ARTIFICIAL INTELLIGENCE FOR CARBON

EMISSION USING SYSTEM OF SYSTEM

THEORY
1st DR.D.UMA NANDHINI 2nd Y.SINDHUJA REDDY 3rd P. PUJITHA 4th M.MONIKA SREE
( Professor) (CSE-AIML) (CSE-AIML) (CSE)
VEL TECH UNIVERSITY VEL TECH UNIVERSITY VEL TECH UNIVERSITY VEL TECH UNIVERSITY
Chennai, India Chennai, India Chennai, India Chennai, India
[email protected] [email protected] [email protected] [email protected]

5th S.VAMSI KRISHNA

(CSE)
VEL TECH UNIVERSITY
Chennai, India
[email protected]

Abstract—The quick development of Artificial Intelligence (AI) advanced means to reduce them through monitoring and
has brought about a massive increase in energy usage, especially optimization [7]. Uses vary from optimizing energy efficiency
during model training and deployment of large models. Such to monitoring emissions through satellite data [2].To address
models are heavily dependent on high-performance computing
hardware like Graphics Processing Units (GPUs) and Tensor these issues, the project employs a SoS strategy incorpo-
Processing Units (TPUs), which require a lot of electrical power rating energy monitoring, carbon accounting, and green AI
and add to growing levels of carbon emissions. With AI systems workflows. The integrative approach is designed to keep the
becoming more sophisticated and ubiquitous, the environmental environmental impact of AI development to a minimum while
impact of their development and operation is also becoming a preserving performance, keeping innovation aligned with in-
central issue. The purpose of this project is to rigorously investi-
gate the enviromental influence of AI training by calculating its ternational climate objectives and advancing the vision for
carbon impact and highlighting the most significant contributing Green AI [6].Additionally, the implementation of ML into
variables, including model design, hardware selection, data center SoS paradigms supports dynamic adaptation and feedback
utilization, and energy procurement. Drawing from a System of among diverse subsystems. With ongoing learning from op-
Systems (SoS) theoretical framework, this study will combine erational information, ML models are able to adapt strategies
interdisciplinary pieces—such as machine learning, energy man-
agement, environmental science, and policy analysis—to consider to decrease emissions, real-time energy optimization, and aid
holistic and scalable solutions for AI carbon reduction. proactive green policies. As a result, ML is no longer merely
Index Terms—KEYWORDS:System of System a method of analysis but also a driving force toward disruptive
theory(SoS),Carbon impact, carbon emission change in green AI approaches.

I. I NTRODUCTION II. LITERATURE REVIEW

Machine Learning (ML) is an important contributor to the Artificial Intelligence (AI) and Machine Learning (ML)
project ”Artificial Intelligence for Carbon Emission Using have grown dramatically, transforming domains like energy
System of System Theory” because it can handle massive, management, environmental observation, and automation. This
intricate datasets from interrelated systems. Within a System achievement, however, comes at enormous environmental ex-
of Systems (SoS) system, where sectors such as energy, pense based on the power-hungry process of training large-
transport, and industry are interconnected, ML can detect scale AI models. Scholars are now heavily concerned with
emission trends, make future output forecasts, and provide quantifying and addressing the carbon price of AI systems.
