Spatiotemporal Deep-Learning Networks for Shared-Parking Demand Prediction
Liu Y, Chunyu Liu, Xia Luo
Abstract
One fundamental issue in managing a shared-parking system is predicting shared-parking demand. Such predictions are very challenging because predicting shared-parking demand usually involves nonlinearities and complex spatiotemporal dependencies. In this paper, we propose a deep learning–based network comprising of three modeling components—CNN-Module, Conv-LSTM-Module, and LSTM-Module—to predict the shared-parking inflow and outflow in each region of a shared-parking system. First, the CNN-Module utilized convolution neural networks to determine local spatial dependencies. Second, the Conv-LSTM-Module leveraged the Conv-LSTM neural network to capture similarities of shared-parking demand across different regions. Finally, the LSTM-Module was applied to model temporal features by using the Long Short-Term Memory (LSTM) network. Moreover, we also divided the input into three components (recent, daily, and weekly) to extract the periodically shifted relations. The model was evaluated using a real-world shared-parking data set in Chengdu, China. Experiments showed that our model outperforms six other well-known baseline methods within an acceptable time frame. Extensive additional experiments and evaluations were conducted to investigate the sensitivity of our model.