Parking occupancy prediction under COVID-19 anti-pandemic policies: A model based on a policy-aware temporal convolutional network

Real-time and reliable parking occupancy prediction is critical for managing transportation infrastructure and services. However, the COVID-19 pandemic and its corresponding response have led to changes in daily routines and behaviors, including travel time distribution, frequency, and parking time. Changes in parking behavior pose significant challenges for accurately predicting parking occupancy. To address this issue, we proposed a novel policy-aware temporal convolutional network (P-TCN) model that considers the policy dependence of time-series data irregularities. The model employed 3D vector data as input to predict future parking occupancy while ensuring reliability through 10-fold cross-validation. The SHapley Additive exPlanations (SHAP) method was also utilized to analyze the impact of the feature variables on the model. Experimental results show that the P-TCN model accurately identifies the effects of anti-pandemic policies and outperforms existing models such as multiple linear regression (MLR), autoregressive integrated moving average (ARIMA), fully convolutional network (FCN), long short-term memory (LSTM), and temporal convolutional network (TCN). Moreover, the feature variables “policy”, “weather”, and “time information” contribute significantly to the model prediction results, with “policy” having the most significant effect. This model has wide applicability for predicting and interpreting parking occupancy in different countries and regions. As the COVID-19 pandemic gradually normalizes, the P-TCN model's parking occupancy information will help managers better understand residents' behavior and respond flexibly to emergencies.