Abstract
The ever-growing bandwidth demands of emerging services urgently drive optical networks toward ultra-high speed and large capacity. Bandwidth demand forecasting has become a critical issue in the planning and construction of optical networks. Taking service characteristics as the starting point, this paper first analyzes the major service types carried by optical networks and their bandwidth features. Then, general network traffic forecasting methods are summarized. On this basis, typical bandwidth demand forecasting methods for optical networks that are oriented to different service characteristics are systematically reviewed, including service section models, queuing theory models, self-similar traffic models, input traffic iteration, and emerging deep learning-based methods. A comparative analysis is conducted from the perspectives of computational complexity, forecasting accuracy, and applicability. Finally, the limitations of existing methods are discussed, and future research directions based on service characteristics are proposed.