LSTM Neural Network Model Based On Hierarchical Clustering And Its Application On Precipitation Prediction Of Jiangsu Province

Atmospheric precipitation is one of the main sources of freshwater on the surface.The inter-annual uneven distribution of precipitation in many areas often causes droughts and floods in various places.Therefore,it is important to investigate the temporal and spatial variation of precipitation,which is important for agricultural production and residents' lives Meaning.This paper presents a prediction model that is comprehensively established by using cluster analysis,principal component analysis,and LSTM-based neural networks.It also predicts the precipitation time series of 62 meteorological observation stations in Jiangsu Province,with prediction efficiency and accuracy.Comprehensive performance is better.The focus of this article is how to simplify the analysis of 62 time series data,while improving the analysis efficiency and ensuring the prediction accuracy,and at the same time,the purpose of efficiently predicting precipitation at these 62 stations using a personal computer can be achieved.The first step of this paper uses cluster analysis to divide the precipitation sequence of 62 stations into several sub-classes with similar local characteristics.The second step uses the sequence compression method based on principal component analysis to find the dominant sequence of each sub-class.The third step uses the LSTM-based neural network time series prediction model predicts the dominant sequence of each subclass separately,and inversely calculates the predicted value of precipitation at each meteorological station in each subclass.Finally,the prediction established by calculating the standardized mean square error NMSE value pair.The accuracy of the model is evaluated.Based on the annual precipitation data of 62 meteorological observing stations in Jiangsu Province from 1961 to 2019,precipitation prediction is performed on the time series data of precipitation using R language and Python language in order to rationally allocate water resources and prevent meteorological disasters,and reduce economic losses Provide some reference.In the first step of this article,62 sites are divided into two categories by cluster analysis.The clustering results are very reasonable corresponding to the geographical distribution and climatic conditions of each site in Jiangsu Province.Based on the clustering results,this paper uses principal component analysis to extract There are two dominant features of each class,and a total of four dominant sequences are obtained.The total variance contribution rate of each class reaches more than 74%.Finally,the LSTM-based neural network time series prediction model is applied to the four dominant sequences.The predictions were made separately,and the predicted values of precipitation at each meteorological station in the two classes were calculated back.Finally,the model prediction accuracy was evaluated by calculating the average value of the NMSE of all stations in each class.The results of empirical research show that it is reasonable to classify the precipitation of Jiangsu Province into two types of north and south,which is very consistent with the geographical and climatic conditions of Jiangsu Province.The accuracy of the precipitation forecast based on the LSTM neural network prediction model is better.The average NMSE of multiple prediction results of similar sites is about 0.42 and 0.48,which are significantly less than 1.At the same time,the running time of the prediction program on the personal computer does not exceed 3 minutes,which is extremely efficient.The study provides an effective forecast method for future precipitation at meteorological observation stations in Jiangsu Province,and can provide some reference for government departments in cities and counties in the province to take effective measures to deal with natural disasters and rationally plan urban water resources.