Prediction Of Precipitation In Jin’an District Of Fuzhou Based On The Informer Model

Rainfall prediction is of great significance to China’s agricultural and economic development,and its related research has attracted much attention in the fields of statistics,data mining,machine learning,hydrometeorology and other fields.Precipitation formation is affected by complex terrain and highly coupled atmospheric systems,which makes it difficult to characterize engineering and precipitation prediction.How to reasonably process meteorological data and excavate the law of predicting precipitation from it to improve the accuracy of precipitation prediction is a hot issue being overcome in related research.In view of the fact that the Informer model has excellent characteristics such as being good at extracting multi-scale features and fast prediction speed in the time series prediction task,the ERA5 reanalysis data of the European Medium-term Weather Forecast Center with a resolution of 0.5*0.5 is proposed as the research data,so as to achieve the refinement of precipitation areas and the fusion of meteorological elements to characterize the causes of precipitation as the starting point for the precipitation prediction of the future day based on the Informer model in Jin’an,Fuzhou.The main research contents and achievements are as follows:(1)What has been explored is that the laws of precipitation in Fujian Province in the past 21 years in time,space and scale by using descriptive statistical analysis methods and visualization.It is found that precipitation in Fujian Province is cyclical in time,and spatially more precipitation in northwest Fujian than in the coastal area of Fujian,which sits in the Wuyi Mountains,and it is found that the rainstorm and heavy rainstorm events in Fujian Province show a downward trend of fluctuating year by year due to the decrease in the frequency of typhoon landfall.In order to reduce the regional scale of precipitation prediction background,a clustering algorithm based on DTW distance value as a K-Means clustering criterion was used to cluster 50 districts and counties in Fujian Province,and the clustering results obtained four cluster sets that could be distinguished by latitude,coastality,mountain range orientation and mountain range height in the area,and the cluster set covering Fuzhou Jin’an was selected as the extracted regional scale.The VECM model was constructed for the purpose of reducing the characteristic dimension and fusing the feature information to screen the meteorological factors with the Granger causal relationship with the precipitation sequence,and fused many meteorological factors into five principal component features as the input characteristics of the precipitation prediction model by analyzing them as principal component characteristics.(2)In the empirical exploration part of the Informer model,four groups of control experiments are designed that can verify the effectiveness of the principal components of meteorological factors and multi-regional precipitation data.By predicting the precipitation of 769 groups of sequences in the next day,the precipitation in the past two years can be predicted one by one,and draw the comparison curve between the predicted precipitation value and the actual precipitation value in Jinan,Fuzhou in the past two years.The comparison results show that the experimental results of the four control groups are quite different,and the Informer model is feasible in the precipitation prediction of multi-regional and multi-characteristic data.It can accurately predict the precipitation trend and some strong precipitation events.At the same time,in order to show the superiority of the Informer model in the precipitation prediction task,this paper sets up a control group experiment between the Informer model and the LSTM model on the same data set.The experimental results show that the RMSE of the Informer model is6.82 and the MAE is 3.21,which is significantly better than the results of the LSTM model.