Urban Parking Demand Prediction With Hybrid Neural Network Architectures And Explainable AI

Abstract

Due to the increasing urbanisation and the associated growing traffic problems and parking shortages, a more accurate prediction of parking demand becomes increasingly important for intelligent traffic management and optimisation of parking space utilisation. The traditional methods of forecasting often fail to capture the complex dynamic relationships over time and the synergies of alternative model designs. The Smart Parking Management Dataset hourly temporal resampling is used in this work. It uses Entry features including temporal, historical, visibility and temperature for parking demand forecasts. Preprocessing includes feature engineering, Min–Max scaling and training-test split, chronologically. Various models are tried such as LSTM, xLSTM, Informer, Transformer, Autoformer, ARIMA, hybrid xLSTM-Informer model and multi-horizon BiLSTM models. The four metrics for evaluating the performance are the RMSE, MAE, MAPE, R2 and training duration. From the results, it can be concluded that the BiLSTM_1h model has the lowest RMSE (0.014), MAE (0.011), MAPE (5.536%) and the highest R2 (0.989), which demonstrates the effectiveness of the short-term parking demand projection. We use LIME and SHAP to analyse the feature contributions. We also offer an interface based on Flask to call the model for prediction. The results demonstrate that DL architectures provide valuable tools for accurate and interpretable urban parking demand forecast.