Hybrid neural system for time series prediction

In this paper, we present an incremental modular system of times series prediction. This system is hybrid and is based on three methods, an incremental self organizing map (e-SOM) for dynamically learning the past of the times series, an ascendant hierarchical clustering (AHC) for optimizing the number of classes forming the map and a set of local multilayer perceptrons (MLPs) for predicting the evolution of data in the future. The number of MLPs depends on the number of classes formed by AHC. Our approach called (ILM) is compared with several other methods like a global approach only based on MLP and modular one using SOM and MLP. We demonstrate that ILM method is rather more efficient than the other previous methods