A Study On The Tourist Arrivals Forecasting Using Improved Echo State Networks

The tourism is the most promising industry at the moment.Accurate prediction on tourism can help relative department formulate reasonable plan and allocate tourism resources.In recent years,neural network has been widely used in tourism forecasting and performs well.However,it has some shortcomings such as the complex training algorithm,the unstable forecasting accuracy and the local optimization.Considering this,Echo State Network has been adopted.With the dynamic reservoir replacing the traditional hidden layer,the training process becomes much easier.Firstly,the training process of echo state networks(ESN),the crucial parameters of the reservoir and the mathematical model of Small World Network are analyzed.Then,a Small World Wavelet ESN(SW-W-ESN)model is designed.Traditional ESN uses randomly generated reservoir topology.Relative researches find that complex networks have more advantages than random networks.So,a small world reservoir is designed to replace the traditional one.Reservoir neurons generally use a sigmoid function as their activation function to remember and translate the information.This confines the types of the activation function in the original reservoir of the ESN.Wavelet-neurons are injected into the reservoir to replace parts of the original reservoir.Finally,the tourist arrivals of Malaysia are adopted to test the effectiveness of the improvement.The results show that SW-W-ESN model achieves better performance than ESN,SW-ESN,W-ESN.Additionally,SW-W-ESN model is applied to forecast the tourist arrivals of Turkey from ten countries.It's verified that SW-W-ESN performs better than the existing methods.