Rainfall Nowcasting Using Machine Learning from Observed Rainfall Data

Abstract

Bangkok, covering an area of 1,568.7 km², operates a dense network of 130 rainfall monitoring stations, each recording data every 5 minutes and covering approximately 12.17 km². Given the city’s frequent occurrence of intense rainfall at the onset of storms, and the need for at least one hour to drain floodwater one meter deep, short-term rainfall forecasting is crucial for urban water management. This study applies the Random Forest algorithm to forecast rainfall 120 minutes ahead, at 15-minute intervals, using data from rainfall stations. The Phaya Thai station, located centrally, achieved the highest detection performance (0.95) when using 540 minutes of historical data. Twelve input scenarios were tested based on three station selection strategies: single station, nearby stations, and randomly selected stations. During the rainy season (12:00–24:00), the detection rates ranged from 0.95 to 0.99. Models trained on data across all time periods achieved forecast accuracies between 0.98 and 0.99. Using 30 nearby stations yielded the best detection capability (0.95), while random station selection produced the highest overall accuracy (0.99). These findings highlight the effectiveness of using spatially distributed rainfall data with Random Forest for high-resolution short-term rainfall forecasting in urban areas.