Hybrid Deep Learning Approach for Multi-Step-Ahead Daily Rainfall Prediction Using GCM Simulations

Abstract

Deep Learning (DL) is an effective technique for dealing with complex systems. This study proposes a hybrid DL approach, a combination of one-dimensional Convolutional Neural Network (Conv1D) and Multi-Layer Perceptron (MLP) (hereinafter referred to as hybrid Conv1D-MLP model), for multi-step-ahead (1-day to 5-day in advance) daily rainfall prediction. Nine meteorological variables, closely associated with daily rainfall variation, are used as inputs to the hybrid model. The causal variables are obtained from a General Circulation Model (GCM). In general, simulation of meteorological variables from GCM is much better than rainfall estimate and observed records of meteorological variables is sparsely available, if not completely unavailable at many locations. Thus, proposed scheme helps to establish the effectiveness of the DL approach in augmenting the quality of rainfall prediction, exploiting the potential of GCM in simulating meteorological variables. The developed hybrid model is applied to twelve different locations in different climatic regimes in terms of daily precipitation characteristics. The proposed hybrid approach is compared with a DL approach namely, Multi-Layered Perceptron (deep MLP) and another machine learning approach namely, Support Vector Regression (SVR). It is also found that the performance of the model gradually decreases as the prediction lead time (in days) increases. Overall, this study establishes the fact that the hybrid Conv1D-MLP model is more effective in capturing the complex relationship between the causal variables and daily variation of rainfall. The benefit is due to the unification of potentials of individual approaches for extracting the hidden features of hydrometeorological association.