Flood Prediction Based On Machine Learning And Fusion Of Space-time Features

Flood disasters occur frequently in China,which seriously affect the safety of people's lives and property,causing huge economic losses and population casualties every year.If we can achieve high precision forecast on the peak of flood and the arrival time of flood,and guide people in disaster areas to avoid emergencies in advance,the loss can be greatly reduced.Traditional flood prediction models involve the physical process of flood generation,and are faced with difficulties such as complicated calculation,high maintenance cost and long development cycle.Even for professionals,after the traditional model is transplanted to the new basin,the determination of more than a dozen main parameters needs a long time of inference and field measurement.With the development of computer technology,especially the improvement of computer computing power,computer science technology led by machine learning has been widely applied.Using the reserved historical hydrological data,the effective flood prediction model built by machine learning technology has achieved rich results.However,the existing flood prediction models based on machine learning and deep learning technology are insufficient in data feature mining,especially for the rainfall data,only the obvious time series features are mined,while the inherent spatial distribution features are ignored.This phenomenon of insufficient feature mining makes the overall flood model prediction accuracy low,and the phenomenon of over-fitting is easy to appear.In this paper,two types of model was designed using the machine learning techniques in order to get the flood peak and flood time accurately forecast,on the one hand,avoids the problems in the traditional model the parameters of the complex rate,on the other hand,with the introduction of convolution operation.It effectively solves the problem of insufficient spatial distribution feature mining of rainfall data in machine learning model,in this paper,the main work is shown below.1)In this paper,the topographic environment of the study area was firstly studied,and the actual geospatial distribution among hydrological stations was clarified.The original flow and rainfall data are preprocessed,missing value completion,data normalization,data feature correlation analysis and other operations are carried out to provide data support for model design and training.2)A space-time feature fusion analysis model CAe-RNN(Convolutional Autoencoder Recurrent Neural Network)is proposed based on the Convolutional Autoencoder structure.In view of the lack of spatial feature mining of rainfall data by existing data-driven models,this paper introduces the Convolutional Autoencoder structure to carry out spatial feature mining and constructs a CAe-RNN model.The specific work is as follows: A.The rainfall data in the study area is processed by two-dimensional meshing,and the generated meshed rainfall is analyzed by using convolutional neural network to extract the spatial features of rainfall data.B.Pretraining is adopted to correct parameters of the convolutional autoencoder to effectively avoid the problem of slow training speed caused by the surge of data volume after mesh and excessive network parameters.C.The extracted spatial features of rainfall are combined according to the sequence of time steps,and their relationship in time sequence is further mined through the cyclic neural network to realize the fusion analysis of space-time features.D.Combined with the historical river flow information,forecast and simulate the flood flow data in a future period.3)It innovatively introduces concepts related to Graph Neural Network into the field of flood prediction,and proposes a spatial-temporal feature integration analysis model of Graph Convolution-Recurrent Neural Network(GC-RNN)based on frequency domain and a spatial-temporal feature integration analysis model of Graph Attention Network(GA-RNN)based on space domain.The above two models can effectively solve the problem of insufficient spatial distribution mining of existing rainfall data and simulate the actual rainfall convergence process to a very high degree.The specific work contents are as follows:A.The original data is abstracted as graph data,the actual rain gauge station and river channel are abstracted as nodes and edges in the graph data,and the rainfall of each station is taken as the attribute value of nodes in the corresponding graph data.Examine the actual flow of the river to construct the direction of the edges between the different nodes.B.Build deep graph convolutional neural network and deep graph attention neural network to extract and mine spatial distribution features of generated graph data.C.The spatial distribution features extracted from the graph data are further mined through the cyclic neural network to realize the fusion of spatial and temporal features of rainfall.D.Combined with the historical river flow information,forecast and simulate the flood flow data for a period of time in the future.In order to test the practical effect of the model constructed in this paper,experiments were conducted based on historical rainfall and flow data from 50 hydrological stations within Xixian County,Henan Province.In terms of the prediction effectiveness of the annual river flow,both the peak flood and flood arrival time prediction errors meet the hydrological production specifications.In order to carry out the model forecast rating evaluation,several typical flood processes in history were selected for further assessment,and CAe-RNN,GARNN,and GC-RNN were rated as Class A,Class B,and Class C flood forecasting models,respectively,all of which can be used to direct the actual flood forecasting operations.In order to demonstrate the advantages brought by the extraction of rainfall spatial features by the model proposed in this paper,the experimental performance is compared with the existing models without spatial feature extraction,and the results show that the model proposed in this paper reduces the prediction error by 15% to 40%,which proves the advancement and reasonableness of the described model design idea.