Hybrid LSTM-XGBoost framework with city embeddings for CO₂ emissions forecasting and scenario comparison

Developments of cities are always accompanied by environmental sacrifice, especially increasing CO₂ emission is one of the significant drivers of climate warming. Accurate city-level CO₂ emissions forecasting is critical for urban carbon management, where those cities are undergoing economic transitions and exhibiting divergent carbon emissions trends. Taking Hebei as an example, this study developed a hybrid deep learning framework combining LSTM networks and XGBoost models via adaptive power-law weighting, validated through a comprehensive ablation study, and further compared six scenario predictions. The model achieved an R² of 0.958 on the test set by balancing overall accuracy with city-level robustness, capturing the non-linear dynamics and spatiotemporal heterogeneity across cities. Monte Carlo uncertainty propagation analysis confirmed model robustness under both moderate (σ = 0.10) and conservative (σ = 0.20) emission data perturbation, with scenario rankings remaining stable across 500 iterations. Feature importance analysis revealed that the relative contributions from urbanization (36.1%), economic indicators (34.4%), and spatial characteristics (25.3%) were the primary predictive drivers, with lagged and differenced features outranking their contemporaneous counterparts. Post-hoc semantic analysis of the city embeddings confirmed that the learned representations encode an interpretable industrialization-to-ecological gradient consistent with independently identified city typologies. The clustering analysis identified three urban emission clusters related to different developing stages. Under the business-as-usual scenarios, provincial emissions are expected to reach 1.01 billion tons by 2028, while the comprehensive low-carbon interventions can achieve 13.2% reduction versus the baseline projection. The results show that, for the carbon emission control in historical heavy industrial provinces, an overall low-carbon strategy centered on industrial transformation and ecological improvement is needed. Collectively, this framework offers a transferable analytical approach for evidence-based carbon management and scenario-driven policy optimization in spatially heterogeneous urban systems.