Machine-learning-based short-term forecasting of daily precipitation in different climate regions across the contiguous United States

Abstract

Forecasting the precipitation is needed for water resources management and planning, irrigation scheduling and flood modeling. The purpose of this research is to obtain short-term forecasts of the daily precipitation (1-, 2-, and 3-day-ahead) at 28 sites located in various climate regions over the contiguous United States. To this end, daily precipitation data for the study sites are collected from 1995 to 2019. Data from 1995 to 2014 and from 2015 to 2019 are considered respectively in the training and forecasting steps. The precipitation forecasts are obtained using three artificial intelligence models: wavelet particle swarm optimization adaptive neuro-fuzzy inference system (WPSOANFIS), wavelet group method of data handling (WGMDH), and wavelet long short-term memory (WLSTM). The 28-site-average mean absolute errors of 1-day-ahead precipitation forecasts from WLSTM, WGMDH, and WPSOANFIS are respectively, 0.65, 0.77, and 1.31 mm/d. The 28-site mean root mean square error (RMSE) and coefficient of determination (R2) for the WLSTM are 1.47 mm/d and 0.91, respectively. The average RMSE and R2 for WGMDH (WPSOANFIS) are respectively 1.66 mm/d and 0.88 (3.09 mm/d and 0.59). WGMDH outperforms WLSTM and WPSOANFIS in 2- and 3-day-ahead precipitation forecasting. The machine learning techniques show their best results in the West and Southwest climate zones that have the lowest standard deviation of daily precipitation measurements. Lower accuracies are observed for sites with the largest standard deviations of daily precipitation observations, particularly in the 2- and 3-day-ahead horizons. Deep learning techniques (i.e., WGMDH and WLSTM) generate more accurate precipitation forecasts than WPSOANFIS. The 1-, 2-, and 3-day-ahead precipitations obtained using the proposed models are more accurate than those obtained using existing models.