Hybrid SARIMA-ELM-based Data Traffic Forecasting: The Case of UMTS Network in Addis Ababa, Ethiopia
No Thumbnail Available
Date
2018-10-25
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
AAU
Abstract
In Universal Mobile Telecommunications Service (UMTS) network planning, data traffic
demand is one critical input in deciding dimension of network elements. Past data
collected from deployed UMTS network can be used to forecast future demand. In the
context of ethio telecom, the sole telecom service provider in Ethiopia, the future demand
forecast is, however, based on number of subscribers growth forecast obtained from
marketing section. This approach assumes uniform data demand per subscriber to obtain
the total data demand. Understandably, it does not utilize the data growth information
which is already available in the network.
Forecasting the traffic demand based on historical data from network can enhance the
marketing inputs and the traffic model accuracy. In this regard, taking data from ethio
telecom’s UMTS network, a prior research has used Seasonal Autoregressive Integrated
Moving Average (SARIMA) model to forecast a one month data traffic demand.
However, the research did not consider the non-linearity observed in the data traffic. This
thesis handles this non-linearity via a hybrid model that accounts the linearity with
SARIMA model and the non-linearity via Extreme Learning Machine (ELM) model; here
after called the hybrid SARIMA-ELM model.
A one and half year (i.e., from April 2015 – June 2016) data traffic collected from five
Radio Network Controllers (RNCs) of the UMTS network in the city of Addis Ababa is
used for the forecast. The forecasting performance metrics are: Root Mean Square Error
(RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and
Mean Absolute Square Error (MASE). The results indicate that the hybrid SARIMA-ELM
model with SARIMA order of (0,1,1) (1,0,1)7 is selected with 3.75% increase in forecast
than SARIMA only model. The outperform SARIMA-MLP, which has the second lower
error, with 24.8% percentage error reduction.
Description
Keywords
Time Series, traffic, prediction, model, forecast, linear, non-linearity, hybrid, residual SARIMA model, UMTS, ELM model