An implementation of a Multi-model predictor based on the qualitative and quantitative decomposition of the Time-series

Abstract

In this paper, an implementation of a Multi-Model Predictor architecture based on a framework composed by modules is proposed. The modules interacts among them to detect and activate forecasting modes given the input data collected so far. The implementation is oriented to exploit the qualitative and quantitative information of periodic time series using K-means, K-Nearest Neighbors and Seasonal ARIMA applied to the water demand prediction that presents periodicity in its behavior showing better performance compared to Seasonal ARIMA with daily pattern and Double Seasonal Holt-Winters.