APPLICATION OF FUZZY C-MEANS CLUSTERING AND PARTICLE SWARM OPTIMIZATIONTO IMPROVE VOICE TRAFFIC FORECASTINGIN FUZZY TIME SERIES

By

Author

Presented To

Department of Engineering

ABSTRACT
Forecasting of voice traffic using an accurate model is important to the telecommunication service provider in planning a sustainable Quality of Service (QoS) for their mobile networks. This work is aimed at forecasting Erlang C – based voice traffic using a hybrid forecasting model that integrates fuzzy C-means clustering (FCM) and particle swarm optimization (PSO) algorithms with fuzzy time series (FTS) forecasting model. Fuzzy C-means (FCM) clustering, which is an algorithm for data classification, is adopted at the fuzzification phase to obtain unequal partitions. Particle swarm optimization (PSO), which is an evolutional search algorithm, is adopted to optimize the defuzzification phase; by tuning weights assigned to fuzzy sets in a rule.This rule is a fuzzy logical relationship induced from a fuzzy set group (FSG). The clustering and optimization algorithms were implemented in programs written in C#. Daily Erlang C traffic observations collected over a three (3) month period from 1 December, 2012 – 28 February, 2013 from Airtel, Abuja region, was used to evaluate the proposed hybrid model.To evaluate the forecasting efficiency of the proposed hybrid model, its statistical performance measures of mean square error (MSE) and mean absolute percentage error (MAPE), were calculated and compared with those of a conventional fuzzy time series (FTS) model and, a fuzzy C-means (FCM) clustering and fuzzy time series (FTS) hybrid model.Statistical results of MSE 0.9867 and MAPE 0.47 %were obtained during training of the proposed hybrid forecasting model. Compared with the training results ofMSE 845.122 andMAPE 13.47 %, for Chen?s (1996) FTS model and; MSE 856.145 and MAPE 13.37 %, for Cheng?s (2008); the proposed hybrid forecasting model resulted in a relatively higherforecasting accuracy and precision. Also, performancemeasures of MSE 59.22 and MAPE 3.85 %were obtained during thetesting phase of the proposed model. Compared with the test results of MSE 1567.4 and MAPE 23.98 %obtained for Cheng?s (2008) FCM/ FTS hybrid model, the proposed hybrid forecasting model also showed a relatively higher forecasting accuracy and precision. Finally, it was determined that reversing the weights of the forecasting rules, during training, resulted to a lesser performance;MSE 42.73 and MAPE 0.88 %. Thus, reversing the weights of forecasting rule affected the forecasting accuracy