Deep Learning for Cyclone Track Forecasting

Cyclone-track forecasting is important for disaster preparation, but storm movement is sequential, nonlinear and affected by complex environmental behaviour. The research examined how different deep-learning architectures performed when learning trajectory patterns from historical cyclone observations.

Akmam contributed to problem formulation, experimentation, comparative analysis, result interpretation and publication-ready technical communication as part of the research team.

• Latitude
• Longitude
• Maximum wind speed
• Radius information
• Time from initial observation
• Storm-year information
• Sequential trajectory records

The research demonstrated that recurrent neural architectures could learn useful trajectory patterns from historical cyclone sequences, while performance varied by architecture and storm context. Using kilometre-based distance error made the results more understandable than relying only on abstract loss values.

The project shows how scientific machine learning can support analysis of physical and geospatial phenomena. The underlying skills also transfer to operational forecasting, predictive maintenance, movement modelling and other sequential prediction problems.

• Historical-data limitations
• Rare and changing storm behaviour
• Limited atmospheric variables
• Prediction uncertainty
• Generalisation across regions
• Deep-learning predictions should complement rather than replace meteorological systems