融合Transformer的图卷积LSTM道路网络分形维数动态预测模型

Fractal geometry offers a robust framework for characterizing the irregular and self-similar structures inherent in road networks. While fractal dimensions effectively quantify geometric complexity, existing models inadequately capture their dynamic evolution over time, limiting proactive urban planning capabilities. This study proposes the Graph Convolutional LSTM with Transformer (GCLSTM-Transformer), a novel spatio-temporal model that integrates attention-based graph convolutions to capture spatial hierarchies and Transformer modules to model long-term dependencies. Using road network data from 200 Indian counties (2020–2024), we construct dual graphs with centrality-based node features and employ Fuzzy C-means clustering to refine fractal dimension calculations. Experiments demonstrate that GCLSTM-Transformer achieves state-of-the-art performance, with R²=0.9249 (geometric) and R²=0.9421 (structural), outperforming baselines. The interpretability of the model, enabled by attention mechanisms, identifies critical network nodes, aiding traffic management and infrastructure design. This work advances spatiotemporal graph modelling and provides a scalable framework adaptable to other network-based fractal analyses.