Open AccessA Correction Method of Environmental Meteorological Model Based on Long‐Short‐Term Memory Neural Network
Author(s) -
Dai Yuliang,
Lu Zhenyu,
Zhang Hengde,
Zhan Tianming,
Lu Jia,
Wang Peng
Publication year - 2019
Publication title -
earth and space science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.843
H-Index - 23
ISSN - 2333-5084
DOI - 10.1029/2019ea000641
Subject(s) - artificial neural network , correlation coefficient , mean squared error , range (aeronautics) , term (time) , computer science , numerical weather prediction , standard deviation , meteorology , scale (ratio) , environmental science , algorithm , statistics , mathematics , artificial intelligence , machine learning , materials science , physics , quantum mechanics , composite material
Correction method can reduce the high deviation between the prediction results of numerical model and the observation results and improve the prediction accuracy. Based on the numerical models, including Rapid Refresh Multi‐scale Analysis and Prediction System‐CHEM and CMA Unified Atmospheric Chemistry Environment, and combined with European Centre for Medium‐Range Weather Forecasts meteorological field model data, a correction method of environmental meteorological model based on Long‐Short‐Term Memory (LSTM) neural network is proposed in this paper. The method mainly includes the following steps: First, the meteorological factors that have the main influence on the PM 2.5 concentration are selected by the correlation coefficient method; at the same time, the forecast results of numerical models are used as additional factors, and these factors are taken as the initial characteristics of the LSTM. Then, the network parameters of the LSTM are trained by initial characteristics and corresponding observation data, and the mapping relationship between the input factors and the output PM 2.5 concentration is established. Finally, European Centre for Medium‐Range Weather Forecasts data of March 2018 are selected to test the prediction performance of LSTM correction method. Results show that compared with single environment meteorological model, the correlation coefficient, the root mean square error, and the mean absolute error between forecasted and observed PM 2.5 concentration in 3–72 hr increased from 0.35–0.7 to 0.55–0.75, decreased from 45.3–67.46 to 37.74–53.7 μg/m 3 , and decreased by 7.86–16.52%, respectively. It indicates that the forecast performance of LSTM correction model is better than single environment meteorological model.