The distributed hydrological model represented by the Grid Xin'anjiang model was used for flood simulation and forecasting in the Qinhuai River Basin. The spatiotemporal simulation is highly refined in the middle and upper reaches, but its application accuracy is limited in the downstream reaches. In this study, the Qianhancun (Qin) station is used as the node, the grid Xin'anjiang model is used to calculate the runoff in the upstream, the hydrodynamic model is used to simulate the water level of the tidal river section affected by the operation of the gate pump in the downstream, and the K-nearest neighbor (KNN) algorithm and feedback method is adopted to correct the simulation results in real time. After the correction of the KNN method, the coefficient of certainty is 0.718-0.975, which meets the forecast accuracy requirements of Class B.

To explore the prediction accuracy and physical properties of the grid water drop confluence method, and verify its applicability in areas lacking of or without data, based on the grid model of Xin’anjiang model, the flow yield calculation is carried out, and the confluence model based on the grid water drops is constructed. Tunxi and its nested basin in the humid area and Dongwan and its nested basin in the semi-arid area are taken as the study basins, and the nested sub-basin is assumed to be the area lacking of data. The simulation validation of Tunxi and the inner sub-basin of Dongwan is carried out without re-setting parameters, and the results indicate that in the same watershed, the comprehensive simulation accuracy of the grid water drop confluence model decreases with the decrease of the subwatershed area. The simulation accuracy of the model in Tunxi River basin reaches grade A and the simulation accuracy in its sub-basins basically meets the requirements, while the simulation accuracy of Dongwan reaches Class B and the accuracy of its sub-basins reaches Class C except for the pass rate of flood peak. This study validates the applicability of the grid-water-drop based confluence method in the area lacking of or without data.

To address the problem that the parameters of the LSTM model do not transpose well in nested watersheds with inadequate records of hydrological observations, a runoff modulus-based LSTM(RM-LSTM) was developed using the Tunxi Watershed as the study watershed and its nested sub-basins as the reference watersheds, with the consideration of the influence of area factor on the parameter transfer method. The flood simulations in the reference watersheds were carried out using the parameter transfer method. The results showed that the RM-LSTM model was well applied in the study watershed, with a determination coefficient of 0.87 for the simulation results, and a qualification rate of 85.0% or more for flood peak, flood volume and peak present time. In the reference watersheds, the RM-LSTM model showed a marked improvement in terms of the determination coefficient, qualification rate of flood peak, qualification rate of flood volume, and qualification rate of peak time after the parameter transplantation, compared to the results obtained from the LSTM model. The RM-LSTM model can better take into account the influence of watershed area changes on the parameter transfer method.

In order to better apply the Grid-Xin’anjiang(Grid-XAJ)model in the flood simulation and forecasting in areas affected by reservoir discharge, the effect of reservoir discharge on the model confluence calculation was compared between the segmented Muskingum(MK) method, grid-based MK method and grid-based Muskingum-Cunge-Todini (MCT) method, based on the real geographical location of the reservoir in the Hengpaitou Watershed. The results show that the flood peak simulation of the three schemes is good, and the average pass rate of all the floods in the periodic and verification period of each scheme rate is 100%. The average pass rate of runoff depth in the grid-based MK method and segmented MK method is 76.9%, which is lower than that of the grid-based MCT method, 84.6%. Compared with the segmented MK method, grid-based MCT method and grid-based MK method can reflect the influence of upstream reservoir discharge on the spatial and temporal distribution of discharge in the catchment process. Grid-based MCT method further considers the real situation of the underlying river channel on the basis of grid-based MK method. In general, grid-based MCT method has strong adaptability and reliability for flood forecasting in basins affected by reservoir drainage.