real-time intelligence for better decision-making and efficiency The infrastructure behind AI systems, particularly data
in resource usage. Its flexibility and scalability enable the centers, has a significant impact on carbon emissions. Wang
development of smart, green solutions. and Chen [6] evaluated the environmental footprint of AI data
Yet, training such large-scale AI models is also associated centers and urged better energy consumption and infrastructure
with high energy requirements, particularly when driven by redesign. Zhou and Li [7] investigated ways in which AI
GPUs and TPUs, causing high emissions and environmental can optimize energy efficiency and lower emissions through
issues[10]. Whereas AI generates emissions, it also provides real-time management of intelligent energy systems.Technical
reports such as Patterson et AI. [10] and Henderson et al. [11] The interconnectedness of these systems explains the so-
offer extensive methodologies for quantifying the carbon and phisticated nature of emissions and how they might be con-
energy footprints of training AI models, providing baselines trolled. For instance, the decarbonization of energy systems
for green practices. Previous seminal work by Lacoste et al. through using renewable sources can decrease both transporta-
[12] and Strubell et al. [13] also brought attention to the policy tion and industrial sector emissions substantially as they rely
and energy issues of deep learning, and especially in natural heavily on energy. Likewise, enhancing urban infrastructure
language processing. Finally, Dhar [14] offered one of the through green transport means and effective energy con-
earliest examinations of AI’s carbon impact, urging developers sumption can decrease city carbon emissions. Cross-sectoral
and researchers to consider environmental factors from the approaches, including the integration of renewable energy into
outset of system design. Together, these studies highlight industry and applying the principles of the circular economy,
the need for balancing technology progress with ecological can also decrease emissions. Successful policy and governance
sustainability. They form a solid basis for further investigation that promotes cooperation among these systems are critical to
into a System of Systems (SoS) strategy combining carbon obtain end-to-end carbon reduction and sustainability objec-
accounting, energy monitoring, and streamlined AI workflows tives.
to mitigate emissions while preserving technological perfor-
mance and scalability. B. Data Collection and Preprocessing
Real-time Data Acquisition: To construct a solid system
III. METHODOLOGY
for tracking carbon emissions, there is a need to gather real-
SoS theory associated with AI applied to Carbon Emission time environmental data using the application of advanced
mitigation combines AI with various subsystems for tracking, IoT sensors and big data technologies. These sensors can
monitoring, and reducing carbon emissions. AI analyzes enor- be used in different subsystems, including transportation,
mous datasets, automates energy management, and aids in the industrial processes, and energy consumption, to continuously
strategic selection processes. SoS theory links self-governing monitor data concerning carbon output. Furthermore, open
systems such as transportation and electricity networks inter- datasets, such as the CHDD (Chronic Heart Disease Dataset)
dependently working towards a goal: emission reduction. This or other expert private datasets, may offer useful multi-feature
strategy offers real-time observation and analysis, predictive data with both real-time measurements and past trends for
analytics, advanced steering, and sustainable environmental carbon emissions. The data can also be supplemented by other
control. environmental parameters, including temperature, humidity, or
air quality indices, to give a holistic picture of the patterns of
emissions. By combining IoT sensors with big data infras-
tructure, organizations can have the assurance of data being
collected efficiently and in real-time so that they are able to
make smart decisions on minimizing their carbon footprint and
adherence to environmental rules and regulations.
Data Preprocessing: Once data is obtained, the following
step is of prime importance where data is preprocessed to
secure its quality and aptness for training AI models. It entails
detecting and filling in missing or incomplete data, cleaning
Fig. 1. Feature importance - random forest. inconsistencies due to sensor faults or data transmission errors,
and normalizing the data to keep it uniform from more
than one source. Data preprocessing methods like imputation
A. System Modeling and Analysis of missing values, outlier detection, and normalization are
The categorization of systems accounting for carbon emis- utilized to improve accuracy and reliability of the dataset.
sions consists of significant sectors like energy, transportation, Ensuring the data is effectively cleaned and normalized makes
industry, agriculture, and urban systems. Energy systems, it even more machine-learning-algorithm-compliant, ensuring
power generation, and fuel production account for consider- the creation of predictive models capable of accurately pro-
able emissions, particularly where dependent on fossil fuels jecting carbon emissions using different factors. By taking
like oil and coal. Transportation systems like road, air, and adequate care of the preprocessing phase, organizations can
shipping contribute to substantial emissions using carbon- escape biases or inaccuracy that may compromise the usability
emitting fuels. Industrial systems, e.g., manufacturing and of AI solutions for carbon management.
