A Multivariate Artificial Neural Network Approach for Rainfall Forecasting: Case Study of Victoria, Australia

El Nino southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) have enormous effects on the precipitations around the world. Australian rainfall is also affected by these key modes of complex climate variables. Many studies have tried to establish the relationships of these large-scale climate indices among the rainfalls of different parts of Australia, particularly Western Australia, New South Wales, Queensland and Victoria. Unlike the other regions, no clear relationship can be found between each individual large-scale climate mode and Victorian rainfall. Past studies considering southeast Australian rainfall predictability could achieve a maximum of 30% correlation. This study looks into the lagged-time relationships of these modes on Victorian spring rainfall. On the other hand, few attempts have been made to establish the combined effect of these indices on rainfall in order to develop a better understanding and forecasting system. Since rainfall is a complicated atmospheric phenomenon, linear techniques might not be sufficient enough to capture its characteristics. This research attempts to find a nonlinear relationship between the Victorian rainfall and the lagged-indices affecting the region using Artificial Neural Networks (ANN). It was discovered that ANN modelling is able to provide higher correlations using the lagged-indices to forecast spring rainfall in compared to linear methods. Using these indices in an ANN model increased the model correlation up to 99%, 98% and 43% for the three case study stations of Horsham, Melbourne and Orbost in Victoria, Australia respectively. it seems that IOD has a higher effect on the centre and west of Victoria more than the ENSO, while ENSO seems to have a stronger effect on the east side. This method can be used for other parts of the world where a relationship exists between rainfall and large scale climate modes which could not be established by linear methods.