Forecasting Time Series by Double Vector Quantization Based on Probabilistic Self-Organizing Map

Abstract

Forecasting time series using nonlinear models has been extensively explored due to their accurate representation of real-world problems. These models incorporate double vector quantization based on Self-Organising Map (SOM), where initially a SOM is used to quantify two spaces: the deformation space and the regression space in the original method. Given the quantifier’s importance, we observe that modifying it can significantly impact the dynamics of the prediction. Therefore, the choice of quantization method becomes a critical area of research for optimizing prediction accuracy. In this paper, we introduce a novel approach: double vector quantization based on the Probabilistic Self-Organizing Map, where we replaced the clustering method SOM in the original model by its probabilistic version (Probabilistic Self-Organizing Map). The stability of this approach is demonstrated to validate its reliability. Furthermore, the effectiveness is assessed through a comparative analysis with traditional double vector quantization models and recurrent neural networks, employing the real-world electricity consumption of Tetouan City, and the benchmark sunspot time series.