Abstract: Real-time error correction serves as a crucial means to enhance the accuracy of flood forecast.To address the problem of poor accuracy and stability of traditional error correction models, we incorporate machine learning technology into the training process of the error sequence mapping function and propose an intelligent error correction method based on SPGA-XGBoost.This method initially constructs an error sequence based on the predicted and measured values of the traditional hydrological forecasting model as the input of the error correction model, and constructs a error correction model to fully explore the nonlinear relationship of the error sequence by the deep learning algorithm XGBoost.Moreover, a hybrid genetic optimization algorithm SPGA which combines particle swarm optimization algorithm and simulated annealing algorithm, is proposed to optimize the hyperparameters of XGBoost model, thereby it can better mine the timing characteristics of error sequences and improve the accuracy of error correction.The experimental results of Luoshan Station on Changjiang River indicate that the RMSE and MAE of the model are decreased by 0.440 m and 0.356 m respectively, and the NSE is improved by 0.016 compared with the uncorrected model, which is superior to the STGCN, GBDT and KNN.The SPGA-XGBoost model can fully discover the correlations within the error sequence and improve the accuracy of water level forecast, and has good applicability and application prospect.