mining, also release considerable carbon as a result of energy-
hungry processes. Agricultural and land use systems also add C. AI Model Development and Analysis
their share in the form of deforestation, cattle rearing, and use Predictive modeling and optimization are two techniques of
of fertilizers. Urban systems, such as infrastructure, waste, AI that can heavily help in analyzing and enhancing efforts
and transportation, also play a role by burning energy and towards sustainability. Using a myriad of machine learning
producing waste. algorithms like classification models, we can study large
datasets to forecast emission patterns and determine possible circumstances and determine the most successful means of
avenues of carbon reduction. For example, regression analysis, lowering emissions.
decision trees, or ensembling methods can be used to analyze Strategic Decision Support is a central capability of the
past data on emissions, energy use, and other environmental platform, designed to enable decision-makers with actionable
indicators to predict future trends. These are then used to intelligence based on AI-driven analysis and scenario simula-
inform the development of policy and assist companies in tion. Through the processing of several possible avenues for
streamlining their activities towards reduced carbon emissions. carbon reduction, the system offers insightful recommenda-
Furthermore, optimization techniques like genetic algorithms tions that can inform leadership in policy setting, investment
or simulated annealing may be applied to determine the most in clean technologies, and prioritization of initiatives based
suitable methods for minimizing emissions across varying on their environmental and economic effects. Decision-makers
variables, for instance, production activities or transportation will be able to use these insights to make informed, data-
networks, in order to enhance efficiency and sustainability. driven decisions that guarantee long-term sustainability and
When dealing with environmental data, some of the frequent adherence to international climate goals. The capacity to model
challenges include class imbalance, especially where carbon- various scenarios also enables a more adaptive and resilient
intensive events or particular emissions-related occurrences are strategy, allowing for flexibility to change strategies as new
uncommon. In such instances, machine learning models may data and technologies become available, ultimately resulting
have difficulty identifying key but underrepresented examples, in more significant and effective carbon emissions reduction
resulting in biased predictions. A strong remedy to this is throughout the system of systems (SoS).
the Synthetic Minority Over-sampling Technique (SMOTE).
SMOTE addresses this issue by creating synthetic examples E. Feedback Loop and Continuous Improvement
of the minority class, enabling the model to identify better The real-time monitoring and assessment system creates
the rare but significant events related to high emissions. By an ongoing feedback cycle that monitors the performance of
balancing the distribution of data, SMOTE enhances model adopted carbon emission mitigation strategies. By gathering
performance so that it can effectively identify emissions and evaluating real-time data from multiple sources, including
spikes, carbon-heavy behaviors, or other vital conditions that sensors, environmental monitoring systems, and operational
could otherwise be missed. This improves the model’s capacity performance measurements, the platform ensures that actual
to provide actionable insights for targeted reduction strategies results of emissions reduction efforts are correctly captured
and ultimately leads to more effective environmental interven- and evaluated. This constant monitoring enables quick de-
tions. tection of any divergence from desired outcomes, allowing
swift adjustment to strategies to maximize results. By making
available to decision-makers real-time intelligence regarding
the success of the mitigation initiatives, the system facilitates
ensuring that emission-reduction goals are always achieved
and adjustments are implemented as needed to enhance results.
Adaptive tuning of strategy is a central feature of this
feedback system, enabling dynamic adjustment of AI models
and mitigation strategies in accordance with the real-time data
collected. Since the feedback loop continuously offers insights
Fig. 2. Total power consumption. into the performance of different strategies, the AI models can
be tuned and recalibrated to improve their predictive accuracy
and effectiveness. This continuous process of optimization
D. Decision Support and Scenario Analysis guarantees that carbon reduction actions are always synchro-
The Simulation and Analysis Platform takes advantage of nized with changing conditions, emerging technologies, and
cutting-edge AI algorithms to simulate and evaluate numerous fresh environmental information. Through the integration of
scenarios for carbon emissions reduction within an existing adaptive learning methods, the system not only enhances the
system or operating paradigm. Through the incorporation of accuracy of future projections but also optimizes the overall
massive amounts of environmental information, the platform performance of mitigation actions, resulting in a more robust
is able to test different parameters like energy usage patterns, and efficient strategy for carbon emission reduction. This
industrial processes, and transport systems, while considering ongoing improvement creates a cycle of improvement, where
the likely effect of varying mitigation methods. These sim- each step taken results in more informed decisions and further
ulations deliver real-time evidence of the potential of certain progress towards combating climate change.
