Rainfall Prediction using Machine Learning Techniques for Sabarmati River Basin, Gujarat, India

Abstract

Rainfall has a direct effect on agriculture, hydroelectric generation, and water resources management, etc. Many natural catastrophes are also closely linked to rainfall intensity and duration, including flood and drought. Therefore, it is essential to have fast and reliable technique of forecasting rainfall intensity and duration for regional water resource management. Timely rainfall forecasting is also required to avoid and mitigate potential harmful impacts of natural catastrophes such as landslides, floods, and droughts. Rainfall prediction is usually based on numerical weather models combined with meteorological radar data. Such models have been used extensively in studies, including multiple regressions and climatology averaging techniques, numerical methods, and empirical formulations. Forecast accuracy depends on uncertainty, and probabilistic forecasting handles the challenge of unpredictability better than deterministic predictions. The behavior of Random Forest, Gradient Boosting, and Decision tree models has been studied to optimize the results generated from data fed into them. Gradient Boosting was found to work best among those tested, with features related to and affecting rainfall predictability giving an accuracy of 93%. Random forest and Decision tree method having 90% and 78.5% of accuracy was achieved respectively. It was also observed that Mean Absolute Error (MAE) is 1.54, Mean Squared Error (MSE) is 24.94, Root Mean Squared Error (RMSE) is 4.99 for the 40 year time period data. This prediction will be useful for the Meteorological Department, State Disaster Management Department, Water Resources Management Department of State including Dam and Reservoir inflow management.