MACHINE LEARNING APPROACH ON INPUT VARIABLE SELECTION FOR MONTHLY RAINFALL PREDICTION CONSIDERING TELECONNECTIONS

抄録

 In Artificial Neural Network (ANN) models for rainfall prediction, the model performance relies heavily on the input data characteristics. Input variables of the model, such as meteorological parameters and their spatiotemporal variations, should be carefully selected. Due to the ability of ANN to detect non-linear and complicated patterns behind the data, a new input variable selection method called Zero Input Test (ZIT) is proposed in this study. A comparative study between ZIT and correlation coefficient (CC) is carried out considering the global data of atmospheric variables from lags of 1 to 12 months. Selections from the two approaches are evaluated by building two ANN models based on CC and ZIT selections to predict monthly rainfall of Upper Chao Phraya River Basin (UCPRB), Thailand. The selection results show that CC and ZIT identified different regions to be important variables. However, the rainfall prediction results of both models show similar accuracy and prediction trends. This suggests that regions outside those selected by CC also contribute to UCPRB rainfall, and there are many regions across time variation that influence the phenomenon. Additionally, the current inconsistency in ZIT selections indicates that there might be room for improvement of the method to further develop the input variable selection.