Multi-output neural network for the temperature forecasting in Semarang

Semarang has an increasing extent of UHI for about 8.4% per year within 1994 to 2002. The high temperature that spread rapidly throughout the city causes a negative impact. The research on cases of climate and weather much has been used neural network model because it can capture nonlinear relationship. The purpose of this study is temperature forecasting for some period ahead by using Neural Network (NN). This study will use temperature data in Semarang and Ahmad Yani station. The temperature in Semarang station allegedly to be related with Ahmad Yani station at the same or different time each other, so need to be done multivariate modeling with VAR modeling. Selection of the optimal input based on VAR modeling. The selected VAR model is VARIMA (3,1,0) based on MPACF identification and the smallest AIC value. NN model that used in this study is Feed Forward Neural Network (FFNN). The best model selection is FFNN (2,2,2) with the smallest AIC value i.e. −444,61. The results of time series plot between forecast and actual data shows that error values still relatively small for each station. It can be concluded that FFNN model has weakness i.e. less good for forecasting at testing data.