Multiprior LSTM (mpLSTM): Predicting Visibility with Uncertainties from Complex Background States

Visibility condition is a crucial factor to airport flight arrangement. Low visibility usually leads to capacity reduction, which incurs delay or cancellation of arriving and departing flights. To keep the airport capacity at a high rate, accurate visibility forecast is urgently needed.

To predict low-visibility states, we propose a multi-prior LSTM (mpLSTM) model to predict future visibility probability distribution based on a variety of prior knowledge.The prior knowledge comes from previous background states, since low-visibility occurrence rate is highly related to previous background states, such as season or special weather conditions. The model will learn from the information of prior knowledge and performs visibility prediction with uncertainties with an end-to-end manner. Experiments show that our forecasting system outperforms human-prediction for history data. With longer time period of input data, prediction accuracy also increases relevantly.