Using temporal neural networks to forecasting of broadband network faults

The goal of this study is to research a possibility of short — and long-term forecasting of the expected number of faults in broadband telecommunications networks using various prediction models. The experiment is conducted on actual live data collected from a telecommunications network in the period between 2009 and 2011 and the number of faults included in the analysis exceeds 1.5 million. Research focus has been placed on dynamic neural networks specialized for prediction in nonlinear systems, with autoregressive integrated moving average method included for comparison, since it previously demonstrated satisfactory results in fault prediction in broadband networks. To achieve the highest accuracy for each specific test case, variable parameters have been adjusted for every method.