Author(s): Iriany, A., Rosyida, D., Sulistyono, A. D., Wong, W.K., Sui, J. M.
A neural network constitutes a non-linear model requiring no statistical assumption. Along with the development of which, the neural network model has been frequently combined with time series and spatiotemporal models. This current research combined neural network and spatiotemporal models. One of the spatiotemporal models is the GSTAR-SUR model. The weight projected in this current research is cross-covariance normalized weight. This sort of weight is deemed suitable for data with high variability. The significant variable in the GSTAR-SUR model containing cross-covariance normalized weight was used as the input layer of the neural network model. The hidden layer made use of 10 neurons fulfilling the criteria of the lowest RMSE value and there was 1 neuron used as output. The data were in the form of 10-day precipitations in Junggo, Pujon, Tinjumoyo, and Ngujung, during the period of 2005 to 2014. The findings in this research were obtained by a model for forecasting precipitations in Junggo, Pujon, Tinjumoyo, and Ngujung is NN (16, 19,1) – GSTAR-SUR (31). The cross-covariance normalized weight on the GSTAR-SUR model as input of the neural network model has yielded better forecasting. This research has found out that the NN-GSTAR-SUR model yielded better and more accurate forecasting, showing a R^2 value of 61.77%. This study recommends research is included subset elements in the NN GSTAR-SUR model to improve forecast accuracy.