Authors - Aleta Fabregas, Nathanael Almazan, Jordan Garcia, Shaina Laman, Paolo Morato, Armin Coronado, Montaigne Molejon, Mariel Leo Violeta Abstract - The Philippines is frequently affected by tropical storms, typhoons, and flooding events that threaten communities located near major river systems. Accurate river level forecasting is essential for improving disaster preparedness and reducing flood-related risks. This study proposes RIVERCAST, a forecasting system that utilizes an Auto-Regressive Transformer with Kernel Principal Component Analysis (Kernel PCA) and Euclidean Kernel to predict Marikina River water levels across the Nangka, Sto. Niño, and Montalban monitoring stations. Meteorological, hydrological, and topographical datasets were collected from PAGASA, MMDA, DPWH, and OpenWeather API records from January 2012 to January 2023. Eighty percent of the collected records were allocated for training while the remaining twenty percent were utilized for testing. The pro-posed model was compared with the Transformer model developed by Xu et al. (2023) using rolling window testing and mean absolute error analysis. Results revealed that the proposed Auto-Regressive Transformer with Kernel PCA and Euclidean Kernel achieved an overall forecasting accuracy of 93.19%, outperforming the Bidirectional Transformer model, which obtained 92.57% accuracy. Findings further indicated that precipitation, rainfall intensity, and temperature significantly influenced forecasting performance, while humidity exhibited the least contribution. The developed model demonstrated reliable twelve-hour river level forecasting capability and may support flood preparedness and early warning initiatives within flood-prone communities along the Marikina River.
Open Zoom