Work

A hybrid artificial neural network approach to forecasting CO₂ emissions in the United States under policy uncertainty and energy consumption

Year: 2026

Type: article

Abstract: This study applies an advanced hybrid artificial neural network framework to predict total CO₂ emissions in the United States using monthly data from 1986M1 to 2025M12. Natural gas consumption, economic policy uncertainty, monetary policy uncertainty, oil prices, and renewable energy consumption are employed as predictors of emissions dynamics. The study compares the conventional multilayer perceptron with hybrid MLP models optimized using Tunicate Swarm Optimization, Moth-Flame Optimization, the Arithmetic Optimization Algorithm, the Sine Cosine Algorithm, and the Gaussian Mutation Specular Reflection Learning with Local Escaping Operator–based Artificial Electric Field Algorithm. It further extends the existing MLP-GRLEO-AEFA framework by integrating kernel regularization, thereby developing the MLP-KR-GRLEO-AEFA model. Model performance is evaluated using separate training and testing samples and assessed through NNSE, R², MSE, RMSE, MSLE, and MAE. Robustness is further examined using benchmark functions, convergence analysis, Friedman ranking, Wilcoxon tests, residual diagnostics, observed-versus-predicted plots, and permutation feature importance. The results show that the conventional MLP achieves the strongest training performance, MLP-SCA records the best out-of-sample testing performance, and MLP-KR-GRLEO-AEFA provides the strongest aggregate full-sample performance. Permutation feature importance also identifies the predictors that contribute most strongly to U.S. CO₂ emissions prediction. Based on these findings the study provides a reliable framework for forecasting U.S. CO2 emissions and supporting climate and energy-policy decisions.

Source: Scientific Reports

DOI: 10.1038/s41598-026-77197-y

Authors: Seyed Alireza Athari, Serdar Neslihanoglu, Berna Uzun, Adeola Praise Adepoju, Mohamed Djafar Henni