Research On Seasonal Hydrological Dynamical Prediction Based On WRF-Hydro

Drought and flood are the serious natural disasters in China,affecting 121 million people,and causing 293.3 billion yuan in direct economic losses every year.In recent years,due to climate change,the frequency and intensity of drought and flood disasters have become higher,and it is urgent to carry out hydrological seasonal prediction to reduce the uncertainty of disaster prevention and protect people’s lives and property.Quantitative numerical prediction based on physical hydrological models is the current development direction of hydrological prediction,and it is also a worldwide problem with major challenges.In this paper,a seasonal dynamic prediction system is created based on the WRF-Hydro(Weather Research and Forecasting Model Hydrological modeling system)model,and the influence of the underlying surface characteristics,the initial state of the system,and the meteorological forcing on the system are studied.Due to the complexity of the climate system and the hydrological process,the output of the prediction system has some errors.And in order to solve this problem,a post-processing method based on machine learning is developed to improve the prediction performance of the system.The research content and results of this paper are as follows:(1)The influence of the underlying surface characteristics on hydrological simulationTaking the Xijiang River Basin as an example,the influence of underlying surface characteristics on hydrological simulation was studied by using two datasets of soil type from FAO(Food and Agriculture Organization)and GSDE(Global Soil Dataset for Earth System Science),and two datasets of land use from MODIS(MODerate-resolution Imaging Spectroradiometer)and CNLUCC(China Land Use Land Cover Remote Sensing Monitoring Dataset).It was found that soil type data and land use data had different effects on the simulation of different hydrological variables.It was found that the root mean square error(RMSE)of soil moisture simulated using GSDE soil type data and CNLUCC land use data was reduced by 3.6%compared with the RMSE of soil moisture simulated using FAO soil type data and MODIS land use data,and the Bias of simulated streamflow at most hydrological stations was reduced by more than 7%.Further applying these two datasets of underlying surface data to the northern region of China,it was found that the RMSE of the model simulated soil moisture decreases by29%.(2)The influence of the initial state on the simulation and prediction resultsIn this paper,the improvement of the initial state after data assimilation on the simulation results was studied.Taking the Xijiang River Basin as an example,the soil moisture data on the first day of each month was assimilated by the Ensemble Kalman filter assimilation method.After assimilation,the Bias of simulated soil moisture was reduced by 60%.The Nash efficiency coefficient(NSE)for simulated streamflow was 0.76 which was improved by 28%,the Bias was reduced by 75%,and the RMSE was reduced by 23%.Further applying assimilation to northern China,it was found that the RMSE of simulated soil moisture is reduced by 13%,and the NSE of simulated streamflow in the Yellow River Basin increased by0.22.After the initial state improvement(assimilation of satellite soil moisture),the NSE of predicted streamflow could increase by more than 40%.(3)Evaluation of meteorological forcing field and deviation correctionThis paper evaluated the applicability of the predicted precipitation(in China)of five mainstream climate models,and corrected the biases in the predicted precipitation.The five models predicted precipitation could show the spatial distribution characteristics of decreasing precipitation from southeast to northwest in China.The Bias of the SEAS5 model prediction was 29% lower than the prediction of other models.The correlation coefficient and anomalous symbol consistency rate of summer precipitation predicted from June and predicted from March to May were different by more than 0.3 and 10%.The bias correction method effectively improved the performance of mode prediction(Bias could reduce 90% and reach 0.01 mm/day),among which the empirical quantile mapping method was better.(4)Evaluation of seasonal hydrodynamic prediction systemBased on the above research,a seasonal hydrological prediction system was constructed,which had the ability to predict summer drought and flood,and the consistency rate of anomaly symbols could reach more than 80%.Among them,the correlation coefficient and anomalous symbol consistency rate of predicted streamflow based on SEAS5 were 0.2 and 7% higher than that predicted by other climate models.The prediction performance of the system for different river basins is different,and the correlation coefficient difference between different river basins could reach 0.3,and the anomalous symbol consistency rate could be up to 10%.The correlation coefficient and anomalous symbol consistency rate of summer streamflow predicted from June was 0.5 and 15% higher than that predicted from March to May.(5)Post-processing of simulated and predicted streamflow using the machine learningBased on convolutional neural network and variational mode decomposition,the streamflow post-processing method was constructed,and it was found that the post-processing method could reduce the model simulation error and improve the correlation coefficient of the simulated streamflow.Compared with the traditional post-processing method,the correlation coefficient of this method increased 0.2,and the NSE of this method increased 0.3.This postprocessing method was not only suitable for the calibrated hydrological models but also for the uncalibrated hydrological models.This method improved the accuracy of seasonal streamflow prediction,and the NSE was larger than 0.6 which was increased by more than 0.3 and RMSE reduced by up to 46%.