Predicting Flight Delays Using LSTM and BILSTM Models with Shap Interpretation

Flight delays represent a critical challenge in air transportation, affecting passenger satisfaction, operational efficiency, and financial outcomes. This study develops predictive models for flight delay duration at Juanda International Airport, Surabaya, using Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) networks integrated with a Shapley Additive Explanations (SHAP) interpretability method. The research utilized 242,638 flight observations spanning January 2023 to October 2025, incorporating flight operational and meteorological variables. The dataset was partitioned into training (62.35%), validation (9.01%), and testing (28.64%) subsets. After Min-Max normalization and preprocessing, models were designed with varying hyperparameters through grid search optimization. Performance evaluation employed Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Results demonstrated that LSTM with three hidden layers and eight neurons achieved superior performance, with an MAE of 16.888 minutes and an RMSE of 22.401 minutes on test data. BiLSTM yielded an MAE of 24.399 minutes and an RMSE of 34.596 minutes, establishing LSTM as the optimal model. SHAP interpretation revealed that operational factors—including flight type, airline, and airport routes—represent the dominant predictors of delays, followed by meteorological factors such as wind speed and temporal factors including scheduling time. Although R² values remained relatively low, this research provides valuable interpretability insights into delay determinants, enabling data-driven decision-making for airport management and airlines to enhance punctuality and operational efficiency.