Research On Time Serie Prediction Using SOM Neural Network

Recently, research on time series prediction using self-organizing map(SOM) neural network has aroused wide concern of researchers at home and abroad, and become a vivid research area with important theory and application value. As a unsupervised competitive learning neural network, SOM has the advantages of simple and intuitive structure, its temporal association memory technique overcomes local optimum problem of traditional methods.This research discussed the improved methods based on SOM neural network and its application in time series prediction, which focusing on prediction accuracy of real-life applications, and it extends the new space in time series prediction using unsupervised neural networks.The main work of this research includes the following aspects:(1) Study the time-series forecasting theory, the structure and the algorithm of SOM neural network. As a generalization to the temporal domain of the associative memory technique of SOM neural network, vector quantized temporal association memory(VQTAM)modeling technique can achieve time series prediction.(2) A class of recursive SOM methods are proposed, include recursive self-organizing map(RecSOM) method and self-organizing map of structured data(SOMSD) method.Recursive SOM methods use context information to reflect statistical properties of data set,RecSOM uses feedback with delay to represent the concept of recursive, SOMSD uses grid coordinates of the winning neuron to represent context information, which is more suitable for structured data. Recursive SOM methods are applied to traffic volume prediction instances,under the same circumstances, compare to other prediction methods, the results validate the proposed methods is feasible and effective.(3) On the basis of VQTAM modeling technique, a class of local auto-regressive(AR)methods based on SOM neural network are proposed. AR-SOM builds multiple local linear AR models, K-SOM computes the K first winning neurons to build a single time-variant local AR model with weight vectors instead of input vector, local linear map(LLM)-SOM updates the coefficient vectors of local models simultaneously with clustering data vectors. In contrast to the global model, the proposed methods are flexible enough to present effective supervisal neural architecture as well as small computing complexity. The methods are then applied to various typical examples of chaotic time series forecasting instances, and further applied to various network traffic forecasting instances and video traffic forecasting instances, compared to existing methods under the same conditions, experimental results confirm the proposedmethods can significantly improve the accuracy of prediction, give considerably better performance, which verified its effectiveness and potential for applications.