Author ORCID Identifier
https://orcid.org/0000-0002-9226-9773
Biosketch
J. Senthil Kumar is currently working as an Assistant Professor in the Department of Computer Science at SASTRA Deemed University, Kumbakonam, Tamil Nadu, India. He teaches courses on programming languages, data structures, and data science. His research interests include data analytics, machine learning, and atmospheric sciences.
He completed his Bachelor of Science in Computer Science from Bharathidasan University, his Master of Science from Alagappa University, and his Master of Technology in Computer Science and Engineering from SASTRA Deemed University. He recently earned his Doctor of Philosophy (Ph.D.) in Computer Science and Engineering.
He has 15 years of teaching experience and 6 years of research experience. He has guided many undergraduate and postgraduate students in their academic projects. He has published nearly 15 research articles in SCI and Scopus-indexed journals and has presented papers at national and international conferences.
Date of Award
17-8-2025
Document Type
Thesis
School
School of Computing
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.V.Venkataraman
Second Advisor
Dr.S.Meganathan
Keywords
Tropical Cyclone Intensity Prediction, Updated empirical model, Deep Multilayer Perceptive Classifier Model, Deep Learning Model, Long Short Term Memory – Cat Swarm Optimization Model
Abstract
The Bay of Bengal region's coastlines have been badly devastated by tropical cyclones, as the region experiences an average of five to six cyclones per year, with about two to three of these intensifying into tropical storms or severe cyclones. Thus it necessitates to study the accurate and efficient forecasting of their intensity to improve preparedness and response to natural disasters. The present study compares and examines three distinct approaches to cyclone intensity prediction using historical datasets from 1998 to 2020: hybrid optimisation, deep learning-based, and empirical approaches.The predicted accuracy, computational effectiveness, and feasibility for real-time scenarios of each model are assessed.
The first study has proposed an enhanced empirical model that used 12-hour wind speed variations and a correction procedure to improve forecast accuracy across 72-hour periods. The model has significantly reduced both the absolute error (from 21.53 to 12.04 knots) and RMSE (from 25.29 to 21.62 knots), improving previous empirical methods. Its affordability and simplicity of usage make it perfect for easy forecasting. It relies on a limited set of criteria, however, limits its ability to predict long-duration and high-intensity cyclone activities.
To address these limitations, the second study introduced LEGEMP, a deep multilayer perceptive classification model that combines feature selection using Herfindahl correlation with classification using Jaccardized similarity-based learning. The LEGEMP framework has an average prediction accuracy of 85.97%, which is significantly better than existing models. The soft-step activation function and Nesterov gradient descent for training has significantly improved the efficiency of the model by reducing the average error rate to 14.15%. Limited observational data and dynamic unpredictability in cyclone activity have been better managed by this approach.
In order to enhance cyclone prediction, the third study used a hybrid deep learning model developed with Long Short-Term Memory (LSTM) networks tuned by Cat Swarm Optimization (CSO). The model successfully demonstrated long-term temporal dependencies without gradient loss, with a high accuracy of 91.38%. It has the lowest MAE (11.32 knots) and RMSE (13.76 knots), as well as the highest Pierce Skill Score (0.79) and Hit Rate (0.97). Furthermore, the LSTM-CSO model has a low false alarm rate (0.20), indicating its robustness and usefulness in real-world cyclone forecasting scenarios.
These models have collectively demonstrated the progress of cyclone intensity forecasting methods. Even if the empirical method yields a fast and intelligible response, the deep learning and hybrid optimization frameworks have significantly increased accuracy and scalability. This thesis emphasizes the necessity of integrating data-driven intelligence with accurate modelling techniques to improve early warning systems and strengthen coastal resilience against tropical cyclone dangers.
Recommended Citation
J, Senthil Kumar Mr, "Cyclone Intensity Prediction in the Bay of Bengal Using Deep Learning Methods" (2025). Theses and Dissertations. 162.
https://knowledgeconnect.sastra.edu/theses/162