Research On Flow Prediction Model Based On VMD-SSA-LSTM

Due to extreme climate change and human and social activities,runoff and other hydrological series are"non-stationary and non-linear",which poses a great challenge for runoff prediction.Efficient and accurate runoff forecasting not only provides a basis for scientific management and rational exploitation of water resources,but also guides the work of flood and drought prevention.This paper takes the hydrological data of Shalizhai hydrological station as a research example,and selects the flow time series from 2010 to 2019 as the research object.A flow prediction model combining data decomposition algorithm and model optimisation algorithm is proposed,and the efficiency as well as accuracy of the model is verified by comparing the error accuracy of different models.The research content and results of this paper are as follows:(1)The hydrographic elements,such as station cross-sections and climatic conditions,were analysed at the Shalizhai hydrographic station.The analysis shows that the water level and flow at Shalizhai hydrographic station are affected by regular siltation and flooding,and that the water level and flow correspond to each other in a counter-clockwise rope-and-loop curve.A simple analysis of water level,rainfall,flow and evaporation over the years shows that there is a correlation between rainfall,evaporation and flow.(2)For the complex rope-set curve relationship between water level and flow at this station,and the non-linear and non-smooth flow series,the modal decomposition algorithm is used to deal with the problem.By comparing the empirical modal decomposition algorithm(EMD)and the variational modal decomposition algorithm(VMD),it is found that the EMD algorithm has the problem of endpoint effect and modal confounding,while the VMD algorithm can effectively solve this problem.The VMD algorithm was chosen to decompose the data,using the observation method to select a modal number K of 12 and a penalty factor ofαof 2500.The parameters were determined and the original sequence of flow was decomposed into 12 IMF components of different frequencies in the simulation.A basis for subsequent modelling is provided.(3)Based on the mutual information method,the flow rate,water level,precipitation and evapotranspiration of the previous period were selected as the input variables of the model,and long and short-term memory neural network(LSTM),kernel limit learning machine(KELM),VMD-LSTM and VMD-KELM flow prediction models were established.The root mean square error(RMSE),mean absolute error(MAE),mean absolute percentage error(MAPE)and coefficient of determination(R-squared)were selected as evaluation indicators of the prediction accuracy of the models.Simulation results:the LSTM model is better than the KELM model in a single model,but the prediction effect of the four models compared is VMD-LSTM model>VMD-KELM model>LSTM model>KELM model,where the RMSE value of the VMD-LSTM model is 27.59 m~3/s,MAE value is 11.34 m~3/s,MAPE value is 17%,and The R~2 value is 94.14%,which is much better than the other three models,indicating that the flow prediction model combined with the data decomposition algorithm is more accurate.(4)The hyperparameters in the two models were optimised by the sparrow search algorithm(SSA)to improve the prediction performance of the models.The results show that compared with the LSTM model and the VMD-LSTM model,the RMSE of the VMD-SSA-LSTM model is reduced by 28.96 m~3/s and 18.39 m~3/s,respectively;the MAE is reduced by9.56 m~3/s and 4.87 m~3/s,respectively;the MAPE is increased by 16%and 7%,respectively;the R~2 is increased by 7.38%and 3.16%,respectively Further comparison between the VMD-SSA-LSTM model and the VMD-SSA-KELM model showed that the prediction accuracy of both models was higher,but the VMD-SSA-LSTM model fitted the original data better and had a better overall prediction capability at the peak of flow and during the flood period.