NEAR REAL-TIME CO 2EMISSIONS BASED ON CARBON SATELLITE ANDARTIFICIAL INTELLIGENCE A P REPRINT

2025-05-02 0 0 1.89MB 11 页 10玖币

侵权投诉

NEAR REAL-TIME CO2EMISSIONS BASED ON CARBON

SATELLITE AND ARTIFICIAL INTELLIGENCE

A PREPRINT

Zhengwen Zhang1, Jinjin Gu2, Junhua Zhao1,3,∗

, Jianwei Huang1,3, Haifeng Wu4

1School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China

2School of Electrical and Information Engineering, The University of Sydney, Sydney, Australia

3Shenzhen Institute of Artiﬁcial Intelligence and Robotics for Society (AIRS), Shenzhen, China

4Shenzhen Finance Institute, Chinese University of Hong Kong, Shenzhen, China

October 25, 2022

ABSTRACT

To limit global warming to pre-industrial levels, global governments, industry and academia are taking

aggressive efforts to reduce carbon emissions. The evaluation of anthropogenic carbon dioxide (CO

)

emissions, however, depends on the self-reporting information that is not always reliable. Society

need to develop an objective, independent, and generalized system to meter CO

emissions. Satellite

observation from space that reports column-average regional CO

dry-air mole fractions has

gradually indicated its potential to build such a system. Nevertheless, estimating anthropogenic CO

emissions from CO

observing satellite is bottlenecked by the inﬂuence of the highly complicated

physical characteristics of atmospheric activities. Here we provide the ﬁrst method that combines

the advanced artiﬁcial intelligence (AI) techniques and the carbon satellite monitor to quantify

anthropogenic CO

emissions. We propose an integral AI based pipeline that contains both a data

retrieval algorithm and a two-step data-driven solution. First, the data retrieval algorithm can generate

effective datasets from multi-modal data including carbon satellite, the information of carbon sources,

and several environmental factors. Second, the two-step data-driven solution that applies the powerful

representation of deep learning techniques to learn to quantify anthropogenic CO

emissions from

satellite CO

observation with other factors. Our work unmasks the potential of quantifying CO

emissions based on the combination of deep learning algorithms and the carbon satellite monitor.

Keywords Carbon satellite ·Anthropogenic CO2emissions ·Artiﬁcial Intelligence

1 Introduction

Human activities contribute to climate change by changing the amount of greenhouse gases, aerosols (ﬁne particulate

matter) and clouds in the atmosphere. The biggest contributor is the burning of fossil fuels Olivier et al. [2005], which

emit carbon dioxide (CO

) gas into the atmosphere. Reducing these emissions is a core objective of the Paris Agreement

of the United Nations Framework Convention on Climate Change (UNFCCC). The Paris Agreement requires all parties

to report anthropogenic greenhouse gas emissions and removals at least every 2 years. In addition, an increasing number

of corporates are required to provide annual report for CO2emission accounts.

Although there are well-established self-reporting mechanisms, we still need alternative carbon emission observation

methods to provide validation for the reported data and eliminate potential biases. Quantifying anthropogenic (CO

)

emissions at the individual facility level has important implications. It can help to monitor emissions reductions or

support regulation of carbon trading/pricing systems or other mitigation strategies. At present, most existing carbon

emission estimation methods are based on corporates’ self-reporting information Gurney et al. [2021], regional or sector

level carbon emission factors Shan et al. [2016], and other publicly available statistics data Shan et al. [2020]. However,

∗Corresponding author

arXiv:2210.09850v2 [physics.ao-ph] 22 Oct 2022

Near Real-time CO2Emissions Based on Carbon Satellite And Artiﬁcial Intelligence A PREPRINT

these methods need extra veriﬁcation and the estimation frequency is unable to meet the requirements for real-time

carbon emission estimation. In some industrial sectors, continuous emission monitoring systems (CEMS) are being

deployed for accurate and real-time carbon emission estimation Jahnke [1997]. But considering their high cost, it is

impractical to build a comprehensive carbon emission monitoring system based on CEMS.

Recently, carbon satellite monitoring is a new technical means that can directly provide CO

column-average dry-air

mole fractions (XCO

), which denotes the column-average regional CO

concentration in atmosphere. Owing to the

potential of supporting CO

emissions estimation, carbon satellite monitoring has been a fast-growing research topic.

Previous studies qualiﬁed CO2emissions by building a Gaussian plume model to simulate the CO2ﬂux movement in

the atmosphere Nassar et al. [2017]. In this paper, to estimate carbon emissions from carbon satellite observations at

the individual facility level, we propose a pure data-driven methods based on a novel deep learning algorithm. To the

best of our knowledge, this is the ﬁrst carbon measurement method based on carbon satellite data and state-of-the-art

artiﬁcial intelligence technologies. We anticipate this work to open up new research paradigms in carbon measurement.

