Abstract
Recent advances in artificial intelligence have enabled effective modeling of complex industrial spatial systems. Predicting dynamic changes in manufacturing industrial land remains challenging due to spatial heterogeneity and temporal evolution across multiple influencing factors. This study develops a computational framework that integrates graph attention networks and Transformer-based temporal modeling with a spatial decoding module to forecast manufacturing industrial land change. Historical multi-source data spanning 2014 to 2023 are structured as a spatiotemporal graph, with nodes representing spatial units and edges encoding spatial proximity, transportation connections, and functional similarity. Graph attention networks learn the influence of neighboring units, while the Transformer captures temporal dependencies across consecutive periods. The spatial decoding module converts node-level predictions into geographic maps, generating interpretable probability and classification outputs. Experiments demonstrate that the proposed method outperforms baseline models in node-level classification with an accuracy of 0.768 and a Macro-F1 score of 0.712, achieves high spatial consistency with a Kappa of 0.668 and a mean intersection over union of 0.498, and accurately predicts manufacturing industrial land change intensity with a mean absolute error of 0.077 and an R² of 0.765. A component-wise ablation experiment further evaluates the contributions of the main model components. The results show that combining spatial graph modeling with temporal deep learning provides a robust computational approach for manufacturing industrial land change prediction and supports data-driven industrial land planning decisions.