Flood forecasting is an effective nonengineering measure to reduce the economic losses brought by floods. Accurately forecasting flood is one of the key technical challenges in hydrological field. There are flood forecasting models based on physical mechanisms used, but the accuracy and efficiency of forecasts need to be improved. At present, forecasting models constructed by using deep learning techniques have been developed rapidly. This article comprehensively reviews the principles and characteristics of deep learning models that have been applied in the field of flood forecasting and summarizes their application progresses and problems in the quantitative and probabilistic flood forecasting. In addition, this article explores the relevance and application prospects of deep learning models combined with flood physics models, particularly in the parameterization of physical processes, interpretability studies, and error correction of flood forecasting models. The results suggest that the deep coupling of deep learning technology with physical models is the developing direction of deep learning models in the future. It will be an important development paradigm for the time series prediction of flood, and also an important research component to achieve intelligent water resource management in the future. Finally, to better apply the deep learning technology in the field of flood forecasting, some thoughts on the difficulties of deep learning in flood prediction are given and corresponding solutions are proposed for the current challenging problems.