STDSE-Net: A GIS-Driven AI Framework for Dynamic Ensemble-Based Energy Emission Prediction

The integration of Geographic Information Systems (GIS) with advanced artificial intelligence (AI) techniques presents transformative opportunities for accurate energy forecasting and developing decarbonization strategies. However, existing forecasting models often fall short in capturing the complex spatial-temporal interdependencies that drive CO₂ emissions such as geographic autocorrelation, delayed policy impacts, and regional socio-economic asymmetries. These limitations hinder predictive performance, particularly in emission-volatile and underrepresented regions. To address these challenges, this study proposes a GIS-Based Intelligent Energy Forecasting System built upon a novel Spatio-Temporal Dynamic Stacking Ensemble (STDSE-Net) framework. The system fuses heterogeneous base models—Random Forest, SVR-RBF, LightGBM and a custom GIS-augmented XGBoost-Geo—through Particle Swarm Optimization (PSO) that dynamically adjusts model weights based on spatial error gradients. A meta-learner (XGBoost) integrates spatially weighted predictions using a hybrid mean-median voting strategy, improving robustness against outliers and data sparsity. Additionally, quantile regression is employed to estimate uncertainty bounds, and a temporal embedding module captures lagged emission trends via rolling policy-aware averages. Trained on the Global Data on Sustainable Energy (2000–2020), STDSE-Net achieves a predictive accuracy of R² = 0.98, RMSE = 3.20 Mt, and MAE = 2.10 Mt, outperforming existing benchmarks by 22–48% in accuracy. Its GIS-informed dynamic weighting framework prioritises regionally optimal models, thereby reducing residual bias in heterogeneous contexts, such as Sub-Saharan Africa and South Asia. Spatial-temporal cross-validation confirms strong generalizability, and residual analysis demonstrates uniform error distribution across diverse geographic clusters. By bridging spatial heterogeneity with AI-driven adaptability, STDSE-Net offers a scalable and interpretable blueprint for next-generation emission forecasting. These advancements have direct implications for renewable energy planning, regional grid optimization, and geographically targeted climate policies. Future research will extend this framework by incorporating real-time satellite imagery, transformer-based temporal encoders, and policy document mining to support more granular, long-horizon predictions in dynamic policy environments.