Echo State Network with Multiple Loops Reservoir and its Application in Network Traffic Prediction

Abstract

Echo state network (ESN), which was proposed as a novel recurrent neural network (RNN), has already been proved to exhibit better prediction ability than traditional neural networks in dealing with time series prediction. However, ESN's randomly generated reservoir structure is of high complexity and can not guarantee the stability of the prediction. In this paper, we propose a novel ESN with deterministic multiple loops reservoir structure (MLR) to avoid the randomness of the reservoir in the classic ESN. In addition, compared with the adjacent-feedback loop reservoir structure (ALR), the novel reservoir structure strengthens the connection of neurons in the reservoir and improves the nonlinear approximation ability of ESN. To test its performance, our MLR-based ESN is applied to network traffic prediction. Extensive simulation results regarding to different prediction steps demonstrate that MLR can achieve higher prediction accuracy, and outperform existing prediction models. Furthermore, we also analyze the influence of parameters of MLR on the prediction accuracy, such as the neuronal interval of multiple loops and the number of loops.