Forecasting Atmospheric Visibility Using Auto Regressive Recurrent Neural Network

Abstract

Atmospheric visibility conditions not only affect traffic on roads, but also aviation operations. Poor visibility at the destination site can reduce airport capacity leading to ground delays, flight cancellations, flight diversions, and extra operating costs. Hence, timely forecast of visibility is important for safe operation in both airports and highways. Visibility is affected by meteorological weather variables such as precipitation, temperature, wind speed, humidity, smoke, fog, mist, and Particulate Matter (PM) concentrations in the atmosphere. This paper is an effort to forecast univariate weather variable visibility and explore the effect of highly correlated meteorological weather variables on visibility, using an Auto Regressive Recurrent Neural Network (ARRNN). By adjusting the number of epochs and the regression horizon, i.e. past time steps used in visibility prediction, we showed that ARRNN outperforms long-short term memory (LSTM) networks and vanilla recurrent neural network (Vanilla RNN) in terms of coefficient of determination (R2).