Modeling Stopped Vehicle Dynamics on Urban Backstreets via GraphLSTM-Attn: A Context-Aware GeoAI Approach

Stopped vehicles on urban backstreets pose significant challenges to traffic flow, emergency response, and pedestrian safety. This study formulates the prediction of stopped vehicle counts as a spatiotemporal forecasting task, leveraging CCTV-derived vehicle trajectories to capture localized stopping behaviors. Each road segment equipped with a CCTV camera is modeled as a node within a graph, and hourly time series of stopped vehicles are constructed from trajectory data. To address the complex interplay of temporal patterns, spatial dependencies, and contextual features, we propose a novel deep learning architecture - GraphLSTM-Attn - that sequentially integrates a two-layer Long Short-Term Memory (LSTM) network and a Graph Attention Network (GAT). The model first encodes temporal dependencies in stopping behavior and subsequently incorporates spatial interactions among adjacent road segments using attention-based graph convolutions. Contextual features, such as road area and vehicle passing time, are selected via SHAP-based feature importance analysis and embedded into the prediction pipeline. The model was trained on CCTV data from 38 road segments in Anyang, South Korea, collected over an 8-day period, and benchmarked against LSTM, GRU, and various GNN-based baselines. Experimental results show that GraphLSTM-Attn consistently outperforms competing models across MAE, RMSE, and R2 metrics, demonstrating its ability to learn short-range temporal trends while adapting to the sparse and irregular connectivity of backstreet networks. The findings highlight the value of integrating attention-driven spatial learning and contextual embedding for fine-grained urban traffic forecasting.