Research On Runoff Simulation Based On CNN-BiLSTM-Attention Precipitation Fusion Algorithm

Precipitation,as one of the most fundamental meteorological data in the water cycle,plays a decisive factor in the variation of watershed runoff.In recent years,the frequent occurrence of extreme weather events brought about by global climate change has intensified the frequency,intensity,and harm of various floods,thereby having a huge impact on China’s agricultural development.Among all the key elements of flood prediction,the most important one is the uncertainty of precipitation space.Therefore,obtaining high-precision precipitation data for driving hydrological forecasting models is of great significance for hydrological process analysis and flood disaster prediction.Compared to the observed precipitation data from traditional ground stations,although satellite remote sensing precipitation has poor accuracy,it can better reflect the spatial distribution of precipitation within the basin.Therefore,conducting fusion based on satellite ground station precipitation information can effectively improve the accuracy of precipitation estimation by learning from the strengths and weaknesses of the two types of data.Currently,in order to obtain high-precision precipitation data,scholars from various countries have proposed many methods for integrating satellite and ground station precipitation,and precipitation fusion based on deep learning is also a hot topic in current research.Based on this,the paper takes the Jinjiang River basin as the research area,optimizes the deep learning model built by predecessors,uses the optimized CNN-Bi LSTMAttention model and Kalman filter to carry out the fusion research of satellite ground station precipitation data,and uses the fused precipitation data to drive the SWAT distributed hydrological model to simulate the runoff process simulation,Explore the advantages and disadvantages of each fusion algorithm and its applicability in hydrological simulation.The experimental results on the Jinjiang River Basin show that:(1)The CNN-Bi LSTMAttention model proposed in this paper has improved in multiple indicators compared to the previous CNN-Bi LSTM model,especially the significant increase in relative deviation from-0.077 to-0.009,an increase of 98.83%,a root mean square error increase of 1.05%,and a correlation coefficient increase of 0.55%.(2)Fusion precipitation can improve the precipitation accuracy of the original TRMM data in the study area,with Kalman filter fusion precipitation data having the highest correlation,which is 15.9% higher than the original TRMM.However,the spatial distribution throughout the entire watershed does not match the actual situation;However,the CNN-Bi LSTM-Attention model can better reflect the uncertainty of precipitation space,and is better than the Kalman filter in terms of relative deviation and average absolute error indicators,which are increased by 46.8% and 12.0%respectively compared with the original TRMM.(3)Various precipitation data were input into the SWAT distributed hydrological model to simulate the runoff process simulation of the driving basin.The results show that the accuracy of the CNN-Bi LSTM-Attention integrated precipitation simulation is the highest.Compared with the simulation results of the original TRMM and the measured precipitation data of the meteorological station,the Nash coefficient in the verification period in the daily scale is increased from 0.56 and 0.31 to 0.63,respectively,12.5% and 103.2%,and can better reflect the extreme value of runoff,It can be used for flood disaster prediction to prevent the impact of flood disasters on the development of agriculture and rural economy in China and the safety of people’s property.