Meng Jia (贾萌)

Meng Jia (贾萌)

PhD student in Statistics at Colorado School of Mines

Talks and presentations

Methane emission localization and quantification on oil and gas sites using physics-informed neural networks

December 12, 2024

Poster, AGU Fall Meeting 2024, Washington, D.C.

Abstract: Mitigating methane emissions from the oil and gas (O&G) sector is pivotal for addressing global climate change quickly, as this sector contributes 22% of global anthropogenic methane emissions and 32% of such emissions in the United States. Continuous monitoring systems (CMS) are becoming increasingly vital for methane emission monitoring on O&G since they provide measurement of methane concentrations, and corresponding wind data, in near real-time, capable of capturing both highly fluctuating emissions and transient super emitters on O&G. However, existing localization and quantification approaches may not be the most effective for several reasons. First, most widely used forward models, such as Gaussian dispersion models, inadequately represent the complexities of atmospheric methane transport governed by the advection-diffusion partial differential equation (PDE), particularly in calm wind conditions. Second, these approaches struggle with multi-source emissions due to the increased complexity within the traditional inversion framework. Here, we propose to use physics-informed neural networks (PINN) to address the source localization and quantification problem. Specifically, a neural network is used to parameterize the solution to the advection diffusion PDE. Concurrently, the source term in the PDE is also treated as a trainable variable which is inverted from the data observed from CMS sensors. Importantly, we leverage the sparsity of potential emission sources to constrain the source term, achieving more accurate localization and quantification results. Finally, we implement a variety of techniques to enhance the performance of the PINN model such as embedding input coordinate variables, using adaptive weighting schemes in the loss function, and respecting causality in training the time-dependent PDE. We evaluate our model’s performance on controlled release data and also present initial results from a real O&G site. iPoster

A data-driven algorithm to optimize the placement of continuous monitoring sensors on oil and gas sites

December 11, 2023

Talk, AGU Fall Meeting 2023, San Francisco, California

Abstract: Methane, the main part of natural gas, is the second biggest cause of climate change after carbon dioxide. Even though it doesn’t stick around as long, it traps heat even better. So, if we can cut down on methane quickly, it can really help slow down global warming. Oil and gas sector gives off about 14% of total methane, so reducing how much they emit can make a big difference. For methane monitoring, we usually put multiple sensors around the site to catch all the emissions. But this method might not work best because it doesn’t take into account the emission and wind behaviors at that specific site. In our research, we’ve come up with a smarter way to place these sensors. We use data to help decide where to put them based on where and how emission occurs and how wind moves around the site. This method involves three steps: modeling how wind and emission behaves and generating emission scenarios based on this information, evaluating emission detection at each possible sensor location, and figuring out the best locations to put the sensors that work together for maximum emission monitoring performance.
Slides iPoster

A data-driven algorithm to optimize the placement of continuous monitoring sensors on oil and gas sites

November 15, 2023

Talk, Air Quality Measurement Methods and Techonology, Durham, NC

Abstract: Methane emissions from the oil and gas sector exhibit high temporal variability, and infrequent, short-lived super emitter events represent a large portion of overall emissions based on current understanding. Therefore, continuous monitoring systems (CMS) will likely play an increasing role in emissions monitoring, as they provide the near real-time measurements that are necessary for capturing both highly variable emissions and short-lived events. The placement of CMS on a given facility is a critical consideration. The most commonly used approach is fence line placement, where a set of sensors are positioned around the perimeter of a site to ideally encompass all potential emissions within. However, this intuitive approach may not be the most efficient, as it does not account for the actual emission patterns and wind distributions specific to a site. In this study, we propose a data-driven algorithm to optimize sensor placement based on emissions and wind distribution which we demonstrate for a specific site. The sensor placement optimization algorithm comprises three steps: wind and emission distribution modeling, concentration simulation and sensor detection, and sensor placement optimization. In the first step, neural networks are employed to approximate the wind distribution on the site using historical wind records, considering temporal correlation. Additionally, an emission event distribution is developed based on empirical emission logs. We then sample many emission scenarios from the modeled wind and emission distributions. In the second step, methane concentrations at all candidate sensor locations are simulated using the Gaussian puff model for all sampled emission scenarios. The sensor detection status is determined by comparing the simulated concentrations at candidate sensors with a detection threshold. In the last step, given the number of available sensors (N), an objective function is defined to maximize the number of detected emission scenarios by at least one sensor from the N-sensor placement. Efficient numerical optimization algorithms can be employed to achieve this task. Alternatively, the algorithm can be adapted to find the minimum number of sensors and their corresponding placement solution that satisfies a given detection probability, and we will show practical results for these different objectives.
Slides

Methane emission detection, localization, and quantification using continuous point-sensors on oil and gas facilities

June 02, 2023

Poster, International Indian Statistical Association Annual Conference, Golden, CO

Abstract: Reducing methane emissions is a key component of short-term climate action. The oil and gas sector provides a promising avenue for methane emission reduction, as it accounts for 22% of global anthropogenic methane emissions and 32% within the U.S. We propose a generic, modular framework for emission event detection (estimate emission start and end time), localization (estimate emission source), and quantification (estimate emission rate) on oil and gas facilities. The framework uses methane concentration and wind speed and direction data collected by continuous point-sensors. The framework is separated into four steps: 1) background removal and event detection, 2) atmospheric transport simulation, 3) source localization, and 4) emission rate quantification. We evaluate our framework by testing it on a set of 85 controlled releases that vary in duration and size. The framework identifies all controlled releases, with 82% being localized correctly. 90% of small events (<= 1kg/hr) are quantified within an error range of [-78.1%, 178.6%], while 90% of large events (> 1kg/hr) are quantified within an error range of [-49.6%, 77.4%]. Based on the framework’s performance, it appears to be useful for near real- time alerting for rapid emissions mitigation and emission quantification for data-driven inventory estimation on oil and gas production sites.
Poster

Determining Crust and Upper Mantle Structure by Bayesian Joint Inversion of Receiver Functions and Surface Wave Dispersion at a Single Station: Preparation for Data from the InSight Mission

December 17, 2017

Poster, AGU Fall Meeting 2017, San Francisco, California

Abstract: The InSight (Interior Exploration using Seismic Investigations, Geodesy and Heat Transport) mission will deploy a geophysical station on Mars in 2018. Using seismology to explore the interior structure of the Mars is one of the main targets, and as part of the mission, we will use 3-component seismic data to constrain the crust and upper mantle structure including P and S wave velocities and densities underneath the station. We will apply a reversible jump Markov chain Monte Carlo algorithm in the transdimensional hierarchical Bayesian inversion framework, in which the number of parameters in the model space and the noise level of the observed data are also treated as unknowns in the inversion process. Bayesian based methods produce an ensemble of models which can be analyzed to quantify uncertainties and trade-offs of the model parameters. In order to get better resolution, we will simultaneously invert three different types of seismic data: receiver functions, surface wave dispersion (SWD), and ZH ratios. Because the InSight mission will only deliver a single seismic station to Mars, and both the source location and the interior structure will be unknown, we will jointly invert the ray parameter in our approach. In preparation for this work, we first verify our approach by using a set of synthetic data. We find that SWD can constrain the absolute value of velocities while receiver functions constrain the discontinuities. By joint inversion, the velocity structure in the crust and upper mantle is well recovered. Then, we apply our approach to real data from an earth-based seismic station BFO located in Black Forest Observatory in Germany, as already used in a demonstration study for single station location methods. From the comparison of the results, our hierarchical treatment shows its advantage over the conventional method in which the noise level of observed data is fixed as a prior.
Poster