Machine Learning Based Disastrous Meteorology Prediction

Disastrous meteorology generally includes meteorological disasters and their derivative disasters and its evolution process usually has certain laws.Therefore,it is important to grasp these laws to forecast the occurrence of disasters for disaster prevention and mitigation.Numerical weather prediction(NWP for short)is the mainstream method,which forecasts disastrous weather by simulating the atmospheric movement process.However,NWP needs large equipment to assist,and the economic cost is very high.There are errors between the model simulation and the actual situation,and its output needs to be further corrected to meet the prediction accuracy.The current popular machine learning model has great advantages in the above prediction cost and efficiency,and the improved machine learning prediction method can also achieve the disastrous weather prediction that meets the prediction requirements.Therefore,for the above two types of disasters,this paper has carried out research on machine learning prediction methods respectively.Machine learning algorithms for meteorological disasters are studied in this paper.Such disasters refer to those directly caused by typhoon,rainstorm,thunderstorm and other factors.We take the rainstorm disaster as an example to study.Affected by the east Asian monsoon,East China is prone to heavy rain disasters in summer.Therefore,it is of great significance to accurately forecast the rainy season precipitation in China.For sparse data of seasonal precipitation,this paper proposes ensemble empirical mode decomposition-Markov(D-Markov)model to achieve high-precision rainy season precipitation prediction.The D-Markov model introduces the wavelet analysis-based ensemble empirical mode decomposition(W-EEMD)into the Markov model to improve the ability of forecasting extreme precipitation in rainy season.The efficiency and effectiveness of our D-Markov model are verified by comparing with the BCC＿CSM1.1(m)prediction results provided by the China Meteorological Administration.Machine learning algorithms for meteorological derived disasters are studied in this paper.Such disasters refer to landslide,debris flow,air pollution and other disasters caused by meteorological factors.We take the atmospheric pollution in meteorological derivative disasters as an example to carry out research.The pollution of fine particles occurs frequently in winter in China,and pollution prevention and emission reduction are one of the important issues facing China.This paper designs a visual analysis system for air pollution prediction based on convolutional gate recurrent unit(Conv GRU),which provides a scientific basis for the formulation of pollution prevention and reduction schemes.The core algorithm of this system is to make the prediction model more suitable for fine particles by designing a comprehensive loss(C-Loss)function.C-Loss function comprehensively considering the absolute error and relative error between the prediction results and the ground truth.Then,by designing an interactive visual analysis system,the correlation between the formation process of air pollution and meteorological factors is explored,which further verifies the effectiveness of our prediction model proposed in this paper.For disastrous meteorology,this paper improves the machine learning algorithm and further analyzes the prediction results,which verifies the effectiveness of the algorithm improvement and provides strong support for disaster prevention and reduction.