In high-density urban areas, flooding affects a large number of people. A rapidly implementable nonstructural measure is the development of an early flood warning mechanism based on observations from ground-based rainfall stations, especially where radars are not yet installed. To increase the lead time for issuing warnings, a reliable short-term rainfall forecasting model is required, specifically for fast-responding urban catchments where the time of concentration is less than 45min. With this objective, a rainfall forecasting methodology has been developed using the least-squares support vector machine (LS-SVM) and the probabilistic global search-Lausanne (PGSL) technique. The study's focus was Mumbai, which receives all of its annual rainfall of 2,430mm during June to September. The Mumbai storm drainage system had been designed to drain 25 mm/h rainfall and is being upgraded for 50 mm/h. Storms less than 25 mm/h do not cause flooding; hence, the proposed methodology was developed using rainfall events greater than 25 mm/h. The model developed in this study has been evaluated using statistical performance criteria for four observed high-intensity rainfall events (50-100 mm/h) during 2011. The results indicate that the proposed methodology using LS-SVM and PGSL can effectively forecast high-intensity rainfall with lead time from 5 to 20min. This study improves upon the 1-h forecast limitation of earlier studies and has the potential to forecast rainfall in real time, especially where radar data are not available.