Accurate traffic flow prediction is critical for intelligent transportation systems but remains a significant challenge due to complex and dynamic spatio-temporal dependencies. While Graph Convolutional Networks (GCNs) are powerful tools for this task, their performance is often limited by a reliance on predefined, static adjacency matrices that fail to capture the dynamic nature of road networks. To address this limitation, we propose LeaGCN, a Learnable Graph Convolutional Network with temporal attention. The cornerstone of our model is an adaptive GCN layer that utilizes a learnable adjacency matrix, allowing it to discover hidden spatial relationships directly from data without prior structural knowledge. This spatial module is integrated with an LSTM network to model temporal patterns and a soft attention mechanism to focus on the most relevant historical information, enhancing long-term prediction accuracy. Extensive experiments on the real-world PEMSD4 and PEMSD8 datasets demonstrate that LeaGCN significantly outperforms state-of-the-art baseline models. For instance, our model reduces the Mean Absolute Error (MAE) by up to 23% on the PEMSD4 dataset compared to the strongest baseline, validating its effectiveness and superiority.

Keywords