GMDH modelling for mobile user throughput forecasting
Isah Abdullahi Lawal · 2020
This paper demonstrates the use of GMDH algorithm as an alternative approach for forecasting mobile user throughput in a cellular data network. We measure the hourly throughput per user for three weeks and use it to train a GMDH-based model for predicting the user throughput for the fourth week. We adopt a modelling strategy that employs a single next-day forecaster iteratively to estimate an entire week throughput. Our experimental results show that the GMDH-based forecaster performed very well with a mean percentage error as low as 1.87%. We also compare the performance of the GMDH-based forecaster with that developed using state-of-the-art LSTM method and show that it can achieve a comparable performance against the LSTM method. Moreover, the GMDH algorithmâs ability to select only effective input variables during model training reduces the dimensionality of the training data by 43% and allows the development of simpler and more interpretable throughput forecaster.