Deep Learning For City-scale Wireless Traffic Prediction

The future wireless communication systems,i.e.,5G and 6G,are continually em-bracing AI technologies,and gradually moving toward intelligence.Accurate wireless traffic prediction plays fundamental roles in intelligent communications and is the key to the self-organization,self-management and self-optimization of the communication sys-tems.Wireless traffic prediction not only can benefit the network management through dynamic congestion control,also can enhance the energy efficiency by achieving cell-zooming and traffic-aware base station sleeping scheme.However,due to the high com-plexity of human mobility,the high randomness of service requests and the complicated spatial constraints of cross-domain datasets on wireless traffic generation,the accurate prediction of wireless traffic is still a very challenging task.The prediction problem is especially hard when we are focusing on city-scale wireless traffic prediction,because it involves many cells and the complicated spatiotemporal dependence should be carefully considered in the prediction model.Unfortunately,there are still many shortcomings in the existing solutions,such as inaccurate prediction,unable to take into consideration the cross-domain big data.Besides,the prediction results are tending to be the average of historical values when adopting lp loss as objective function.Current solutions can hard-ly meet the requirements of future intelligent communication systems.Hence,the deep understanding of spatiotemporal characteristics of wireless traffic and the accurate traffic prediction problem based on cross-domain big data are the key issues of this dissertation.Consequently,this dissertation mainly focuses on the accurate wireless traffic predic-tion problem based on deep learning techniques.Specifically,the simultaneously spatial and temporal modeling of wireless traffic using convolutional neural networks is firstly in-vestigated.Then,thi,s dissertation proposes a novel prediction framework aiming to solve the cross-domain data fusion by leveraging deep transfer learning.Finally,how to use generative adversarial networks to improve prediction performance is further explored.The main contributions are summarized as follows.1)To simultaneously capture the complex spatial and temporal dependence of wire-less traffic,a city-scale wireless traffic prediction framework,namely STDenseNet,is proposed based on deep learning.STDenseNet adopts convolutional neural networks to model the spatial dependence among different cells.Besides,two deep networks with shared network structure are designed to capture the temporal dependences,i.e.,closeness dependence and period dependence.Based on these two networks,a parametric-matrix-based fusion scheme is further proposed to effectively fuse different kinds of features.Furthermore,the densely connected pattern is adopted in STDenseNet to enhance fea-ture propagation and reuse.The performances of STDenseNet are verified on real world datasets and the results show that STDenseNet can well capture the spatiotemporal depen-dence and effectively reduce prediction error.The parametric-matrix-based feature fusion scheme can significantly improve the performance.2)In addition to the spatial and temporal dependence,the patterns behind wireless traffic generation are very complicated.The cross-domain data of a cell,such as the num-ber of BSs,social activity level and point of interests,has also strong correlations with wireless traffic and will be spatial constraints to traffic generation.Thus,to jointly model the spatiotemporal dependence and this kind of constraints,a novel prediction framework named STCNet is proposed based on cross-domain data and transfer learning,STCNet takes advantages of convolutional long short-term memory networks to capture the spa-tiotemporal dependence and the sequences information between traffic frames.The spa-tial constraints of cross-domain data are well modeled by convolutional neural networks.Different kinds of datasets are fused through feature level fusion scheme.Meanwhile,to characterize the different traffic patterns of different city functional zones,a spectrum clustering-based algorithm is adopted to segment a city into different areas.In order to make full use of the knowledge obtained by one cluster,a successive inter-cluster trans-fer learning strategy is further proposed.Finally,the transfer learning is also explored between different kinds of wireless traffic aiming to enhance prediction accuracy.Ex-periments demonstrate that STCNet can well capture the complicated characteristics of wireless traffic,the prediction errors are relatively small.In particular,the transfer learn-ing can drastically reduce prediction errors and will be a new possibility for enhancing wireless traffic prediction performance.3)The wireless traffic at city-scale does not strictly follow Gaussian distribution,but multi-modal distribution.When adopting lp loss as objective function,the predic-tion results tend to be the average of multi-modal distribution.To solve this problem,a wireless traffic prediction model named TPGAN is proposed in this dissertation based on generative adversarial networks.The generator network of TPGAN adopts STCNet as its core component aiming to simultaneously model the spatiotemporal characteris-tics and also the cross-domain data constraints on traffic generation.Besides,TPGAN combines adversarial loss and lp loss as the final objective function to avoid the inherent drawbacks of mean square error loss.Experiment results show that TPGAN can effective-1y improve prediction performance compared with traditional lp loss-oriented methods.The distribution of prediction results has a high similarity with ground truth.The results also demonstrate that the weights of adversarial loss have great influence on prediction performance.In summary,aiming to solve the shortcomings of current wireless traffic prediction methods,this dissertation focuses on the simultaneously modeling of spatial and temporal dependence of wireless traffic,cross-domain data fusion for enhancing prediction perfor-mance and accurate prediction under multi-modal distribution.The proposed methods can effectively improve prediction performance compared with traditional ones.