Deep learning application for disaggregation of rainfall with emphasis on preservation of extreme rainfall characteristics for Indian monsoon conditions

摘要

Abstract Temporal disaggregation of rainfall data is an important topic of research due to the lack of good quality high temporal resolution data for several regions and locations. Long sequences of high-resolution rainfall data is extremely valuable for various purposes in water resource engineering, especially in urban hydrology. Several methods of disaggregation have been studied worldwide to convert low temporal resolution rainfall data to a higher temporal resolution to generate rainfall data with a reliable degree of statistical accuracy. Artificial neural network (ANN) is a machine learning model which is currently being widely utilised in several different fields due to its wide adaptability and versatility in modelling different physical phenomena. However, its application in the disaggregation of rainfall data from daily to hourly temporal levels has not been extensively tested, especially in a monsoon-based rainfall system like the Indian monsoon climate. This paper studies the feasibility of a deep learning ANN in the Indian monsoon system while combining it with a K-means clustering algorithm for the purpose of disaggregation. The results show that the model can effectively conserve hourly rainfall statistics while also capturing extreme rainfall characteristics post disaggregation. This is seen in its ability to conserve the hourly mean and the number of dry hours while being able to generate Intensity–Duration–Frequency (IDF) curves which closely resemble the IDF curves generated from the observed data. The model can generate values at high temporal resolutions of up to 1-h durations using daily rainfall inputs which may be aggregated to different temporal aggregation levels.