[Objective] In China, flood detention basins serve as the “last line of defense” for flood control. In major flood events, if the incoming flood volume still exceeds the discharge capacity of river channels already with reasonable reservoir operation, flood detention basins will be activated to store and detain floodwater and prevent catastrophic levee breaches. During other periods, limited production activities are permitted within these areas. Currently, flood detention basins in the middle and lower reaches of the Yangtze River accommodate a large residential population, and flood diversion would cause substantial inundation losses. This imposes higher demands on scientific operation. Effectively predicting the water level reduction effect of flood detention basin operation is fundamental for real-time operation. While process-based hydrodynamic models are challenged by a heavy computational burden and does not meet real-time computational requirement, data-driven models are lack of real data. This paper proposes a deep learning-based surrogate modeling framework to forecast the influence of flood detention basin operations on mainstream water level. [Methods] Using calibrated and validated hydrodynamic models coupled with the Monte-Carlo simulation method, a large number of operational scenarios under varying hydrological and engineering conditions were simulated to generate training samples. Based on the generated samples, prediction models are constructed. Deep learning models with outstanding performance in time series prediction, including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), are selected. The prediction accuracies of these models are compared to identify the deep learning model most suitable for predicting the impact of flood detention basin operation. Various combinations of factors related to the operation strategy of the flood detention basin, upstream inflow, downstream backwater effects, and flood rise and recession conditions are considered. [Results] In the case study of the Honghudong Detention Basin (HDB) operation on water level reduction at Hankou station, the Long Short-Term Memory network demonstrated the highest prediction accuracy and stability compared to GRU and CNN. The best combination of predictive factors includes the flood diversion discharge, diversion duration, pre-diversion water level process at Chenglingji Station, water level process at Hankou Station, flow process at Luoshan Station, and flow process of the Han River over three days. Considering the gain in prediction accuracy brought by adding input factors, a predictive model can also be constructed using the flood diversion discharge, diversion duration, and the water level process at Hankou Station for the three days prior to the diversion, under data-scarce conditions. Using the selected combination of optimal predictors, a Nash-Sutcliffe efficiency coefficient of 0.983 9 and a root mean square error of 0.024 cm in predicting the water level reduction process at Hankou Station over seven days following diversion are achieved. Meanwhile, the computation time was reduced from 1.5 minutes to less than 1 second compared to the one-dimensional hydrodynamic model. [Conclusion] The method not only provides high predictive accuracy with fewer computation times but also requires fewer input factors compared to hydrodynamic models, significantly enhancing practical applicability. The proposed method is well-suited for coupling with an operation model, enabling more informed and timely real-time operation decisions. The proposed method helps basin authorities to improve detention basin real-time operation.