A Platform-centric Framework for Intelligent Parking Traffic Prediction and Resource Optimization in Shared AVPC Systems

To address the challenges posed by uncertainties in the parking behaviors of private owners and temporary users, as well as the complexities involved in integrating shared parking with Electric Vehicle (EV)-charging, this paper proposes a novel platform-centric intelligent framework for shared Automated Valet Parking and Charging (AVPC) systems. The framework leverages Long Short-Term Memory (LSTM) and Deep Reinforcement Learning (DRL) to optimize both parking traffic prediction and resource allocation, respectively. Specifically, to mitigate uncertainties in vehicle parking and EV-charging demand and supply, we utilize an LSTM prediction model to forecast the average day-ahead arrival times, departure times, and service pricing for parking space owners (O-users) and temporary users (R-users). Then, we design an improved Proximal Policy Optimization (PPO)-based algorithm with large warm-up training steps that integrates the LSTM prediction results with real-time supply and demand information from O-users and R-users to determine optimal shared AVPC resource allocation. Extensive simulations using real-world parking datasets demonstrate that the LSTM model achieves an average Mean Absolute Percentage Error (MAPE) of 1.71% and 0.08% for O-users and R-users' parking traffic predictions, respectively. Additionally, the proposed LSTM-PPO-based approach improves platform profit and parking resource utilization by at least 9% and 15%, respectively, compared with state-of-the-art.