Hybrid Systems to Select Variables for Time Series Forecasting Using MLP and Search Algorithms

Research on time series forecasting has been an area of considerable interest in recent decades. Several techniques have been researched for time series forecasting. There is a fundamental task in any area of knowledge of time series: use past values to predict future values from the available historical series. Thus, a very important step is to define which of these past values will be considered in the prediction process. In this paper it is proposed two hybrid systems to select variables: Harmony Search and Neural Networks (HS + MLP) and Temporal Memory Search and Neural Networks (TMS + MLP). The variables selections improves the performance of learning models by eliminating redundant or irrelevant attributes. To perform a comparative study between the techniques, ten real-world time series were used.