[A-AS04] Machine Learning Techniques in Weather, Climate, Hydrology and Disease Predictions

[AAS04-03] Potential of Deep Learning Approach to Predict Multi-step ahead Daily Rainfall

*Mohd Imran Khan¹, Rajib Maity¹ (1.Indian Institute of Technology Kharagpur, Kharagpur, West Bengal, India.)

Keywords:Deep Learning, Machine Learning, Multi-step ahead daily rainfall, Deep One-dimensional Convolutional Neural Network (deep Conv1D), deep Multi-Layer Perceptron (deep MLP), Support Vector Regression (SVR)

Prediction of daily rainfall is a challenging task and holds paramount importance for many operational and scientific applications. Recent advancement in Artificial Intelligence (AI) / Machine Learning (ML) approaches are proven to have outperformed many existing approaches in understanding some complex phenomena, such as image processing, sequence learning, speech recognition etc. This study portrays the potential of a supervised Deep Learning (DL) approach for multi-step ahead (1-day to 5-day in advance) daily rainfall prediction by using a set of nine meteorological precursors as input to the model (Khan and Maity, 2020). These precursors, closely associated with daily rainfall variations, are obtained from climate model simulations. In general, simulation of meteorological variables is much better as compared to rainfall by any climate model. Moreover, observed records of meteorological variables are sparsely available, if not completely unavailable at many locations. Thus, the proposed DL approach, based on deep one-dimensional Convolutional Neural Network (deep Conv1D), helps in augmenting the quality of rainfall prediction by utilizing simulated meteorological variables from the climate model. The developed model is applied to twelve cities in Maharashtra, India, located in different climatic regimes in terms of daily precipitation characteristics. A 7-fold cross validation indicates satisfactory model performance for rainfall prediction of 1-day to 5-day in advance. The performance of the deep Conv1D model is also compared with deep Multi-Layer Perceptron model (deep MLP), a popular DL based approach and with Support Vector Regression model, an ML approach in terms of three statistical measures viz. coefficient of correlation, root mean square error and Nash–Sutcliffe efficiency. The improvements in the proposed model vary from marginally to reasonably across the cities. It is also noticed that deep MLP and SVR models are not able to predict the daily rainfall values exceeding 150 and 180mm respectively during the testing period. Overall, this study establishes the fact that deep Conv1D is more effective in capturing the complex nonlinear relationship between the meteorological variables and rainfall variability. The benefit is due to the unification of potentials of convolutional and fully connected layers in extracting the hidden information from hydrometeorological association of variables. However, the performance of the model gradually decreases as the lead time advances. The proposed approach is expected to be helpful in agriculture, irrigation scheduling, and even in managing flood due to heavy rainfall.

Reference:
Khan, M.I. and R. Maity (2020), Hybrid deep learning approach for multi-step-ahead daily rainfall prediction using GCM simulations, IEEE Access, IEEE, 8(1), 52774-52784, doi: 10.1109/ACCESS.2020.2980977.