Our research is based on the observations of NASA’s Orbiting Carbon Observatory 2 (OCO-2) satellite Crisp et al.

[2017]. OCO-2 makes high spectral resolution measurements of reﬂected solar radiation at wavelengths in the 0.76, 1.61,

and 2.06

m regions to derive XCO

. OCO-2 has limited imaging capabilities, it measures XCO

in 8 parallelogram

footprints (each is about 1.29 ×2.25 km2) over a narrow swath (less than 10.3 km).

plumes derived from large emission sources may cause local enhancement in the OCO-2 observational near-source

data because of diffusion and ﬂow of gases, separating it from the background XCO

. This local enhancement recorded

in satellite data is viewed as reﬂecting the patterns of CO

emissions and subsequent movement. Thus we can utilize

this pattern to estimate the emissions of carbon source. Detecting this enhancement and establishing the map from

near-source satellite data to CO

emissions are central to this task. We choose to design a deep neural network suitable

for carbon satellite’s unique data structure based on a Transformer architecture Vaswani et al. [2017]. After well-

training, given a range of carbon satellite data with local enhancement, carbon source location, and some environmental

information, this neural network directly predicts carbon emissions for the location it queried.

However, several challenges make this work not a naive utilization of deep learning on carbon satellite data.

•

First, the OCO-2 has only limited imaging capabilities, and its data is in the form of discrete measurement

locations and the carbon concentrations measured at that location. This presents a challenge to the design of

deep learning methods.

•

Second, the excess XCO

generated by large emission sources typically reaches 1% at the best, which is

about 4ppm compared with an instrument noise typically around 0.3–0.6ppm. This non-negligible noise in the

XCO

measurement hampers the imaging and detection of emission plumes and the precision of emission

quantiﬁcation. In the case of severe measurement noise, it is not even possible to conﬁrm whether there is

signiﬁcant local enhancement with a single measurement. Similar to the case of carbon measurements, it

is also extremely difﬁcult to obtain an accurate wind ﬁeld situation, which is another important basis for

quantifying the plumes caused by carbon emissions. Usually, we can only obtain average wind speed and

direction data over a large spatial scale and over a long period of time. And this only provides a very limited

clue for estimating the emission plume.

•

Third, data acquisition is also challenged due to space satellites’ unique measurement orbital limitations. Only

on rare occasions do the OCO-2 tracks cross CO

plumes downwind of large cities or power plants, limiting

the possibility of quantifying the corresponding CO

emissions to a few cases within a year. This also limits

our possibilities for multiple measurements of the same carbon source. By matching the location of known

carbon sources, wind direction, and the location of the satellite detection swath, we are only able to match the

hundreds of available data from millions of OCO-2 records, which also suffer from noise.

•

The last challenge is the complexity of emission source data. In this work, we select real hourly emission

sources from continuous emission monitoring system (CEMS) data for our research. However, not only are the

emissions data recorded by CMES facing missing and inaccurate challenges, but the situation of their emission

sources is sometimes extremely complex. Multiple closely located emission sources may inﬂuence each other,

making estimation more difﬁcult.

To solve above problem, this work makes a three-step data retrieval algorithm and a two-step data-driven solution based

on AI. The data retrieval algorithm is designed to ﬁnd effective satellite data that contains local enhancement and to

match auxiliary information geographically and temporally. We apply three steps to achieve this goals:(1) retrieval

based on carbon sources for extracting satellite data near carbon source; (2) processing of abnormal data for ﬁltering

extremely noisy data; (3) retrieval based on pattern detection for ﬁnding effective satellite data with local enhancement.

Near Real-time CO2Emissions Based on Carbon Satellite And Artiﬁcial Intelligence A PREPRINT

During data processing, we ﬁnd that there is scarcity of data with real emission label. In contrast, we have massive

amount of data without matched real emission label. To sufﬁciently utilize the properties of retrieved multi-modal data,

We propose a novel data-driven solution. Our solution includes two steps: (1) masked pre-training for carbon emission

estimation on large-scale of data without real emission label; (2) linear probing for the ﬁnal prediction on a small-scale

data with real emission labels.

