|Recently, in Japan, many people have died from delays in evacuation during flash floods in small and medium-sized rivers for which flood forecasts have not been implemented. Damage caused by such flash floods occurs not only in Japan but also all over the world, and countermeasures are needed. For these reason, this study aims to develop flood prediction system that is fast in calculation, low-cost, simple, and adequately accurate and then standardize the provision of water-level prediction information for small and medium-sized rivers. Specifically, to improve the accuracy of the water-level prediction model, the particle filter method was applied to continuously calibrate the RRI (Rainfall-Runoff-Inundation) model using real-time observed water levels. Additionally, the SCE-UA method was applied to the model as an automatic parameter tuning function to reduce the effort of model construction. The system was built and verified for 216 events on 125 rivers, and the target lead time was obtained for approximately 50% of the floods. Additionally, when actual rainfall was used as the predicted rainfall, the target lead time was obtained for about 85% of the floods. We could develop a flash flood prediction system that can be implemented in small and medium-sized rivers.