Dynamical behavior of multivariate time series for SOI, precipitation/temperature in Fukuoka and its prediction by artificial neural networks

Large atmospheric circulation has affected local/regional hydro-meteorological variables such as precipitation and temperature. The large-scale circulation represented by Southern Oscillation Index (SOI) in the present study has played a driving force affecting the local variables. The underlying interaction among them is difficult to detect directly due to the existence of noise and strong nonlinearity. In the present study, simultaneous predictability of SOI, precipitation, and temperature at Fukuoka was verified through noise reduction by low pass filtering and training of artificial neural networks (ANNs), presenting remarkable properties that can represent the nonlinearity in a system. Two types of transfer function (i.e., hyperbolic tangent and pure linear functions) were applied to hidden nodes, while only pure linear function was used for output layer. Possible extrapolation beyond the extreme values in training was verified with the testing and validation datasets. The observed and predicted values for the two cases were depicted in three-dimensional phase space to reveal the dynamical behavior of the interaction among the regional driving force and local hydro-meteorological variables, as well as shown in the respective time series plots. The identified parameters from training of ANNs were verified in the testing and validation phase as well.