Forecasting Carbon Emissions With External Drivers: Comparative Linear–Machine Learning Models With Global Change Assessment Model ( <scp>GCAM</scp> ) Mitigation Scenarios in West Africa

ABSTRACT Carbon dioxide emissions in West Africa are rising under population growth, urbanization, and persistent fossil‐fuel dependence, yet forecasting tools remain limited for heterogeneous, data‐constrained energy systems. This study develops a comparative framework that applies autoregressive modeling, machine learning, and attention‐based deep learning to forecast emissions across 16 West African countries from 2002 to 2023, with projections to 2050. Ridge Regression identifies GDP per capita, population, electricity access, fossil‐fuel use, and renewable energy shares as key predictors for SARIMAX, Random Forest, Transformer Encoder, and Attention‐Gated GRU models. Attention‐based architectures perform best, achieving MAPE values of 6.14%–6.59% and R 2 above 0.98. Business‐as‐usual projections show substantial mid‐century per capita emissions growth. Benchmarking against GCAM‐derived Net‐Zero 2050 and Net‐Zero 2070 pathways shows that the aggressive pathway limits cumulative emissions to 15.3 tCO 2 per capita by 2050, avoiding 2.4 tCO 2 per capita relative to the moderate case.