DenseNet-Transformer: A deep learning method for spatial-temporal traffic prediction in optical fronthaul network

The rapid evolution of 6G Radio Access Networks (RAN) towards virtualization and intelligence is driven by the widespread adoption of high-bandwidth services and the proliferation of mobile communications. Active Antenna Units (AAU) interface with Distributed Units (DU) over Passive Optical Networks (PON), serving as crucial bearers that transmit extensive data through the optical fronthaul network to the core network. However, the escalating demand for high-bandwidth services, coupled with the uneven spatial-temporal traffic patterns influenced by factors such as geographical location and urban functionalities across various base stations, poses significant challenges to the optical fronthaul network in terms of operational costs and resource allocation efficiency. To tackle these challenges, we introduce the DenseNet-Transformer spatial-temporal traffic prediction algorithm tailored specifically for the traffic characteristics of optical fronthaul networks. In DenseNet-Transformer, DenseNet captures spatial feature correlations among nearby traffic in adjacent regions, facilitating enhanced learning of traffic characteristics across distant areas through dense connections. The Transformer component learns both long and short-term temporal dependencies, enhancing the algorithm's temporal prediction capabilities using multi-head attention mechanisms and positional encoding. We validate the effectiveness of DenseNet-Transformer through a series of ablation experiments and comparative tests against other algorithms under identical conditions. Experimental results on real datasets demonstrate that, in most scenarios, DenseNet-Transformer outperforms existing algorithms for time traffic prediction in both wireless and optical communication domains, as well as spatial-temporal prediction algorithms.