Dispersion based Recurrent Neural Network Model for Methane Monitoring in Albertan Tailings Ponds

TL;DR

This work tackles methane emissions from Alberta oil sands tailings ponds by marrying mechanistic degradation models with atmospheric dispersion through a physics-constrained recurrent neural network (DIRNN). It introduces an inverse-dispersion variant (iDIRNN) to estimate emissions from measured concentrations, enabling source attribution around weather stations, including inactive ponds. The approach leverages WBEA atmospheric data and MM-derived diluent degradation to predict both $CH_4$ concentrations and emissions, achieving forecasts that capture seasonal patterns and sectoral emission contributions; it also reveals substantial underestimation in official CH$_4$ tallies. The results demonstrate the method's potential for scalable, physics-informed environmental monitoring that supports regulatory and industry decision-making, with future work extending coverage to all wind directions and integrating remote sensing.

Abstract

Bitumen extraction for the production of synthetic crude oil in Canada's Athabasca Oil Sands industry has recently come under spotlight for being a significant source of greenhouse gas emission. A major cause of concern is methane, a greenhouse gas produced by the anaerobic biodegradation of hydrocarbons in oil sands residues, or tailings, stored in settle basins commonly known as oil sands tailing ponds. In order to determine the methane emitting potential of these tailing ponds and have future methane projections, we use real-time weather data, mechanistic models developed from laboratory controlled experiments, and industrial reports to train a physics constrained machine learning model. Our trained model can successfully identify the directions of active ponds and estimate their emission levels, which are generally hard to obtain due to data sampling restrictions. We found that each active oil sands tailing pond could emit between 950 to 1500 tonnes of methane per year, whose environmental impact is equivalent to carbon dioxide emissions from at least 6000 gasoline powered vehicles. Although abandoned ponds are often presumed to have insignificant emissions, our findings indicate that these ponds could become active over time and potentially emit up to 1000 tonnes of methane each year. Taking an average over all datasets that was used in model training, we estimate that emissions around major oil sands regions would need to be reduced by approximately 12% over a year, to reduce the average methane concentrations to 2005 levels.

Figures (10)

• Figure 1: Region of Wood Buffalo with all the weather monitoring stations of interest and main OSTPs and/or EPLs. Selected stations: Mannix ('Pond 2/3' approximately 1.4 km northwest of the station); Lower Camp ('Pond 2/3' approximately 3.5 kms southwest and 'Pond 5' about 1.4 kms west of the station, respectively); Mildred Lake ('Mildred Lake Settling Basin'/MLSB approximately 1.7 kms northwest and 'Pond 5' approximately 2.1 kms southeast of the station, respectively); Buffalo ('West-In-Pit'/WIP at 0.8 kms northwest of the station).
• Figure 2: Graphical representation of the proposed modeling framework. Left: Representation of how the different input variables interact. Solid lines depict a direct connection and the dashed lines represent an indirect affect. Right: A flowchart of DIRNN.
• Figure 3: Results for predicting CH$_4$ concentrations and emissions using a neural network based representation of the dispersion/advection terms in ADM. The green dashed line indicates the data split between training and validation set. For concentrations, the model accurately captures the seasonal nature of concentrations as well as the emissions based on given input data. The model predictions always fall within the 95% confidence of interval.
• Figure 4: Scaled values (between 0 and 1) of true versus predicted CH$_4$ concentrations and emissions in 2023 for stations Mannix, Lower Camp and Mildred Lake using a neural network representation of the dispersion/advection terms in the ADM. The trained model accurately forecasts concentrations and emissions for one year when trained on historical data, suggesting future trends based on different levels of input data.
• Figure 5: Estimation of yearly CH$_4$ emissions (in t) from all wind directions around each station for 2020 (left) and 2023 (right). Estimated emissions are highest from the direction of OSTPs, with MLSB emitting 4500t per year and Pond 2/3 emitting more than 850 tonnes (t) per year every 20 degree interval. Abandoned lakes such as WIP and Pond 5 also emits more than 2000t of CH$_4$ per year. Please note that the round plots are for demonstration purpose and are not to scale. Latitude and longitude coordinate ranges given in $X$ and $Y$ axis.