Research On VMD-OOYO-CNN-GRU Based Hybrid Deep Learning Model And Application

Runoff is an important part of water resources,and runoff forecasting can provide decision support for efficient utilization and scientific management of water resources.By forecasting runoff,water resources can be reasonably dispatched and the efficiency of water resources utilization can be improved.Runoff forecasting has been an important research topic in the field of hydrology.Previous runoff forecasting methods are mainly based on hydrological models and statistical methods,but these methods face many difficulties,such as the past runoff forecasting methods often require a large amount of hydrological data and model parameters,high model complexity,and are not easy to operate and maintain.In contrast,data-driven artificial intelligence technologies can use powerful learning capabilities to automatically extract features from the data,making the models less complex.In recent years,the development of deep learning technology has provided a new way for runoff forecasting.Therefore,the study of deep learning technology in runoff forecasting is of great practical significance and academic value.In this paper,based on deep learning technology,short-term flood forecasting is carried out in the Pei River basin,and medium-and long-term monthly runoff forecasting is carried out in the Zhengyixia hydrological station and the Fengle River hydrological station in the Heihe River basin,respectively.The main research contents and results are as follows:(1)In order to solve the problem that flood forecasting is more complicated when using process-driven models,this paper proposes to construct a deep neural network model for flood forecasting.Using the Pei River as the study basin,a single convolutional neural network(CNN),long and short-term memory neural network(LSTM),gated recurrent unit neural network(GRU),and their combined models CNN-LSTM and CNN-GRU are constructed for rainfall runoff flood simulation.Using the relative error of flood peak,peak present time error,correlation coefficient and Nash coefficient as evaluation indexes,the results show that all the above five deep learning models show certain reliability in flood simulation,among which,CNN outperforms the other four models in predicting flood flow.(2)To address the problem of improving the accuracy of runoff time series prediction,this paper proposes to apply deep learning techniques in machine learning,which are superior to traditional models,to monthly runoff prediction,and to couple convolutional neural networks with recurrent neural networks.In this hybrid prediction model,the advantages of both are fully exploited,as the spatial features of the data are first extracted by the convolutional neural network model and then the extracted features are passed to the next layer;while the recurrent neural network has a long memory capability to obtain important information from the previous data and effectively simulate complex long time series.The CNN,LSTM,GRU,CNN-LSTM and CNN-GRU models are used for monthly runoff forecasting at the hydrological stations of Zhengyixia and Fengle River in the Heihe River basin,for example.Using root mean square error,mean relative error,and correlation coefficient as evaluation indexes,the results show that the coupled forecasting models have better fitting effects than single forecasting models,among which the CNN-GRU model has the best forecasting performance.(3)To solve the problems of low efficiency in determining model parameters by manual trial calculation and low forecast accuracy due to nonlinear non-smoothness of runoff time series in runoff forecasting,OOYO-CNN model for flood forecasting and VMD-OOYO-CNN-GRU model for monthly runoff forecasting are proposed,respectively.Through example analysis,it is concluded that data decomposition and model parameter optimization are effective strategies to improve deep learning.OOYO can automatically find the optimal parameter combination,avoiding the problem of tedious and time-consuming manual adjustment of parameters;VMD decomposes the original runoff series into multiple components with low complexity and high periodicity,reducing the complexity of the original runoff series and providing a good basis for the model prediction.