טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentVolodin Maryana
SubjectArtificial Neural Network Application to Time Series
Forecasting
DepartmentDepartment of Quality Assurance and Reliability
Supervisor Dr. Leonid Burstein
Full Thesis text - in Hebrew Full thesis text - Hebrew Version


Abstract

It is only recently that quality engineers, following the statisticians, have started to use artificial neural networks (ANN) for solving various problems in their areas. Nevertheless, up to now ANN application to time-series modeling and forecasting remains outside the mainstream quality-science literature. This paper examines the accuracy of prediction methods based on such applications and assesses their advantages and disadvantages. Also studied is the effectiveness of data preprocessing by detrending of time-series. The forecasting problem in this study is domestic power consumption in Israel, using data for the year 2000 to 2005 with a forecast for 2006. Each applied ANN model was studied using two versions: the original data and the detrended data. Moreover, two different data sets, obtained on a monthly and a bi-monthly basis, were used for each model. Three reliable versions were chosen the Linear Neural Network (LNN), the Generalized Regression Neural Network (GRNN) and the Focused Time Delay Neural Network (FTDNN). The mean error (ME) and the standard deviation (STDV), the autocorrelation coefficient (ACC) and the hypothesis testing for comparison results were determined for all ANN methods. In the Introduction, the principal concepts of time series analysis, ANN methodology and a suitable computerized toolbox for time series prediction are presented. Chapter 2 presents the goals of the work and their realization methods. Chapter 3 is the Literature Review. The central chapters (4-7) present the theoretical background, a description of the main methods used in this study and an analysis of the prediction results. The prediction method was verified against a classical one (Fourier transform method). The final chapter presents the comparison results, conclusions and recommendations. It is indicated that bi-monthly data are preferable in long-term predictions and monthly data in short-term ones. All of the models are suitable for domestic power consumption: the GRNN (ME 5.8%, STDV 4.4%, ACC 91.2%, bi-monthly detrended data), which give the most accurate results and the LNN (ME 6.9%, STDV 5.1%, ACC 92.7%, bi-monthly detrended data) are most suitable. The statistical analysis did not show a difference in results between these methods. However, the FTDNN (ME 5.8%, STDV 7.7%, ACC 92.3%, bi-monthly detrended data) entailed difficulties with the network training with attendant excessive computer time, as well as predictive instability with the need for repetition. In these circumstances, the LNN and GRNN were rated best. The Appendices present the detailed data tables, algorithms, computation results and the specialized Matlab programs.