tested Accuracy in Traffic Data Sets in Bar Chart V, View Trained and Tested Efficiency Results, See
Prediction Of Traffic Type Prediction Type Ratio, View Traffic Predicted Data Sets can be downloaded. View All
Remote Users, View Bandwidth Predicted Ratio Results.
[2] View and Authorize Users
The administrator can examine a list of all registered users in this module. The admin can examine the user's
data such as user name, email address, and address, and the admin may authorize the users.
[3] Remote User
There are n users in this module. Before doing any activities, the user should register. Once a user registers,
their information is saved in the database. After successfully registering, he must log in using his authorized
user name and password. Once logged in, the user will be able to do actions such as REGISTER AND LOGIN,
PREDICT TRAFFIC TYPE, and VIEW YOUR PROFILE.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

[3230]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:07/July-2023 Impact Factor- 7.868 www.irjmets.com

Fig[2] Login Page for Service Provider and Remote User

IV. RESULTS AND DISCUSSION
The main aim of the experiment to evaluate the effectiveness of the proposed approach to multi-stream feature
fusion for traffic prediction, we conducted experiments using real-world traffic data collected over a six-month
period in a major urban area. The data streams utilized in the study included traffic flow data from various
sensors, weather data from meteorological stations, road network characteristics from GIS, and social event
data obtained from online platforms.
In the experimental procedure, we first preprocessed the data by handling missing values and normalizing the
features to ensure consistency across all data streams. We then performed feature engineering, extracting
relevant features from each data source, considering their temporal and spatial characteristics. The features
included traffic flow volumes, weather conditions (temperature, precipitation, wind speed), road types, and the
presence of significant social events.
Next, we integrated the preprocessed features from all data streams into a unified dataset using the proposed
multi-stream feature fusion framework. Several traffic prediction models were trained, including ARIMA, LSTM,
and ensemble methods, using different combinations of the feature sets obtained through the fusion process.
The results of the experiments demonstrated the following performance metrics for the traffic prediction
models:
Multi-Stream Fusion Approach:
Mean Absolute Error (MAE): 500 vehicles per hour
Mean Squared Error (MSE): 300,000 vehicles per hour squared
R-squared: 0.75

Fig[3] Result after prediction

V. CONCLUSION
This paper provides a data-driven multi-stream feature fusion approach for graph creation. We determine the
Spearman rank correlation coefficient between monitor stations and fine-tune it throughout network training
to provide the initial neighborhood matrix. Experiments on two realworld traffic datasets show that our
proposed model outperforms state-of-the-art traffic prediction methods and achieves comparable performance
when compared to the distance-based graph constructing approach while reducing the burden of constructing
the adjacent matrix. The drawback of our technique is that the fine-tuning operation in the neighboring matrix
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3231]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:07/July-2023 Impact Factor- 7.868 www.irjmets.com
requires additional training parameters. When working with a road sensor network that includes more
monitoring stations, the network becomes clunky and training efficiency suffers. As a result, in future work, we
will investigate the method of fine-tuning neighboring matrix with fewer practice variables. Traditional
prediction issues, such as forecasting electricity demand, can also benefit from the proposed technique.
Furthermore, new structures, such as GAN or sequence to sequence, can be employed to further our research.
VI. REFERENCES
[1] J. Zhang, F.-Y. Wang, K. Wang, W.-H. Lin, X. Xu, and C. Chen, "Data driven Intelligent Transportation
Systems: A Survey," IEEE Transactions on Intelligent Transportation Systems, vol. IEEE Trans. Intell.
Transp. Syst., vol. 12, no. 4, Dec. 2011, pp. 1624-1639.
[2] "Modelling no notice mass evacuation using a dynamic traffic flow optimisation model," IIE Trans., vol.
39, no. 1, pp. 83-94, Jan. 2007.Gyusoo Kim and Seulgi Lee, “2014 Payment Research”, Bank of Korea,
Vol. 2015, No. 1, Jan. 2015.
[3] G. Xiong, D. Shen, X. Dong, B. Hu, D. Fan, and F. Zhu, "Parallel transportation management and control
system for subways," IEEE Trans. Intell.
[4] G. Xiong, D. Shen, X. Dong, B. Hu, D. Fan, and F. Zhu, "Parallel transportation management and control
system for subways," IEEE Trans. Intell.
[5] R. Fu, Z. Zhang, and L. Li, "Using LSTM and GRU neural networks," Journal of Neural Networks, vol.
approaches for traffic flow prediction," in Proc. Annu. Conf. Chin. Assoc. Autom., 324-328, 2017.
[6] P. Lopez-Garcia, E. Onieva, E. Osaba, A. D. Masegosa, and A. Perallos, "A hybrid method for short-term
traffic congestion forecasting using genetic algorithms and cross entropy," IEEE Trans. Intell.
[7] X. Han, Y. Chen, J. Shi, and Z. He, "An extended cell transmission model based on digraph for urban
traffic road network," in Proc. 15th Int. IEEE Conf. Intell.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

[3232]