The proposed data-driven method makes two important technical advances. Firstly, we present a new network

architecture called CarbonNet for carbon emission estimation from OCO-2 XCO

measurements. CarbonNet abandons

the traditional neural network design paradigm using fully-connected layers or convolutional layers and adopts a novel

strategy based on Transformer. The OCO-2 measurements can be viewed as independent measurements in different

locations. We treat each measurement as a token containing the location information, the XCO

number, the global

wind information and the statistical background XCO

number. In our CarbonNet, these tokens calculate self-attention

and interact with each other, i.e. each token can interact with others, greatly improving its computing efﬁciency. More

importantly, in CarbonNet, the order of tokens is variable, and tokens can be masked ﬂexibly. This enables our second

technical design called mask pre-training. Recall that of the vast amount of available satellite data, only a small fraction

can be matched with recorded emissions sources. We have very little satellite data with emission labels, but a lot of data

without labels. The distribution of these data is complex and full of noise. Learning directly with a small amount of

data will undoubtedly lead to severe overﬁtting. To address this problem, we propose a masked pre-training method

using satellite data that cannot match the emission record. For data without labels, we randomly mask a portion of

it, e.g., 25% of the measurement points, and train the CarbonNet to predict the masked measurements based on the

data provided. This forces the network to learn the distribution of satellite-measured XCO

and can learn an efﬁcient

representation of the satellite data. Finally, we use linear regression to predict carbon emissions from satellite data

representations on the labeled data. Our experiments show that the proposed solution can effectively predict carbon

emissions and can avoid overﬁtting and noise data interference to a certain extent.

2 Data

In this section, we ﬁrst introduce three types of raw data and their sources. Then we discuss the provided three-steps

data retrieval algorithm in detail. Finally we give the information of generated dataset.

2.1 Data Acquisition

We collect three different modal data that contributes to CO

emissions estimation: (1) carbon satellite observations;

(2) the position and emissions of carbon source; (3) some environmental information that inﬂuences CO

movements.

Firstly, we use version 9r of the OCO-2 bias-corrected XCO

retrievals, which is a level-2 satellite product that records

XCO

and a lot of supplementary information. We extract several key recordings from this data, including position,

time, XCO

, XCO

evaluation parameters, and recording angles of the carbon satellite. The position information is

consist of the center and corner coordinates (in the geographic coordinate system) of each scan area. XCO

is the

most important key, which reﬂects the column-averaged regional concentration of carbon dioxide in the atmosphere

detected by the carbon satellite. Two XCO

evaluation parameters in OCO-2 product (

xco2_quality_flag

and

xco2_uncertainty

) are used to evaluate the quality of XCO

recordings. Besides, we choose

solar_zenith_angle

and

sensor_zenith_angle

in recordings that can help us select worth data. The detailed explanations of these

selected keys can be referred to Eldering et al. [2017]. We could use these parameters both to ﬁnd effective data and to

estimate the emissions of carbon sources. On the other hand, we choose real hourly emission data from continuous

emissions monitors (CEMS) released by EPA Air Markets Program Data [EPA]. It covers over 1303 power plants in

the United States and their positions. Finally, some environmental information, including the average wind speed of u

and v direction below 50m, solar radiation and surface pressure, is derived from the ERA5 reanalysis dataset [C3S].

2.2 Data Retrieval

There are two reasons to develop a data retrieval algorithm before building model to estimate emissions. The ﬁrst reason

is that the quality of data is low, which indicates the low signal-noise rate of satellite observations, incomplete CEMS

emissions recordings and low resolution of environmental data. The second reason is that our inversion framework

relies on multi-modal data to jointly link the observed CO

local enhancement with upwind local emission source.

Matching of these data signiﬁcantly limit the the number of effective data. We did a lot of analysis and ﬁltering of

the data. The data retrieval is mainly composed of three steps: (1) retrieval based on carbon source; (2) processing

of abnormal data; (3) retrieval based on pattern (local enhancement in satellite data) detection. The details of this

three-step data retrieval algorithm are depicted as following.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

NEARREAL-TIMECO2EMISSIONSBASEDONCARBONSATELLITEANDARTIFICIALINTELLIGENCEAPREPRINTZhengwenZhang1,JinjinGu2,JunhuaZhao1;3;,JianweiHuang1;3,HaifengWu41SchoolofScienceandEngineering,TheChineseUniversityofHongKong,Shenzhen,China2SchoolofElectricalandInformationEngineering,TheUniversityofSydney,Sydney,Au...

展开>> 收起<<

NEAR REAL-TIME CO 2EMISSIONS BASED ON CARBON SATELLITE ANDARTIFICIAL INTELLIGENCE A P REPRINT.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

NEAR REAL-TIME CO 2EMISSIONS BASED ON CARBON SATELLITE ANDARTIFICIAL INTELLIGENCE A P REPRINT

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: