Multi-Scale Causal Inference and Scenario Prediction for Carbon Emissions from Energy Consumption in the Pearl River...

Accurately identifying the driving mechanisms of carbon emissions from energy consumption is crucial for regional low-carbon transition. However, existing studies predominantly focus on statistical regression or trend extrapolation, lacking systematic characterization of multi-scale causal mechanisms and comprehensive evaluation of multi-scenario pathways. Therefore, this study integrates multi-source data on socioeconomic conditions, land use, and climate in the Pearl River Delta urban agglomeration (PRD) from 2000 to 2022. It constructs an integrated multi-source data-deep learning model evaluation-interpretability analysis -causal identification-scenario prediction. Key variables were identified using Sharpley Additivity Explanation (SHAP), followed by quantification of net effects and transmission pathways via causal graphs. Scenario prediction for carbon emissions from energy consumption (CE) were conducted using sub-scale optimal models. The results indicate (1) Scale-specific modeling indicates that the Sparrow Search Algorithm-optimized Convolutional Long Short-Term Memory Network (SSA-ConvLSTM) performs best at the pixel scale, while the Ister-Transformer deep learning model-which integrates temporal and spatial features-is optimal at the county and city scales. These models can be utilized for rolling evaluation and calibration. (2) SHAP analysis reveals a "scale synergy-morphological threshold/saturation-efficiency offsetting" relationship. Causal inference further demonstrates that pixel-scale PGDP exerts a negative effect on carbon emissions. At the county scale, GDP and JCQ primarily influence emissions through population, while at the city scale, population's direct impact on emissions is most significant. (3) Scenario simulations indicate that policy implementation should be stratified: municipal-scale "total volume constraints," county-scale "compact development and stock efficiency enhancement," and pixel-scale "energy efficiency and technological offsetting." The integrated framework of causal identification-interpretative analysis-scenario forecasting proposed in this study provides robust support for scientific carbon emission prediction and regional low-carbon policy formulation.