الصفحات

Alexandria Engineeing Libraries

بحث هذه المدونة الإلكترونية

شرح طريقة تحميل الامتحانات من المدونة

الثلاثاء، 25 مارس، 2014

A Hybrid Time Series Forecasting Model Using Auto Regressive Integrate Moving Average and Stacked Heterogeneous Neural Networks

Title:   A Hybrid Time Series Forecasting Model Using Auto Regressive Integrate Moving Average and Stacked Heterogeneous Neural Networks
Author         :  Sherin Metwally Goda Mohamed 
Collection   M.Sc. Computer

Abstract:  
Obtaining accurate forecasting results for the time series data is an important and sometimes critical task in many fields, the theoretical and the empirical findings have suggested that combining different models can be an effective way to improve the forecasting accuracy of each individual model. Hybrid techniques that combine between linear and nonlinear forecasting models are one of the most important and the most accurate kinds of the hybrid models for time series forecasting, especially when the time series contains both linear and nonlinear patterns. These models can outperform single models. The most widely used models are the models which
combine between the autoregressive integrated moving average (ARIMA) as linear model and artificial neural network model as nonlinear model. In this thesis, we proposed a new hybrid models that combine between the autoregressive integrated moving average (ARIMA) as linear model and the stacked heterogeneous neural network as nonlinear model (SHNN). The autoregressive integrated moving average (ARIMA) model can be considered as a hybrid model of two linear forecasting models which are the autoregressive model and the moving average model with an integrated test for the data stationary. Also the stacked heterogeneous neural network (SHNN) can be considered as a hybrid model of two nonlinear forecasting models which are two neural networks with different activation function at the output layer. So the proposed model is a hybrid model of these linear and nonlinear hybrid models. We also use the genetic algorithms to find the optimal structure of the network, which in turns improve the prediction accuracy of the network. The empirical results with ten well-known real data sets, some of these data sets are stationary series and the others are non-stationary ones, indicate that the proposed model can be an effective way to improve the forecasting accuracy achieved by the individual models. Therefore, it can be used as a convenient alternative model for the time series forecasting, especially when higher forecasting accuracy is needed.

ليست هناك تعليقات:

إرسال تعليق