Real-Time and Long-term Forecasting Tools for Flood Management and Quick Response

• Main Authors: De Lara-Tuprio, Elvira, Queaño, Karlo L, Tuprio, Elijah Patrick, Alvarado, Renz, Bandong, Annie, Gatdula, Carleen, Cayas, Ryan Roi, Doria, Christell, Sombilon, Junifer
• 格式: text
• 出版: Archīum Ateneo 2024
• 主題: Computer Sciences; Data Science; Mathematics; Statistics and Probability
• 在線閱讀: https://archium.ateneo.edu/mathematics-faculty-pubs/295; https://doi.org/10.1063/5.0230616

The forecasting tools presented in this paper were applied in the flood-prone city of Marikina, a highly urbanized city located in the National Capital Region (NCR) of the Philippines. The city is traversed by Marikina River, which inundates nearby communities during heavy downpours. Marikina City has a long history of flooding. Perhaps the most devastating among these not only in Marikina but in the entire NCR is the Typhoon Ondoy which resulted in hundreds of deaths and huge amount of damage to properties. Despite the constant threat of flooding, residents stay in the city and have learned to adapt to the realities of the environmental hazards. For its part, the local government of Marikina has developed an alert system to warn residents of impending flood during heavy rains. Depending on the water level recorded at Sto. Niño station, which is located at a strategic point along Marikina River, residents are advised to either prepare or go to evacuation centers. The disaster response efforts would be best supplemented by real-time forecasts of water level to give both the local government and the residents ample time to prepare. This is the motivation for this study. In this paper, machine learning algorithms such as random forest, artificial neural network and support vector machine were applied on historical data of rainfall and water level recorded at different stations to develop tools to forecast water level at Sto. Niño at least one hour ahead. Results indicated that the random forest model performed better overall than other forecasting models since it has the lowest error metric and the highest correlation coefficient. To supplement the disaster management program and for budget planning purposes, a probabilistic approach was used to obtain long-term forecast in terms of annual frequency and duration of water level reaching at least a specified height.