Improving Parking Occupancy Prediction in Poor Data Conditions Through Customization and Learning to Learn

Haohao Qu, Sheng Liu, Zihan Guo, Linlin You, Jun Li

Abstract: Parking occupancy prediction (POP) can be used for many real-time parking-related services to significantly reduce the unnecessary cruising for parking and additional congestion. However, accurate and fast forecasting in data-poor car parks remains a challenge. To tackle the bottleneck, this paper proposes a knowledge transfer framework that can customize a lightweight but effective pre-trained network to those data-deficient parking lots for POP. The proposed approach integrates two novel ideas, namely Customization: select source domain utilizing reinforcement learning based on parking-related feature matching; and Learning to Learn: extract insightful prior knowledge from the selected sources using Federated Meta-learning. Results of a real-world case study with 34 parking lots in Guangzhou City, China, from June 1 to 30, 2018, show that compared to the baseline, the proposed approach can 1) bring approximately 21\(\%\) extra performance improvement; 2) improve the model adaptation and convergence speed dramatically; 3) stabilize predictions with error minor variance.