Hybrid Deep Learning Framework for Short-Term Urban Parking Occupancy Prediction Using Multi-Model Architectures

Abstract

Urbanization has resulted in heightened traffic congestion and parking deficiencies, rendering precise parking demand forecasts crucial for effective traffic management and best use of parking resources. Conventional prediction methods frequently encounter difficulties in capturing intricate temporal correlations and do not utilize the complimentary advantages of various model architectures. This research employs the Smart Parking Management Dataset, incorporating characteristics such as Temperature, Visibility, and Entry, resampled to hourly intervals. Preprocessing encompassed feature engineering, normalization via MinMax scaling, and the division of data into training and testing sets. Various deep learning models, such as LSTM, xLSTM, Informer, Transformer, Autoformer, and ARIMA, were employed, in addition to a hybrid xLSTM-Informer model and multi-horizon BiLSTM models. Performance was assessed by RMSE, MAE, MAPE, R²-score, and training duration. The BiLSTM_1h model exhibited the highest performance, attaining an RMSE of 0.014, MAE of 0.011, MAPE of 5.536, and a R²-score of 0.989, so illustrating its proficiency in accurately capturing short-term parking demand trends. Explainable AI methodologies, including LIME and SHAP, were utilized to elucidate feature contributions, while a Flask-based interface facilitated user-driven predictions via model inference. The incorporation of sophisticated hybrid architectures and interpretable modeling markedly improves the precision and dependability of urban parking demand predictions.