Short-term Prediction of the Near-space Wind Speed by Multi-layer Perceptron and LSTM Network

The deep learning method is suitable for dealing with complex characteristics of wind speed in the near space due to its higher globality, lower fitting error and higher operational efficiency, which can theoretically realize the nonlinear fitting of any function. It is conducive to solving the short-term and imminent prediction problems of the atmospheric wind field in the near space. Deep learning contains numerous algorithms and fully connected neural networks, which are also known as multilayer perceptrons (MLPs) and are commonly used at present. In this paper, we built a MLP network model and trained the wind speed at heights of 72-106 km and then realized the short-term prediction of wind speed for 48 hours based on the continuous observation of the meridional and zonal wind speed data in the near space by the Kunming Meteor Radar in 2012. The wind speed is divided into training dataset and validation dataset at a ratio of 9:1, and then we put them into MLP for several repeated iterative trainings and finally obtain the internal parameters when the network effect approaches to the best. According to the ultimate MLP network, the wind speed at each height is predicted to obtain the wind speed in the next 48 hours. Then, we analyse the error of the short-term prediction results and evaluate the accuracy by RMSE and MAPE. Overall, the accuracy and robustness of the 82-90 km height range are the best. The fitting effect basically conforms to the linear distribution. The predicted value is basically consistent with the observed value at low wind speeds, and there are slight differences at high wind speeds. The amplitude of the predicted value near the positive and negative peaks is slightly smaller than the observed value. The minimum RMSE of the meridional wind speed is 3.16 m/s at a height of 86 km, and the corresponding MAPE is equal to 13.59%. The minimum value of RMSE of zonal wind speed is 3.66 m/s at 84 km, at which time MAPE is equal to 18.40%.