Research On Meteorological Intelligent Algorithm Based On Machine Learning

More accurate observation and forecasting of current weather conditions in the field of meteorology has been a key concern.With the accumulation of massive historical base data and the advancement of supercomputer computing power,weather forecasting is now becoming more and more accurate.However,as the demand for finer grid point forecasting gradually increases,the forecasting area continues to be refined,gradually transitioning from the original municipal forecasts to county-level forecasts and neighborhood forecasts.The pressure of refinement modeling and solving of numerical models is also increasing,and gradually encountering bottlenecks.In recent years,the rapid development of big data technology and artificial intelligence technology has provided opportunities for the meteorological field to fully exploit the value of massive historical data,and gradually realize the data-driven ideas to solve the problems in the meteorological field.The application topics of artificial intelligence in the field of meteorology have become a research hotspot,combining the latest machine learning related algorithms in climate research,marine typhoon forecasting and warning,sea fog forecasting,short time proximity weather forecasting and other operations.Specific research topics include quality control of weather radar,signal extrapolation,quantitative precipitation estimation,strong convective weather identification and warning,environmental forecasting,storm environment identification,and weather system identification.In this paper,we address three important and challenging problems,and the main research contents are as follows.(1)Study of refined grid point forecast revisions.Firstly,in cooperation with the meteorological operation department of Hunan Province,we developed a standard data set format for weather forecast grid point data in the meteorological field,which provides a standard processing method for similar research in the future.A hybrid attention mechanism based on the time-space point forecast revision algorithm is proposed.The 12-hourly forecast data from the European Centre for Medium-Range Weather Forecasts are revised.The practical effectiveness of the algorithm is also verified by using traditional machine learning validation metrics and various methods such as statistical tests and weather tests commonly used in the meteorological field.The experiments show that the method in this paper is an effective objective revision method in practical operations and can provide an important reference for forecasters.(2)Research on the problem of full classification recognition of complex groundbased clouds.This problem has been a difficult project in ground observation,because of the large number of cloud categories,the small difference between some classes,and the influence of various weather conditions,which is subject to subjective and objective factors in the manual observation project.Most of the previous studies on objective observation of cloud shapes can only identify the cloud family,or cloud genus,or some of the cloud shapes,and basically can only identify about 10 categories,and the accuracy is not high,which cannot be compared with the manual observation.In this paper,we propose a hybrid complex cloud classification algorithm based on the combination of a priori visibility features and global features to identify all 29 cloud types,and optimize it for different regions and different shots.The algorithm won the first place in the weather phenomenon video smart observer test organized by China Meteorological Administration in 2019.(3)Research on radar quantitative precipitation estimation.Aiming at the various types of error situations in the quantitative precipitation estimation of radar,this paper proposes a quantitative precipitation estimation algorithm based on spatio-temporal feature learning model,using a convolutional neural network with encoder-decoderlike structure to extract the spatial features of radar reflectivity factor to improve the input accuracy,and combining with the spatio-temporal features,a variety of spatiotemporal network structures are designed for testing experiments,and statistical tests and weatherological methods are used to analyze The inversion effect of the quantitative precipitation estimation algorithm based on spatio-temporal networks is analyzed by statistical tests and weather science methods.The experimental results show that the algorithm in this paper can invert the corresponding rainfall information better and improve the accuracy of quantitative rainfall estimation.