Deep Learning-Based Prediction and Forecasting of Extreme Typhoons in the Philippines: A Fully Connected Neural Network Approach

This study introduces a deep learning-driven method of forecasting super typhoons in the Philippines using a Fully Connected Neural Network (FCNN) to improve forecasting accuracy. Using historical meteorological data from 2009 to 2023, the required variables-such as wind speed, gustiness, central pressure, and precipitation, were included in the model. Feature engineering methods were added to the model, including rolling statistics, lagged variables, and interaction terms, to improve its capability to learn complex patterns. A walk-forward cross-validation strategy was employed to ensure valid temporal training and testing of the model. Experimental findings show that FCNN outperforms traditional and recurrent deep learning models with an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.9937, MAE of 2.7767, and RMSE of 3.3681. The proposed deep learning model effectively predicted the super typhoons in 2024 based on the historical records’ captured linear and non-linear patterns. The findings reflect typhoons' increased intensity and unpredictability due to climate change. The FCNN-based proposed model can be used to improve disaster preparedness and risk reduction efforts. Additional research will incorporate additional meteorological predictors to improve prediction accuracy.