•  
  •  
 

Keywords

Mobile traffic forecasting, Prophet additive decomposition model, Radio access network optimisation, K-nearest neighbours clustering, Weighted barycentre, Artificial intelligence for network planning, Machine learning, Fourth and fifth generation mobile networks, Orange Cameroon, Sub-Saharan Africa

Document Type

Article

Abstract

The exponential growth of mobile data traffic in Sub-Saharan Africa is placing increasing pressure on cellular operators to modernise their radio access network (RAN) planning. Traditional approaches, namely Excel-based linear extrapolation and manual congestion-map analysis, are inadequate for the non-linear, seasonally driven traffic dynamics observed in heterogeneous environments such as Cameroon. This paper presents RAN-OptimForecast, an artificial intelligence (AI)-based decision-support framework that jointly addresses two problems: (1) accurate long-term evolution (LTE) mobile traffic forecasting using machine learning, and (2) optimal geographical positioning of new radio base stations. Because proprietary operator traffic measurements were not available for release, the framework is developed and validated on a reproducible, statistically-calibrated synthetic dataset (240 cells, 36 months) built from a documented additive time-series model (piecewise trend, yearly seasonality, Gaussian noise); the cell geography (coordinates, azimuths, bands) reflects the real Orange Cameroon (OCM) network layout, while traffic volumes are synthetic. The study is therefore a methodological and proof-of-concept contribution, with the complete pipeline released so the identical protocol can be re-run on operator data. Three time-series models (seasonal autoregressive integrated moving average (SARIMA), long short-term memory (LSTM) networks, and the Prophet additive decomposition model) are benchmarked under rolling-origin (walk-forward) validation across multiple horizons. On the synthetic dataset, Prophet attains the lowest error (sMAPE approximately 5% and R2 approximately 0.92 at a one-month horizon, stable to six months), outperforming SARIMA and a seasonal-naive baseline, with all pairwise differences significant under a Diebold-Mariano test (p < 0.001). A K-nearest neighbours (KNN) clustering module combined with a forecast-weighted barycentre positions new sites, and a sensitivity analysis over the neighbour count, search radius, and weighting coefficients characterises the placement behaviour. All modules are embedded in an interactive web application.

DOI

10.30684/2412-0758.2416

First Page

20

Last Page

37

Download

Share

COinS