Accurate visibility prediction is crucial for aviation safety, transportation operations, and environmental monitoring. This study proposes a data‐driven forecasting framework based on long short‐term memory (LSTM) neural networks and their hybrid variants. Using a dataset of 387,672 meteorological records collected from three observation sites in Taiwan, the models leverage key meteorological features—including wind speed, temperature, dew point, atmospheric pressure, and cloud‐cover—to capture temporal dependencies and nonlinear interactions that influence visibility. Four deep learning (DL) architectures were developed and evaluated: standard LSTM, bidirectional LSTM (BiLSTM), stacked LSTM, and convolutional neural network LSTM (CNN‐LSTM). Among them, the BiLSTM achieved the best performance, with the lowest validation mean squared error (MSE) (0.002), mean absolute error (MAE) (0.026), and root MSE (RMSE) (0.050). The results indicate that BiLSTM offers improved accuracy and stability over other architectures, demonstrating its potential as a practical and scalable tool for operational visibility prediction in meteorological applications.
Chuang et al. (Thu,) studied this question.