policies, technological advances, or shifts in practices to lower
carbon emissions over a period. Additionally, the platform F. Implementation and Validation
features machine learning models that are trained on past Pilot testing serves as an initial key step toward the deploy-
data and make estimates of future emissions, allowing users ment of AI-based carbon emission reduction strategies, as it
to model the results of a range of approaches under varying provides a test bed to test the feasibility and performance of
 After pilot testing has proved the efficacy of the AI-based
solutions in carbon emission reduction, the next is scalable
deployment in other subsystems of the SoS. The pilot tests’
success offers a basis for scaling the use of the best practices
and optimized strategies to a wider scale, with the assurance
that the best solutions are implemented to achieve meaningful
carbon emission reductions. The process of scaling entails
modifying the AI systems to operate in perfect harmony
across multiple subsystems while maintaining consistency,
reliability, and compatibility. As the solutions are rolled out on
a wider scale, they will continue to be supported by continuous
validation, refinement, and integration with legacy systems.
The outcome is a more sustainable, data-driven method of
carbon emission reduction that can be successfully scaled
Fig. 3. statistics of sos.
across various sectors, resulting in systemic and enduring
reductions in carbon footprints across the entire network or
organization.

G. Model Interpretability and Transparency (SHAP)

Including SHAP (Shapley Additive Explanations) in AI-

based carbon emission forecasting models provides trans-
parency and interpretability, which are crucial for building
trust in AI systems. SHAP gives an explanation of the contri-
bution of every feature or variable in a model’s prediction by
providing a ”Shapley value” to every input. This enables stake-
holders to see how various factors, including energy usage,
transportation, or industrial processes, affect carbon emissions.
By providing simple, quantifiable reasons for model outputs,
SHAP helps users understand which variables have the greatest
effect on predictions of emissions, thereby demystifying the AI
decision-making process. The explanation is especially useful
in complicated systems where carbon emission drivers may
be difficult to discern, as it bridges the gap between machine
learning models and the effective implementation of their
knowledge in emission abatement policies.
The actionable understanding that SHAP delivers not only
makes AI models more interpretable but also augments user
Fig. 4. Number parameters in billion confidence in the system by giving unambiguous reasons
behind the decisions that have been taken. Knowing which
specific drivers cause emission predictions will allow organi-
suggested strategies. With the choice of specific subsystems zations to target their reduction efforts at those most impactful
within the SoS, firms can pilot-test the AI-supported solutions determinants, aligning strategies for maximum impact against
in the real world by emphasizing their capabilities to predict the most determinant areas. Furthermore, SHAP explanations
accurately, manage, and minimize carbon emissions. At this can facilitate regulatory compliance by providing a breakdown,
stage, the performance of the system can be closely tracked explainable description of how emission cuts are being real-
to determine any technical problems, possible limitations, or ized and what factors are driving progress. This transparency
unforeseen effects that could occur. The feedback from these guarantees not only the effectiveness of the AI system but
pilot tests assists in streamlining the system and making sure also accountability, allowing stakeholders to interact, monitor,
that the AI models and approaches are feasible, scalable, and verify the outputs of the system more easily. Ultimately,
and consistent with the general sustainability objectives. Pilot SHAP increases the usability and value of AI-driven solutions
testing not only ensures the efficacy of the solution but also by making the outputs more understandable and actionable for
offers significant insights into any operational issues that must decision-makers, increasing confidence in the system and its
be solved before large-scale deployment. potential to drive significant carbon emission reductions.
