Application of Machine Learning in Methane Emissions Modelling

Methane emissions reduction activities are vital for reducing greenhouse gas emissions in the oil and gas industry. The Alberta Energy Regulator (AER) has been collecting air quality data throughout the province since 1986. Although the AER data is available to the public, the analysis of this data has not been thorough. Furthermore, there are many papers on reported emissions, and as yet, it remains unclear how to use and analyze this time series data. Machine learning is a state-of-the-art and effective method to forecast and understand methane emissions from the oil and gas sectors. The research documented in this thesis examined the methane emissions data from multiple monitoring stations in time by building machine learning models for prediction performance comparison. The first study compared Autoregressive Integrated Moving Average (ARIMA), Fully Connected Neural Network (FC-NN), and Long Short-Term Memory (LSTM) neural networks regarding total hydrocarbon non-methane hydrocarbon in general. The second study expanded the research by adding climate variables to build LSTM models to learn deep feature relationships between temperature, wind speed and wind directions regarding the methane concentration data. The third study examines the prediction performance of Gated Recurrent Units, Stacked LSTM, LSTM and Bidirectional LSTM neural networks with different scales of data for training to compare forecasting performance. The analysis of the experiments reveals 1. The LSTM neural network model provides better predictive performance than the other methods. With respect to the data itself, the average methane concentrations measured at the majority of Alberta airshed stations are higher than the global methane average. In addition, the methane concentration data itself exhibits both increasing and decreasing trends depending on the station. 2. Extra ambient climate variables can improve the predictive performance of the LSTM model: temperature improves the predictive performance of the methane concentration more than that of wind speed and direction. 3. GRU performs better when trained with shorter datasets, while the Stacked LSTM and the LSTM slightly outperform GRU and BiLSTM when training with more historical data. Also, more training data does not necessarily mean a significantly better prediction model but more training time. The results provide insights for the use of Predictive Emissions Monitoring System (PEMS) for estimating methane concentration emission data.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.