Predicting Carbon Dioxide Emissions in the United States of America Using Machine Learning Algorithms

Abstract

In the past few decades, several nations have experienced air pollution and environmental difficulties because of carbon dioxide (CO₂) emissions. One of the most crucial methods for regulating and maximizing CO₂ emission reductions is precise forecasting. Four machine learning algorithms with high forecasting precision and low data requirements were developed in this study to estimate CO₂ emissions in the United States (US). Data from a dataset covering the years 1973/01 to 2022/07 that included information on different energy sources that had an impact on CO₂ emissions were examined. Then, four algorithms performed the CO₂ emissions forecast from the layer recurrent neural network with 10 nodes (L-RNN), a feed-forward neural network with 10 nodes (FFNN), a convolutional neural network with two layers with 10 and 5 filters (CNN1), and convolutional neural network with two layers and with 50 and 25 filters (CNN2) models. Each algorithm's forecast accuracy was assessed using eight indicators. The three preprocessing techniques used are (1) Without any processing techniques, (2) Processed using max-min normalization technique, and (3) Processed using max-min normalization technique and decomposed by variation mode decomposition (VMD) technique with 7 intrinsic mode functions and 1000 iterations. The latter with L-RNN algorithm gave a high accuracy between the forecasting and actual values. The results of CO₂ emissions from 2011/05 to 2022/07 have been forecasted, and the L-RNN algorithm had the highest forecast accuracy. The research makes significant literary contributions. One of the first studies to focus on predicting CO₂ emissions in the USA using a combination of three preprocessing approaches and four machine-learning algorithms, predicting the number of overall CO₂ emissions while also accounting for a broader range of inputs.