Page 110 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4

However, traditional BFS cannot visit the visited In order to further improve the prediction accuracy,
node, and there is a situation in which the node is Zezhong Feng tried to add more features to improve
visited multiple times in the simulation data. To the traffic forecast model. In the time dimension, we
solve this problem, Lin Xi modified the BFS search add the two features of the day’s weather and the
status part so that the improved algorithm can weekend to make it have traffic and weekend
repeatedly visit the node and adapt the search of the features; in the space dimension, we add the node
link set. traffic at a certain distance around the node as a
feature, and we analyze that it is within a certain
range (tentative 2 Km) where nodes have an impact
on the node traffic of the current link, and such node
traffic is added to features.
Zezhong Feng’s experimental results indicate that:
By adding weather features, the average forecast
Fig. 8 – Improved BFS
accuracy of the model is increased by 0.4%；
For TFM,mainstream time series prediction
algorithms include multiple linear regression, by adding the features of working day/off day, the
Autoregressive differential Moving Average (ARIMA) average forecast accuracy of the model is increased
and Long-Short-Term Memory network (LSTM) by 1.6%；
model [7]. By analyzing the characteristics and data by adding the traffic chrematistics of surrounding
of this research, the influence between the data is nodes, the average forecast accuracy of the model is
almost non-linear. Among the above prediction increased by 1.8%；
models, the LSTM model is designed with a special
structure to memorize and filter the changes in after all three features were added, the average
traffic on the time scale. Therefore, we choose the forecast accuracy of the model increased from
LSTM recurrent neural network model for 91.7% to 95.3%.
prediction to meet the requirements of this
prediction scenario.

Fig. 9 – LTSM adopted Fig. 11
Zezhong Feng believes that the bottleneck for 2.5 Optimization strategy
improving the accuracy of traffic forecasting is that In terms of Topology Optimization Strategy (TOP),
there are few samples and few influencing factors, Zhouwei Gang proposed an optimization strategy of
many nodes of different types, so the model is in two "(ultra-low)*3". First calculate the utilization rate of
layers to reduce overfitting and under-fitting. We all links and the average value of the entire network,
build a model for each node and increase the flow and set a threshold to divide all links into three
information carried by each neuron from 1 to 24, so categories: overload, low load and normal. The
that the model can fit more flow changes. In the end, overload links are processed first, and then the low-
the average accuracy of TFM increased from 91.7% load links are processed. After finishing, adjust the
to 95.3%. threshold value, do it twice again, and finally obtain
the optimized topology by completing three
threshold adjustments.

Fig. 10 – Improved input Fig. 12 – Topology optimization strategy

94 © International Telecommunication Union, 2021