Multi-step LSTM Prediction Model for Visibility Prediction

In this paper, we present a deep learning framework with attention mechanism for visibility prediction. We firstly formulate visibility prediction as a temporal prediction problem. An encoder-decoder architecture based network is proposed to generate a multi-step prediction. To adaptively focus on different parts of the input and output sequence, we incorporate input attention and temporal attention into the network. Experiments verify the feasibility of the proposed model. We produce state-of-the-art prediction accuracy (68.9%) on the runway visual range prediction in our customized data set collected at observation stations of the airport.