Time Series Forecasting using a hybrid Jaya-FLANN model

Abstract

Time series forecasting (TSF) is one of the challenging problems encountered in various domains of science and engineering. This is because of the uncertainty and vagueness associated with the underlying nonlinear data producing process. Therefore to handle the nonlinear patterns well, over the last three decades, a variety of artificial neural network (ANN) models have been applied for TSF. However, the choice of ANN model, its architecture and algorithm used for its training drastically affects the forecasting accuracy. As compared to other ANN models, functional link artificial neural network (FLANN) model is a single layered feed forward higher order neural network (HONN) with only architectural parameter to determine is the number of inputs. Although it has a simple structure, the functional expansion module of FLANN model expands the inputs to higher dimensions and thus enhances the chances to capture the nonlinear patterns well. In addition, being a single layered network it has less number of trainable weights. Motivated by all these characteristics, in this paper, a FLANN model is used for time series forecasting. In addition, an improved version of Jaya algorithm, called self adaptive Jaya (SAJaya) is being proposed to determine the near optimal weight set of the FLANN model. The proposed hybrid method is implemented using MATLAB and the simulation results are compared with CRO based FLANN (CRO-FLANN), PSO based FLANN (PSO-FLANN), TLBO based FLANN (TLBO-FLANN) and DE based FLANN (DE-FLANN) using ten benchmark univariate time series datasets. The proposed hybrid Jaya- FLANN method provided statistically better result when compared with its counterparts for the time series datasets used.