Lattic Forecast Revision Of Meteorological Elements Based On Deep Learning

With the arrival of the era of big data,the awareness of data collection has been strengthened,which makes data mining more useful in various fields.As an important tool of data mining,machine learning is well known by the public,and deep learning is an important development direction of machine learning.Deep learning aims at interpreting data rules by imitating the transmission mechanism of human brain transmitters,and establishes a deep neural network structure model.The research on feature extraction and fusion is developing rapidly in time and space.Deep learning has achieved amazing results in image classification and recognition.The previous meteorological data can not be effectively processed and preserved.Otherwise,the training of the network model requires a lot of historical data to achieve better results.So deep learning algorithm has not been widely used in the field of meteorology except some small-scale tests.This paper takes deep learning as the main algorithm and tries to make a breakthrough in weather element grid point prediction.Firstly,according to the prior knowledge of meteorology,some meteorological elements are selected and processed into data sets.Then the network structure suitable for meteorological field is designed,generates the forecast of a meteorological factor in the future.The main work is as follows:1.A new deep learning model framework,KM-ConvGRU,is proposed.This model can extract features from lattice data both in time and space.The KMConvGRU model consists of two parts: the initial fusion module PIM and the depth fusion module DIM.PIM module is to extract and fuse the spatial features of multi-layer and multi-element meteorological lattice data.Firstly,K-means clustering is used for multi-layer data.Then convolution operations are carried out within and between classes to achieve spatial feature extraction.DIM module extracts spatio-temporal comprehensive features of data blocks fused by preliminary feature extraction,mainly using multi-layer ConvGRU structure.2.Collect and organize the data set suitable for neural network training.Collect the EC-based zero-field data,forecast-field data and real-time data of meteorological elements based on ground observation from 2015 to 2017.Communicated with relevant professionals,six meteorological elements including height,specific humidity,relative humidity,temperature,U-wind component and V-wind component were selected.The target area ranged 5-60 degrees in latitude and 75-135 degrees in longitude.The collected data are cleaned and standardized.According to the network structure requirements,the data is designed as a five-dimensional space-time data block.3.9-KM-ConvGRU is the best model by internal optimization experiments.The experimental prove that 9-KM-ConvGRU model is better than PIM and DIM as independent models.Contrast experiment shows that 9-KM-ConvGRU is better than GRU network and Conv4 D network.The corrected results of 9-kmconvgru model were closer to the actual data than the corrected data,indicating that the model had an obvious correction effect.