Efficient prediction of runway visual range by using a hybrid CNN-LSTM network architecture for aviation services

Visibility is the primary criterion for the landing and takeoff of an aircraft. At all major airports, a procedure called the low visibility procedure (LVP) is adopted in cases of marginal visibility, in which aircraft are permitted to land or take off based on the observed value of visibility, or specifically the runway visual range (RVR), and the forecast on the tendency of the RVR. Since the observed and forecasted value of the RVR is crucial for critical decisions on the landing or takeoff of an aircraft at airports, particularly at Patna airport, where landing is not permitted below the RVR of 1000 m, reliable prediction of the RVR is of great importance. As the predominant factors for the reduction of visibility at Patna Airport are fog, haze, and thunderstorms, this article proposes a novel, fully connected network architecture of a hybrid CNN-LSTM (convolutional neural network-long-short-term memory) framework to predict short-term RVR based on past instrumentally derived RVR values while considering exogenous meteorological factors that support the formation and intensification of fog, haze, etc. The proposed hybrid CNN-LSTM framework to function reliably and precisely, time series hourly observation data of meteorological elements, and the observed RVR values were used for the two forecasting horizons for a thorough examination. Extensive experimentation with the proposed hybrid CNN-LSTM framework architecture has been carried out along with the benchmark models of CNN, GRU (gated current unit), BiLSTM (bidirectional LSTM), and LSTM. Compared with CNN, GRU, BiLSTM, and LSTMs, the experimental findings reveal that the proposed hybrid CNN-LSTM framework achieves the best prediction performance in two random datasets with two different forecasting horizons, totaling four assessment criteria. Also, we look into how CNN, LSTM, and their hybrid network combinations might be used to make such predictions with reliable accuracy. We improve upon earlier models for short-term RVR prediction by optimising the loss function and network structure of the original CNN and LSTM models, making them more amenable to being used in actual operational environments.