Research On Hydrological Forecast Method Based On Deep Learning

Floods result in substantial damage throughout the world every year.An accurate predictions of floods can significantly alleviate the loss of lives and properties.As a time series,hydrological flow elements are affected by natural factors and human factors in the formation process,which presents very complex characteristics of certainty and uncertainty.It is difficult to make accurate and reliable prediction results when traditional prediction methods are used to solve complex problems such as hydrological flow prediction.Therefore,how to effectively obtain the effective characteristics of hydrological series,construct a suitable hydrological flow prediction model,and improve the accuracy of medium and longterm hydrological flow prediction is the focus and difficulty of hydrological research.In this paper,the upper and middle reaches of the Huaihe River Basin Xixian station is the research object,and the hydrological historical data of Xixian sub basin from 2011 to 2018 is used to study and design an effective medium and long-term hydrological flow prediction model.To avoid the issues of hydrological feature extraction,the complex physical relationship in the process of hydrological process which is difficult to be simulated by traditional models,and the independence of input and output of the flow prediction model of artificial neural network,we use the feature extraction algorithm of mutual information analysis,long short-term memory(LSTM)neural network,gated recurrent unit(GRU)neural network and attention mechanism to build LSTM cyclic model with mutual information for hydrology forecasting and Bi GRU multi-step prediction model based on attention mechanism.Specific research work is as follows::(1)As the change of hydrological flow is a time series process,it is greatly affected by the hydrological elements in the early stage in the process of flow prediction,and mutual information can better reflect the complex relationship between the input characteristic factors and flow.In this paper,the mutual information analysis algorithm is used to dynamically obtain the effective hydrological input characteristics,which makes a basic research for the subsequent modeling.(2)A LSTM cyclic model with mutual information for hydrology forecasting is proposed.On the basis of mutual information analysis,aiming at the problems that the traditional model is difficult to effectively simulate the complex relationship in the hydrological process,the input and output of the current artificial neural network are mutually independent,and the accuracy of medium and long-term hydrological flow prediction is reduced,a LSTM cyclic model with mutual information for hydrology forecasting is constructed by making full use of the time series processing ability of LSTM neural network.The results show that the model can automatically and effectively obtain the effective correlation characteristics of the input sequence,and has a good prediction effect in the process of long-term traffic prediction.(3)A Bi GRU multi-step prediction model based on attention mechanism is proposed.As the attention mechanism can automatically match the weight parameters of different time steps of the input sequence,extract more useful information,and the GRU neural network is simpler than the LSTM network structure,which can greatly improve the training efficiency.Therefore,compared with the previous model,from another point of view,a Bi GRU multistep prediction model based on attention mechanism is constructed.When the model is used for multi-step prediction,it not only has high accuracy of flood flow prediction,but also has faster prediction speed in training set and test set.(4)The experimental comparative analysis and related evaluation of the two models are designed and implemented.In this paper,tensorflow and keras framework are used in the experiment to realize the previous popular flood flow prediction model,and compared with the above-mentioned model.The results show that the model proposed in this paper is better than other models.In addition,according to the requirements of the national hydrological forecast standard,the proposed model can achieve good results in the evaluation of the peak arrival time error,peak value error and root mean square error,which has a certain practical significance for flood control and disaster alleviation.