Research On Runoff Prediction Model Based On Variational Mode Decomposition And Recurrent Neural Network

River runoff is affected by many factors such as climate change,man-made activities and underlying surfaces,and presents characteristics such as nonlinearity,chaos,and randomness,which brings great challenges to runoff prediction.Runoff prediction plays an important guiding role in reservoir operation,flood control and drought resistance.Scientific,accurate and timely runoff forecasting can not only reduce the pressure of flood control and reduce the losses caused by floods during the flood season,but also help scientific planning and rational use of the huge hydropower resources brought to us by nature,and fully tap water resources.Economic and social benefits.Therefore,this article proposes a combination of signal processing technology based on frequency conversion modal decomposition and cyclic neural network technology for the problem of low accuracy of traditional mid-and long-term runoff forecasting methods for time series with high volatility and weak correlation between the front and back.The combined model adopts the method of "decomposition-prediction-reconstruction" to predict the runoff,and also introduces the particle swarm algorithm to solve the problem of low efficiency in determining the hyperparameters of the neural network.The case study shows that the runoff sequence at Hankou Station Compared with traditional methods,the prediction accuracy has been improved,and it has the potential to be applied in actual engineering.The specific research content and innovative research results of this paper are as follows.(1)In the data processing link,this paper treats the original runoff sequence as a random signal for processing,which has the characteristics of nonlinearity,chaos,large randomness and disorder.This paper proposes to use the current more advanced frequency conversion modal decomposition in the field of signal processing.Technology(Variational Mode Decomposition,VMD)is used to process the original runoff sequence.The research results show that the frequency conversion modal decomposition technology has a good decomposition effect.The components obtained by the decomposition are stable and orderly,with strong regularity and low volatility,which is conducive to later construction.mold.For the components after decomposition and prediction,a weighted recombination method is proposed to preserve the characteristics of the original signal as much as possible and improve the accuracy of the prediction results.The case study results show that this method can well retain the characteristics of the original runoff sequence.(2)This article compares the more commonly used Recurrent Neural Network(RNN),Long Short-Term Memory(LSTM),and Gated Recurrent Unit(GRU)in runoff sequence simulations.The effect of and the accuracy of the prediction results have been found through research that the long and short-term memory network performs better under the same conditions and is more suitable for runoff simulation.Therefore,this paper selects this network model for research.At the same time,in view of the low efficiency of the trial algorithm and the difficulty in finding the optimal solution for the neural network parameter setting,this paper introduces the particle swarm optimization(PSO)to optimize the model parameters,and introduces a callback function mechanism to improve training effectiveness.(3)Apply the combined model to case studies for verification,use frequency conversion modal decomposition technology to decompose the time series to obtain different modal component sequences,and use an improved long and short-term memory network to perform training simulations based on the component sequences.The component prediction results are reorganized by the K-means weighted sum method to obtain the actual data prediction results.The case study results show that the certainty coefficient reaches 0.965 when the prediction period is 10 days,and the relative error percentage is only 0.071;the prediction period is determined when the prediction period is one month.The performance coefficient reaches 0.934,and the relative error percentage is only 0.087.Compared with the LSTM model and the EMD-LSTM model,there is a big improvement,and it has the potential for practical application.