Qualitative and quantitative Multi-Model forecasting with nonlinear noise filter applied to water demand

An application of a nonlinear simple filter in a Multi-Model Predictor based on the decomposition of its qualitative and quantitative components is proposed. We propose a filter that increases the prediction accuracy when a nonlinear structure is found in the residuals of the qualitative component. The residuals are the mismatch between the real and the predicted qualitative component of the time series. The information given by these residuals is used to find unmodeled structures to correct a new prediction. The proposed implementation is applied to predict the next 24 hours ahead of hourly water demand for control purposes.