• Developed a deep learning-based framework to predict flight delays, with a focus on weather-related delays. • Introduced Liquid Neural Networks (LNN) to the flight delay prediction domain for the first time. • Compared LNN with LSTM and MLP models using real-world flight and weather datasets. • Applied SMOTE to handle class imbalance and used mutual information and Pearson correlation for feature selection. • Conducted systematic hyperparameter tuning via grid search and trial-and-error. • LNN achieved the best overall performance in predicting delays, particularly for the delay-related classes. • Demonstrated superior generalisation of LNN over traditional DL models, despite higher computational cost. • Results suggest LNN's potential in enhancing flight delay forecasting, especially in complex classification tasks. In this paper, we developed and analysed a model for predicting flight delays, focusing on those caused by weather conditions. Flight delays pose a significant problem for airlines, as the growth of air traffic often results in financial losses and passenger inconvenience. Unlike several previous research papers, this study provides a more detailed analysis of delays, distinguishing between weather-related and non-weather-related delays. Deep Learning (DL) models, including the proposed Liquid Neural Networks (LNN), Long Short-Term Memory (LSTM), and Multi-Layer Perceptron (MLP), were utilised to make predictions. We compare the LNN model, which has not been used previously for flight delay prediction, with the other two DL models. For this research, we employed two reliable datasets: flight data from the Bureau of Transportation Statistics and weather data from the Weather Underground site. Our methodological approach is comprehensive in addressing challenges encountered with the data. One major challenge was managing class imbalance, which can affect prediction accuracy, using the Synthetic Minority Over-sampling Technique algorithm. The selection of features for training and evaluation was based on mutual information scores and Pearson correlation coefficients. Hyperparameter optimisation was carried out systematically using grid search and trial-and-error. Models were evaluated with standard performance metrics, including precision, recall, F1-score, accuracy, and a confusion matrix. The results reveal that the LNN model outperformed the other DL models in predicting flight delays, particularly for the two classes related to delay occurrence. Overall, all three models produced excellent results in predicting on-time flights and acceptable outcomes when forecasting delays due to non-weather factors. However, the LNN model exhibited computational limitations when predicting weather-related delays, with notably lower scores for this class and significantly longer training durations. Despite these limitations, the LNN model demonstrated promising performance compared to prior models, even surpassing them on some metrics. These findings indicate that the LNN model has great potential for future flight delay prediction.
Bisandu et al. (Sun,) studied this question.