Neural Network Prediction Model – Applied to U.S. Industrial Greenhouse Gas Emissions

Abstract

This study explores the use of deep learning neural network models for predicting greenhouse gas emissions, focusing on small-sample time-series data sets, an area with limited prior research. It utilizes Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), and Transformers combined with Genetic Algorithms to forecast CO2 emissions from industrial sources in Texas, a major contributor to U.S. greenhouse gas emissions. The analysis is based on the Environmental Protection Agency's (EPA) "Inventory of U.S. Greenhouse Gas Emissions and Sinks" dataset, spanning 1990 to 2020. The results indicate that LSTM and Transformer models are particularly effective, with LSTM outperforming Transformers in computational efficiency by 6.97 times. These findings highlight the potential of LSTM and Transformer models as accurate and stable tools for predicting CO2 emissions in small-sample time-series data, offering valuable insights for future research and policy development in environmental management.