The Typical Applications' Study In Weather Forecasting Based On Machine Learning

Intelligent weather forecasting,armed with machine learning,is helpful for describing the global weather accurately in each aspect.As a synthetic of numerical weather forecasting,statistical inference,machine learning and interpreting numerical weather forecasting's products,intelligent weather forecasting will provide seamless and delicacy forecasting.However,when it comes to industrial forecasting,three problems remains to be tackled considering the challenges caused by systematical preprocessing,efficient data transportation and delivery,and the promise of forecasting seamlessly and delicately.(1)To forecast in real time,how to denoise data efficiently.(2)To deliver and process data in efficient,how to compress and reconstruct them.(3)To analyze the weather processes in small scale based on forecasting's products,how to interpolate the reanalysis data.Based on these issues,this paper carried out the following works.(1)Denosing infrared hyper-spectral data contaminated by clouds.A feature construction methods based on featured channels is proposed,and cloudy data is distinguished by selected featured channels based on logistic regression.The experiment demonstrates that the features construction method is reasonable;in terms of predicting,the accuracy after predicting the samples in sea areas are more than 95%,while it in land areas with a rise of 12% in recall after utilizing surface emissivity data;the logistic regression based on featured channels is one the most practical models for real-time cloud detection.(2)The compression and reconstruction of infrared hyper-spectral data.To compress the spectrum dimension and spatial dimension simultaneously,a new compression model named HCR is designed based on convolution.According to the experiment,the mean squared error of HCR is decreased 13.82% compared with that of principle component analysis.(3)Designing an anisotropy multi-scale kernel,and two multivariate interpolation models for weather processes based on the new kernel.The effects of models are verified.For the weather processes without cyclones,the mean root mean squared error of the new model declines in 20% at least;for the weather processes with cyclones,the mean root mean squared error of the new model for weather processes with cyclones decreases about 55% compared with that of spline,and around 95% compared with back propagation neural networks.