An Interpretable AI-Based Smart Engineering Framework for Carbon Emission Prediction in Coastal Port-Industrial Zones

Abstract

Carbon emission prediction of coastal port-industrial zones integrates industrial production, maritime logistics and spatial environmental governance. Existing models fail to simultaneously capture multi-scale temporal fluctuations and spatial correlations of industrial units under meteorological and policy interference. This paper constructs an interpretable dual-branch framework combining Ensemble Empirical Mode Decomposition–XGBoost (EEMD-XGBoost) and a Temporal Graph Neural Network (T-GNN). The EEMD-XGBoost branch extracts multi-scale temporal features from non-stationary emission sequences, while the T-GNN models spatiotemporal dependencies of industrial sub-units; weighted fusion integrates the two branches, and the framework outputs interpretable indicators including feature importance, spatial contribution and temporal attention weights to analyze emission driving factors. Two 2023 datasets are adopted: a coastal-port prefecture-level subset derived from the national regional carbon dataset (28 coastal port-related prefectures), split chronologically into 70% training and 30% test sets; the Yangtze River Delta ship dataset integrating AIS, ship properties and fuel data is partitioned via tonnage-based stratified sampling into 7:2:1 training–validation–test subsets, with an independent test subset for short-term ship-type forecasting. On the independent ship test set for maritime greenhouse gas prediction, the proposed model achieves an R2 of 0.96, 0.95 and 0.93 for container, bulk and oil ships respectively, which only applies to ship-scale forecasting rather than regional carbon prediction. Special ablation experiments show that single T-GNN converges within 72 epochs, single EEMD-XGBoost has a high-frequency fitting error of 6.82%, and the complete dual-branch model reaches a spatial feature capture rate of 94.65% with only a 3.47% fitting error, despite a 27.52 ms single-sample inference time, proving the synergy of the two branches. The model outperforms baselines under abnormal and sparse data conditions. This interpretable framework supports traceable refined carbon management and provides quantitative engineering implications for port zoning control, differentiated ship emission reduction and regional low-carbon policy implementation.