Forecasting Electricity Demand by Neural Networks and Definition of Inputs by Multi-Criteria Analysis

The planning of efficient policies based on forecasting electricity demand is essential to guarantee the continuity of energy supply for consumers. Some techniques for forecasting electricity demand have used specific procedures to define input variables, which can be particular to each case study. However, the definition of independent and casual variables is still an issue to be explored. There is a lack of models that could help the selection of independent variables, based on correlate criteria and level of importance integrated with artificial networks, which could directly impact the forecasting quality. This work presents a model that integrates a multi-criteria approach which provides the selection of relevant independent variables and artificial neural networks to forecast the electricity demand in countries. It provides to consider the particularities of each application. To demonstrate the applicability of the model a time series of electricity consumption from a southern region of Brazil was used. The dependent inputs used by the neural networks were selected using a traditional method called Wrapper. As a result of this application, with the multi-criteria ELECTRE I method was possible to recognize temperature and average evaporation as explanatory variables. When the variables selected by the multi-criteria approach were included in the predictive models, were observed more consistent results together with artificial neural networks, better than the traditional linear models. The Radial Basis Function Networks and Extreme Learning Machines stood out as potential techniques to be used integrated with a multi-criteria method to better perform the forecasting.