Application of deep learning to multivariate aviation weather forecasting by long short-term memory

Weather forecasts are essential to aviation safety. Unreliable forecasts not only cause problems to pilots and air traffic controllers, but also lead to aviation accidents and incidents. This study develops a long short-term memory (LSTM) integrating both multiple linear regression and the Pearson's correlation coefficients to improve forecasting. A numerical dataset of 10 weather features (sea pressure, temperature, dew point temperature, relative humidity, wind speed, wind direction, sunshine rate, global solar radiation, visible mean, and cloud amount) is applied on every calendar day in a year to train and validate the LSTM for temperature forecasting. It is shown that data standardization is necessary to rescale the data to improve training convergence and reduce training time. In addition, feature selection by multiple linear regression and by Pearson's correlation coefficients are shown effective to the forecast accuracy of the LSTM. By selecting only the sensitive features (sea pressure, dew point temperature, relative humidity and relative humidity), the temperature forecasting errors can be reduced from RMSE 4.0274 to 2.2215 and MAPE 23.0538% to 5.0069%. LSTM deep learning with data standardization and feature selection is effective in forecasting for aviation safety.