Performance Evaluation of Machine Learning Regression Models for Rainfall Prediction

Accurate prediction of rainfall has always been the most demanding task involved in weather forecasting in view of significant variations in weather patterns. With the advent of machine learning algorithms, it is now possible to predict rainfall with higher precision by extracting hidden patterns from past hydro meteorological data. However, it can be challenging to select a suitable algorithm for the prediction of daily, monthly, or annual rainfall estimates. In this study, three data-driven machine learning (ML) regression models; Random forest regression (RFR), Support vector regression (SVR), and CatBoost regression (CBR) were applied to predict daily and monthly rainfall for Aligarh District, Uttar Pradesh, India. Weather datasets from 1980 to 2020 were utilized, which included maximum and minimum temperature, dew point, relative humidity, wind speed, and cloud cover as input variables and rainfall as the target. Results revealed that CBR surpassed RFR and SVR in both daily and monthly rainfall predictions. The CBR and RFR models predicted daily rainfall with a moderate correlation, while the SVR model could not predict rainfall on daily timescale data. All three ML models predicted monthly rainfall with strong correlations, with the CBR exhibiting the strongest. The study instigates the necessity to understand model complexity for rainfall prediction by investigating the disparity of results between daily and monthly rainfall estimates.