Engineering and Applied Science Research (Mar 2021)
Long-term rainfall forecasting using deep neural network coupling with input variables selection technique: A case study of Ping River Basin, Thailand
Abstract
Long-term rainfall forecast is essential for water resources planning and management. Various approaches for forecasting have been developed, however, the accuracy of the forecast is still not satisfied for real engineering practice, particularly for monthly rainfall forecast with leading time of one-year ahead. This study aims to investigate the capability of the machine learning approach in the forecasting of monthly rainfall by using Deep Learning Neural Network (DNN) as a tool for calculation. Ping river basin, situated in the northern part of Thailand, was selected as a study area due to its availability of long time series of rainfall data. Six rainfall stations, distributed over the river basin, were selected for analysis using monthly rainfall from 1975 to 2018. The stochastic efficiency (SE) and correlation coefficient (r) were used for evaluation of model performance. Based on previous studies in this area, it has been revealed that 24 large-scale atmospheric variables (LAV), which were used as predictors in the DNN model, have correlations with seasonal rainfall over the Ping river basin. The result of the first simulation using all 24 LAV during the validation period (2009-2018) in predicting monthly rainfall for six rainfall stations for one-year ahead indicates that DNN is capable of forecasting with an accuracy of the forecast ranging from 58% to 72% with correlation coefficient from 0.59 to 0.82. Further improvement of the forecast was also conducted by the input selection technique resulting in a reduction of input LAV from 24 to 13 LAV. The second simulation of DNN with the input selection technique reveals that DNN provides better accuracy of the forecast for one-year ahead with the stochastic efficiency of the forecast ranging from 69% to 78%, with correlation coefficient from 0.75 to 0.82 for all stations.
