Máquina de estado líquido para previsão de séries temporais contínuas: aplicação na demanda de energia elétrica

Abstract

Among of several aspects of the natural intelligence is its ability to process temporal information. One of major challenges to be addresses is how to efficiently develop intelligent systems that integrate the complexities of human behavior. In this context, appear the Liquid State Machines (LSMs), a pulsed neural architecture (liquid) that projects the input data in a high-dimensional dynamical space and therefore makes the analysis of input data all through a classical neural network (readout). Thus, this thesis presents an innovative solution for forecasting continuous time series through LSMs with reset mechanism and analog inputs, applied to the electric energy demand. The methodology was applied in the short-term and long-term forecasting of electrical energy demand. Results are promising, considering the high error to stop training the readout, the low number of iterations of training of the readout, and that no strategy of seasonal adjustment or preprocessing of input data was achieved. So far, it can be notice that the LSMs have been studied as a new and promising approach in the Artificial Neural Networks paradigm, emergent from cognitive science.