H. Sustainability Considerations for AI the sustainability of the AI-led mitigation efforts.While AI has
Energy efficiency analysis is an essential part of verifying potential to help achieve carbon emission savings, the research
that AI models, especially deep learning models, are aligned identified that the potential depends on the energy efficiency of
with sustainability objectives within the system of systems AI models and their deployment. System of Systems Analysis:
(SoS). Whereas AI-based carbon reduction measures have Interaction Between AI and Environmental Systems .The
the potential to deliver significant environmental gains, the application of the System of Systems (SoS) theory provided
computational resources involved in training and deploying deeper insights into the interdependencies between AI tech-
deep learning models tend to have their own energy expenses. nologies and environmental factors. The key findings from this
Through a review of the energy use and carbon footprint analysis are as follows. AI technologies that help maximize
of these models, organizations are able to pinpoint where energy consumption in industrial processes can reduce carbon
optimization should take place to reduce their footprint. This emissions. But the large amounts of energy used to train AI
encompasses examining the data center power consumption, models, particularly deep learning models, form a feedback
the hardware efficiency being utilized, and computational loop that may offset some of these gains unless sustainable
resources necessary for training models. By measuring these methods are incorporated into AI development.Environmental
factors, one can make it a point that the AI systems are not Impact of AI Hardware: The carbon footprint of the hardware
causing the very issue they are trying to mitigate by default, infrastructure underpinning AI systems, including data centers,
thus getting both the mitigation action and the operational was discovered to be a significant proportion of the overall
processes pointed in a direction of low-carbon. carbon footprint. Energy-consuming processes and cooling
Green AI practices are critical to the extent that AI sys- demands of large-scale data centers additionally increased the
tems deployed for reducing carbon emissions not only yield environmental footprint of AI deployments.Sustainability of
effective outcomes but also reduce their own footprint on AI Models .A key discovery of the research was the need to
the environment. This is possible by creating energy-efficient create sustainable AI models with energy efficiency as a top
algorithms that minimize computation without affecting per- priority.
formance, like through model pruning or neural network
architecture optimization to make it more efficient. Moreover,
using renewable sources of energy, such as solar or wind, to
power AI infrastructure reduces the carbon footprint associated V. ACKNOWLEDGMENT
with operating such systems. Additionally, taking a complete
lifecycle carbon footprint approach to AI systems—from
I would like to thank everyone who has supported and
the production of hardware to deployment and later decom-
guided me during the course of completion of this research
missioning—is crucial to long-term sustainability. Executing
work on Artificial Intelligence for Carbon Emission Reduction
methods like recycling hardware and reducing the carbon
Using a System of Systems Approach.Above all, I would like
footprint of data processing and storage can guarantee that AI-
to express my sincerest gratitude to my project supervisor,
driven mitigation methods are sustainable. By incorporating
Dr.D. Uma Nandhini Mam, for her exceptional guidance, valu-
these sustainability factors into AI system development and
able feedback, and constant encouragement. Her knowledge
operation, organizations can establish a more comprehensive
and support have been pivotal in charting the course of this
carbon reduction strategy that goes beyond the near-term
study and sharpening its results.I am thankful to Vel Tech
objectives of emissions management and includes the overall
University as well for offering the required academic materials,
environmental effect of AI technologies.
computational equipment, and great research environment that
IV. E XPERIMENTAL RESULTS AND ANALYSIS facilitated the smooth completion of this project. Support
AI Model Carbon Emissions Model Interpretability and from peers and faculty members has been greatly valuable
Transparency (SHAP) Explainable AI: Use SHAP (Shapley in improving my knowledge of AI sustainability and energy-
Additive Explanations) to offer transparency and interpretabil- efficient computing.
