Clustered complex echo state networks for traffic forecasting with prior knowledge

For mobile communication traffic, an accurate prediction result plays an important role in network management, capacity planning, traffic congestion control, channel equalization, etc. This paper proposes a clustered complex echo state network for mobile communication traffic forecasting with prior knowledge. In order to reflect some learning mechanisms of real world organization, various complexities such as small-world features and scale-free node degree distribution were introduced to the dynamic reservoir of original Echo State Network (ESN) and viewed as promising tools for function approximation, modeling of nonlinear dynamic systems, and chaotic time series prediction. Our further observation on the real traffic collected by China Mobile Communications Corporation (CMCC) Heilongjiang Co. Ltd., showed the property of multi-periodicities. Based on this characteristic, Fourier spectrum was chosen as the prior knowledge in this paper for generating multiple functional clusters, corresponding to the frequency components revealed in traffic series of each cell, in the dynamic reservoir of complex echo state network. Experiment results on real traffic showed that: the forecasting accuracy of proposed prior knowledge based clustered complex echo state network (PCCESN) model is superior to that of original ESN models and provides decision-making support for network planning and optimization of mobile network.