面對溫室氣體導致全球暖化等氣候變遷議題，各國政府及企業紛紛投入碳盤查以及長期減碳的行動，而有效預測溫室氣體排放也將有助於政府與企業即時檢視與調整減少溫室氣體的策略，達到預期的目標。
　　在預測溫室氣體排放的技術上，過去研究文獻較少是應用類神經網路模型於小樣本的時序列資料集。本研究旨在運用深度學習類神經網路模型如: 循環神經網路(Recurrent Neural Network：RNN)，長短期記憶網路(Long Short Term Memory Network：LSTM)， 門控循環單元(Gated Recurrent Unit：GRU) 和Transformer等模型結合基因演算法，預測美國工業溫室氣體二氧化碳排放量－以德克薩斯州作為示範。
美國德州是美國工業溫室氣體排放量居高的城市，本研究採用美國Environmental Protection Agency (EPA) 的公開資料集 ’Inventory of U.S. Greenhouse Gas Emissions and Sinks’ 在1990-2020年共31筆資料作為訓練預測模型的基礎。
　　綜合實驗結果，本研究結果顯示LSTM與Transformer模型對於小樣本時序列之工業溫室氣體排放資料集具有顯著的成效，並且是四個模型中預測成效最精準也是穩定性最高的模型。 另外，在本研究採用之相同設備與環境下的運算時間估算，依據不同模型特性，我們觀察到LSTM模型運算時間相較Transformer節省約6.97倍的時間。
　　本研究的成果提供了適用於小樣本時序列資料集的有效預測模型，可應用於未來的溫室氣體排放研究領域。並且可協助工業界和政策制定者密集監控二氧化碳排放量趨勢，制定降低碳排放的策略，進而實現有效減少溫室氣體排放的目標。

關鍵字: RNN, LSTM, GRU, Transformer, Greenhouse Gas Emission

In response to the challenges posed by climate change, particularly the issue of greenhouse gas emissions contributing to global warming, governments and companies worldwide have committed to carbon inventory and long-term carbon reduction actions. Accurately predicting greenhouse gas emissions can aid in the real-time review and adjustment of reduction strategies, helping these entities achieve their desired environmental goals.
Regarding technology for predicting greenhouse gas emissions, there was little research literature in the past that applied neural network models to small-sample time-series datasets. This study aims to use deep learning neural networks models such as Recurrent Neural Network (RNN), Long Short Term Memory Network (LSTM), Gated Recurrent Unit (GRU), and Transformer combined with Genetic Algorithms to predict the carbon dioxide emissions of industrial greenhouse gases in the United States - taking Texas as a demonstration.
Texas is a city with high industrial greenhouse gas emissions in the United States. This study uses the public data set 'Inventory of U.S. Greenhouse Gas Emissions and Sinks' of the Environmental Protection Agency (EPA) in the United States from 1990 to 2020. A total of 31 data were used as training predictions—the basis of the model.
Based on the experimental results, the results of this study show that the LSTM and Transformer models have significant effects on small-sample time-series industrial greenhouse gas emissions data sets. They are the most accurate and stable models among the four models. Furthermore, it has been observed that LSTM networks surpassed the Transformer networks by 6.97 times in terms of average computational time.
The results of this study provide an effective prediction model for small-sample time-series data sets, which can be applied to the future research field of greenhouse gas emissions. Moreover, it can assist industry and policymakers to intensively monitor the trend of carbon dioxide emissions and formulate strategies to reduce carbon emissions in a rolling manner to effectively reduce greenhouse gas emissions.

Keywords: RNN, LSTM, GRU, Transformer, Greenhouse Gas Emission

A. C. Riekstin, A. L. (2020, Jan.-March 1). Time Series-Based GHG Emissions Prediction for Smart Homes. IEEE Transactions on Sustainable Computing, 5, pp. 134-146.
Ashish Vaswani, N. S. (2017, December 6). Attention is All you Need. (I. G. Garnett, Ed.) Advances in Neural Information Processing Systems, 30.
David E. Rumelhart, G. E. (1986, Oct 9). Learning representations by back-propagating errors. Nature, 323, pp. 533–536.
Hussain Alibrahim; Simone A. Ludwig. (2021, August 09). Hyperparameter Optimization: Comparing Genetic Algorithm against Grid Search and Bayesian Optimization. IEEE.
Hyndman, R. J. (2018). Forecasting: principles and practice. OTexts.
Ludwig, H. A. (2021). Hyperparameter Optimization: Comparing Genetic Algorithm against Grid Search and Bayesian Optimization. IEEE Congress on Evolutionary Computation (CEC), (pp. 1551-1559). doi:10.1109/CEC45853.2021.9504761.
Mitchell, M. (1998). An Introduction to Genetic Algorithms. The MIT Press. doi:https://doi.org/10.7551/mitpress/3927.001.0001
Parag Sen, M. R. (2016, December 1). Application of ARIMA for forecasting energy consumption and GHG emission: A case study of an Indian pig iron manufacturing organization. Energy, 116, pp. 1031-1038.
Potvin, J.-Y. (1996, June). Genetic algorithms for the traveling salesman problem. Annals of Operations Research, 63, pp. 337–370.
Sepp Hochreiter, J. S. (1997, November 15). Long Short-Term Memory. Neural Computation, pp. 1735-1780.
Sharaf AlKheder, A. A. (2022, May). Forecasting of carbon dioxide emissions from power plants in Kuwait using United States Environmental Protection Agency, Intergovernmental panel on climate change, and machine learning methods. Renewable Energy, 191, pp. 819-827.
Texas State Energy Profile. (2022, May). Retrieved from U.S. Energy Information Administration (EIA): https://www.eia.gov/state/print.php?sid=TX
Utkucan Şahin. (2019, December 1). Forecasting of Turkey's greenhouse gas emissions using linear and nonlinear rolling metabolic grey model based on optimization. Journal of Cleaner Production, 239.
Wei Sun, M. L. (2016, May 20). Prediction and analysis of the three major industries and residential consumption CO2 emissions based on least squares support vector machine in China. Journal of Cleaner Production, 122, pp. 144-153.