Development of a methodology to forecast time series using few input variables

This paper aims to develop a methodology for choosing the inputs of a multilayer fuzzy inference system to forecast time series power demand values in a substation feeder. The forecast is done by analyzing past time series data. On an iteration process, older data with greater correlation with the previous forecast errors are the inputs of the fuzzy system, which has as output a future demand value. It is attempted to estimate the largest possible horizon reaching the minimum forecast error. The obtained results are satisfactory, showing that the developed methodology is capable of picking a small number of inputs to forecast with accuracy different horizons. Thus, it is intended that this paper can generate contributions in the fields of intelligent systems, dynamical systems and electricity market.