Intelligent Short-Term Multiscale Prediction of Parking Space Availability Using an Attention-Enhanced Temporal Convolutional Network

Abstract

The accurate and rapid prediction of parking availability is helpful for improving parking efficiency and to optimize traffic systems. However, previous studies have suffered from limited training sample sizes and a lack of thorough investigation into the correlations among the factors affecting parking availability. The purpose of this study is to explore a prediction method that can account for multiple factors. Firstly, a dynamic prediction method based on a temporal convolutional network (TCN) model was confirmed to be efficient for ultra-short-term parking availability with an accuracy of 0.96 MSE. Then, an attention-enhanced TCN (A-TCN) model based on spatial attention modules was proposed. This model integrates multiple factors, including related dates, extreme weather, and human control, to predict the daily congestion index of parking lots in the short term, with a prediction period of up to one month. Experimental results on real data demonstrate that the MSE of A-TCN is 0.0061, exhibiting better training efficiency and prediction accuracy than a traditional TCN for the short-term prediction time scale.

Prediction of Parking Space Availability Using Improved MAT-LSTM Network

“…Recent studies have explored the application of CNNs to solving time series prediction problems. For instance, TCN achieved a mean squared error (MSE) of 0.96 in ultra-short-term single-input and single-output prediction tasks for NAP [19].…”

“…After integrating a spatial attention mechanism, the multi-input and single-output A-TCN network achieved an accuracy of 0.0061 MSE in short-term prediction tasks involving congestion indices. Moreover, for small-sample datasets (those with fewer than 400 sets of data), TCN networks can converge very rapidly [19]. Therefore, in tasks requiring network efficiency, CNN-based TCNs are preferred.…”

“…When selecting the network baseline, it is a key prerequisite to discuss the efficiency and accuracy of CNNs and RNNs in comparing these two network architectures on the same dataset, which is missing in previous studies. Moreover, different articles use different evaluation metrics, such as MAE in [16], RMSE in [17], and MSE in [19], which makes it difficult to compare the two frameworks in the same study or through the results of different studies.…”

“…As depicted in Figure 2, the TCN model [19] is built upon the CNN architecture and employs causal convolutions to suit sequential modeling. It employs dilated convolutions and residual blocks to capture historical memory.…”