Time Series Forecasting Using Differential Evolution-Based ANN Modelling Scheme
- PDF / 2,732,351 Bytes
- 18 Pages / 595.276 x 790.866 pts Page_size
- 47 Downloads / 224 Views
RESEARCH ARTICLE-COMPUTER ENGINEERING AND COMPUTER SCIENCE
Time Series Forecasting Using Differential Evolution-Based ANN Modelling Scheme Sibarama Panigrahi1 · H. S. Behera1 Received: 2 May 2020 / Accepted: 4 October 2020 © King Fahd University of Petroleum & Minerals 2020
Abstract Over the past few decades, time series forecasting (TSF) has been predominantly performed using different artificial neural network (ANN) models. However, the performance of ANN models in TSF has not yet been fully explored due to several issues like the determination of near-optimal ANN architecture for a time series and the efficiency of training algorithm used to determine the near-optimal weights of ANN. Motivated by this, we have proposed an adaptive differential evolution (DE)-based modelling scheme to automatically determine the near-optimal architecture of ANN for a time series under study. Additionally, we have proposed an adaptive differential evolution-based ANN training algorithm (ADE-ANNT) to determine the near-optimal weights of ANN. To make the adaptive modelling scheme consistently effective, several comparisons are made between different alternatives in the treatment of trend component and normalization techniques. Twenty-one benchmark time series datasets are being considered to assess the comparative performance of the proposed method with the established forecasting models, namely autoregressive integrated moving average, exponential smoothening with error, trend and seasonality, deep belief network and multilayer perceptron + Levenberg–Marquardt (LM) method. To assess the efficiency of the proposed ADE-ANNT training algorithm, comparisons are made with the ANN training algorithms based on recently developed evolutionary algorithms, such as TLBO-ANNT, DE-CRO-HONNT and DE-ANNT+; and the most popular LM training algorithm. Extensive statistical analysis on simulation results reveal the statistical superiority of the proposed training algorithm and proposed method when compared with their counterparts for the datasets used. Keywords Artificial neural network · Multilayer perceptron · Differential evolution · Evolutionary neural network · Time series forecasting
1 Introduction Future of most of the phenomena is a consequence of its past. Therefore, future values of a phenomenon can be extrapolated by systematically analysing its past values. Such a process of predicting the future values of a phenomenon by analysing the past observations is known as time series forecasting (TSF). Accurate TSF plays a vital role in almost every area of study including finance, economics, engineering and management science. Conventionally, statistical models like exponential smoothing, autoregressive integrated moving average (ARIMA) have been widely used in TSF [1]. These models work under the assumption of the linear corre-
B 1
H. S. Behera [email protected] Department of InformationTechnology, Veer Surendra Sai University of Technology, Burla, Odisha 768018, India
lation structure of time series data and often fail to provid
Data Loading...
