Short-term forecasting of spread F using LSTM network

Considering the main mechanism of spread F and the available data related to the occurrence of spread F in our country, a model based on long short-term memory (LSTM) network is built to forecast spread F in 3 hours. The models for Manzhouli, Beijing, and Haikou were built and tested using data in 2015 and 2016. The mean accuracy rates in Manzhouli, Beijing and Haiou are 92.4%, 95.3%, and 96.0%. The mean precision rates are 75.0%, 61.2%, and 62.6%, and the mean recall rates are 73.0%, 50.6%, and 31.5%, respectively. It is shown that the model has a high prediction ability in most cases, but it still needs to improve the ability of the model in some cases, especially in low latitude like Haikou station with a lower recall rate.