Research on Visibility Forecast Based on LSTM Neural Network

For series problems in the meteorological field, the long-short-term memory neural network (LSTM) model is applied to the visibility forecast in the Beijing, Tianjin and Hebei region. First of all, the historical meteorological data during the months (Oct.-to-Dec. and Jan.-to-Feb.) of years 2015-2016 in the Beijing, Tianjin and Hebei region is selected as a dataset. Then, the Pearson Correlation Coefficient method is applied to select meteorological factors that have main influence on visibility to construct the training set, and adjust the network model parameters to train the neural network, and establish the input meteorological factors and the visibility of the output. Finally, European Centre for Medium-Range Weather Forecasts (ECMWF) data of the Beijing, Tianjin and Hebei region from October to December in 2017 is used to test the forecast effect of the LSTM model, and compared with the prediction results of the BP neural network. The result shows the visibility forecast based on the LSTM model is significantly better than BP neural network. The TS score in 0-1 km is 0.22, and its error is 0.34 km. The TS score in 1-10 km is 0.51, and its error is 2.18 km. The TS score above 10 km is 0.38, and its error is 6.07 km