Pengembangan Neural Networks Model Aditif Untuk Peramalan Data Time Series Stasioner Dan Non Stasioner

Abstract

Statistical methods result in less accurate performance when forecasting stationary and non stationary time series data. In this regard, this article proposes systematic development of neural networks and additive neural networks. Systematic development of neural networks includes stationerity identification, differencing technique, determining architecture utilizing Partial Autocorrelation Function (PACF). Stationerity identification uses Dickey Fuller test. Differencing technique eliminates non stationerity. PACF determines the number of input nodes. The additive neural networks cascades neural networks by considering errors. The computational studies show that systematic development enhances the performance of the neural networks. The additive neural networks significantly reduce errors, especially for forecasting non stationary time series data.