An AI/ML-based study on North Indian Ocean Cyclonic Disturbances and City-specific rainfall: Climatological Analysis and Future Prediction

Abstract

Extreme weather events including the cyclonic disturbancs (CDs) can cause significant damage to livelihood. Hence, a precise prediction may have great positive impacts on the life and economy. This study is an attempt to use AI/ML-based frameworks for performing city-specific rainfall analysis, and studying North Indian Ocean (NIO) CD activity, and their landfalling over the Indian coasts (both eastern and western coasts) along with the rim countries. The NIO CD-related analysis involves both seasonal and annual variations along with categorization. Also, an effort is made for forecasting as well. NIO CD-related studies are carried out using the ARIMA model, whereas rainfall-related analysis is carried out using techniques like Long Short Term Memory (LSTM) networks, Gated Recurrent Unit (GRU) and Bidirectional LSTM (BiLSTM). The ARIMA model prediction shows a decreasing trend over the North Indian Ocean. An increasing trend is quite visible over the AS, whereas over the BOB, a decreasing trend is seen. On the east coast of India, an increase in landfal is predicted by the ARIMA model over Tamil Nadu, whereas on the west coast, Konkan and Goa is going to experience an increase. In case of the RIM countries, an increasing trend is seen in Bangladesh and IAA (Iran, Arabian Peninsula and Africa). The rainfallrelated analysis is carried out over 45 smart cities using 121 years of IMD gridded data by computing monthly averages for the years 1901 to 2021. The model was trained using actual grid point value along with 8 neighbourhood grid points data based on region wise and city wise performance of the three deep learning (DL) models (i.e., LSTM, GRU, and BiLSTM) in terms of two performance indicators i.e. Root Means Squrae Error (RMSE) and Mean Absolute Error (MAE). BiLSTM model found to perform relatively better for all cities and GRU could predict better in places of large range of rainfall variation. Therefore, a combination of BiLSTM and GRU may be considered. In case of univariate rainfall prediction, the RMSE values range from 1.476 (Davangere) to 5.34 (Shillong), and MAE values range from 0.8232 (Jaipur) to 3.72 (Shillong). And, for multivariate forecasting, LSTM model performed better than others for all cities considered where RMSE values range from 1.376 (Tumukara) to 4.951 (Shillong) and MAE values range from 0.821 (Udaipur) to 3.057 (Shillong).