ity to the predictions of the AI models on carbon emissions. In addition, I would like to acknowledge the efforts of the
Actionable Insights and Trust: Use SHAP insights to explain wider research community, including the creators of open-
what drives emission projections, improve user trust in the AI source tools like CodeCarbon, MLCO2, and Carbontracker,
system, and supply actionable information to direct mitigation which made it possible to estimate carbon footprints accurately
strategies and regulatory requirements. in AI model training. They provided an essential basis for
Sustainability Considerations for AI Energy Efficiency this work.Last but not least, I thank my family and friends
Review: Review the energy use and carbon emissions relating for their constant encouragement, motivation, and patience
to the use of AI models within the SoS, including deep during this process. Their support has been a cornerstone in
learning models. Sustainable AI Practices: Highlight the sig- successfully completing this research. This work is an effort
nificance of creating energy-efficient algorithms, powering AI towards making AI more sustainable, and I hope it is helpful in
infrastructure with renewable sources of energy, and taking the contributing to the development of environmentally sustainable
lifecycle carbon management of AI into account to maintain artificial intelligence technologies.
 VI. REFERENCES
R EFERENCES
[1] Zhao, X., Wang, Y. (2025). Can Artificial Intelligence Technology
Reduce Carbon Emissions? A Global Perspective. Energy Economics,
143, 108285. https://doi.org/10.1016/j.eneco.2024.108285
[2] Akinlabi, E.T., Akinlabi, S. (2024). Monitoring Carbon Emissions
Using Deep Learning and Statistical Process Control: A Strategy
for Impact Assessment of Governments’ Carbon Reduction Poli-
cies. Environmental Monitoring and Assessment, 196, Article 231.
https://doi.org/10.1007/s10661-024-11356-3
[3] Zhong, J., Zhong, Y., Han, M., Yang, T., Zhang, Q.
(2024). The Impact of AI on Carbon Emissions: Evidence
from 66 Countries. Applied Economics, 56(25), 2975–2989.
https://doi.org/10.1080/00036846.2024.1002376
[4] Luccioni, S., Dandres, T. (2023). Quantifying the Carbon Footprint of
AI: A Case Study on BLOOM. Journal of Machine Learning Research,
24(1), 1–25.
[5] Wang, L., Chen, H. (2023). AI and Energy Consumption: Evaluating the
Environmental Impact of AI Data Centers. Energy Policy, 165, 112927.
https://doi.org/10.1016/j.enpol.2022.112927
[6] Zhou, F., Li, X. (2023). AI in Energy Management: Reducing Carbon
Footprint through Intelligent Systems. Renewable and Sustainable En-
ergy Reviews, 152, 111678. https://doi.org/10.1016/j.rser.2021.111678
[7] Schwartz, R., Dodge, J., Smith, N.A., Etzioni, O. (2022). Green AI: Pro-
moting Environmental Sustainability in AI Research. Communications
of the ACM, 65(12), 27–29. https://doi.org/10.1145/3490895
[8] Hao, K. (2022). Training a Single AI Model Can Emit as Much
Carbon as Five Cars in Their Lifetimes. MIT Technology Review.
https://www.technologyreview.com
[9] Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M.,
Rothchild, D., So, D., Texier, M., Dean, J. (2021). Carbon Emissions
and Large Neural Network Training. arXiv preprint arXiv:2104.10350.
https://arxiv.org/abs/2104.10350
[10] Henderson, P., Hu, J., Romoff, J., Brunskill, E., Jurafsky, D., Pineau,
J. (2020). Towards the Systematic Reporting of the Energy and Carbon
Footprints of Machine Learning. Journal of Machine Learning Research,
21(1), 1–43.
[11] Lacoste, A., Luccioni, A., Schmidt, V., Dandres, T. (2019). Quan-
tifying the Carbon Emissions of Machine Learning. arXiv preprint
arXiv:1910.09700. https://arxiv.org/abs/1910.09700
[12] Strubell, E., Ganesh, A., McCallum, A. (2019). Energy and Policy
Considerations for Deep Learning in NLP. Proceedings of the 57th
Annual Meeting of the Association for Computational Linguistics,
3645–3650. https://doi.org/10.18653/v1/P19-1355
[13] Dhar, P. (2018). The Carbon Impact of Artificial Intelligence. Nature Ma-
chine Intelligence, 1(1), 1–3. https://doi.org/10.1038/s42256-018-0001-0