The increasing popularity of Internet of Things applications, cloud computing services and 5G mobile extensive deployment have led to a tremendous amount of data transported over the Internet. Nowadays, the optical fiber is the most reliable and appropriate data transport system to support the era of Big Data. However, optical network operators are facing significant challenges to meet the exponential demand for bandwidth in secure and cost-efficient ways. To guarantee an error free transmission, a static and relatively large safety margin is reserved when designing optical networks. This margin accounts for factors such as fiber aging and power fluctuations. As a result, the use of the available physical infrastructure is suboptimal. Squeezing the security margin to a near-zero level may help maximize the delivered bandwidth. Therefore, a rigorous understanding of the network behavior is essential. In this context, collecting performance monitoring data in operational optical networks could be very advantageous. Indeed, it makes it possible to monitor the evolution of the quality of performance in optical lightpaths on a daily basis. Real field monitoring data collected over several seasons could be used for quality of performance prediction. Existing researchs have been limited to doing such prediction using only synthetic data.
This project studies the quality of performance prediction measured by the signal to noise ratio during the next 24 hours in operational optical networks. A Machine Leaning apprach is used. The study was limited to a few optical lightpaths whose choice will be justified. Five time series prediction algorithms are proposed and evaluated according to the following performance metrics : the biais, the mean absolute error, the root mean squared error and the coefficient of determination. The neural network architectures being evaluated in here are : Long Short-Term Memory and Gated Recurrent Unit(GRU) which are recurrent neural networks and a one dimensional convolutional neural networks (1D-CNN). The architectures are compared with an Auto Regressive lntegrated Moving Average (ARIMA) model.
Results show that the effectiveness of the evaluated methods to model and predict SNR changes depends on the lightpath itself. It is not guaranteed that all operational ligthpaths contain predictable patterns. Some ligthpaths are just predictable in the short to medium term, others are predictable 3 days ahead. The stateful LSTM performs better than the other methods. Better prediction accuracy is obtained by applying transfer learing from two different data sources. The ARIMA, a simple method, also produce satisfactory prediction results for non-seasonal lightpaths with linear patterns.
Mezni, A. (Author),
Desrosiers (Supervisor) &
Tremblay (Co-supervisor),
3 Sept 2020Student thesis: Master's thesis › Master in Engineering: Engineering