PaperlayerView exact revision on arXiv

Source-linked AI summary

Carbon Footprint Accounting Driven by Large Language Models and Retrieval-augmented Generation

Haijin Wang, Mianrong Zhang, Zheng Chen, Nan Shang, Shangheng Yao, Fushuan Wen, Junhua Zhao

arXiv:2408.09713v2cs.IR

TL;DR

Traditional LCA-based carbon accounting relies on expert effort and is difficult to update in real time as production and related information change. LLMs-RAG-CFA combines LLMs with retrieval-augmented generation to provide current information for accounting; across five industries, the authors report higher retrieval rates and lower information and accounting deviations than other LLMs.

  • Problem

    Traditional LCA-based CFA depends on manual expert work, making it difficult to reflect real-time changes in production processes and environmental impacts.

    Source: §I · p. 1, A. Introduction of Carbon Footprint… · p. 2

  • Method

    LLMs-RAG-CFA retrieves relevant carbon-footprint information with RAG to supplement LLM prompts for automated accounting.

    Source: §I · p. 2

  • Results

    Across five industries, LLMs-RAG-CFA achieved higher information retrieval rates and significantly lower information and CFA deviations than other LLMs.

    Source: §VI · p. 10

  • Takeaways & Limitations

    The system’s RAG-based design balances real-time updates with cost-effectiveness for carbon-emission management.

    Source: §I · p. 2, §VI · p. 10

Abstract

from arXiv · show

Carbon footprint accounting is crucial for quantifying greenhouse gas emissions and achieving carbon neutrality.The dynamic nature of processes, accounting rules, carbon-related policies, and energy supply structures necessitates real-time updates of CFA. Traditional life cycle assessment methods rely heavily on human expertise, making near-real-time updates challenging. This paper introduces a novel approach integrating large language models (LLMs) with retrieval-augmented generation technology to enhance the real-time, professional, and economical aspects of carbon footprint information retrieval and analysis. By leveraging LLMs' logical and language understanding abilities and RAG's efficient retrieval capabilities, the proposed method LLMs-RAG-CFA can retrieve more relevant professional information to assist LLMs, enhancing the model's generative abilities. This method offers broad professional coverage, efficient real-time carbon footprint information acquisition and accounting, and cost-effective automation without frequent LLMs' parameter updates. Experimental results across five industries(primary aluminum, lithium battery, photovoltaic, new energy vehicles, and transformers)demonstrate that the LLMs-RAG-CFA method outperforms traditional methods and other LLMs, achieving higher information retrieval rates and significantly lower information deviations and carbon footprint accounting deviations. The economically viable design utilizes RAG technology to balance real-time updates with cost-effectiveness, providing an efficient, reliable, and cost-saving solution for real-time carbon emission management, thereby enhancing environmental sustainability practices.

I. INTRODUCTION

Traditional LCA-based carbon footprint accounting depends on extensive data collection and expert analysis, making near-real-time updates difficult. The paper proposes combining LLMs with RAG to retrieve current professional information and support automated, cost-effective accounting.

I. INTRODUCTION

  • LCA-based CFA relies on extensive data collection and expert experience, making it difficult to reflect real-time production conditions.The paper links this limitation to challenges for carbon-reduction decisions.Source: §I · pp. 1–2
  • LLMs-RAG-CFA combines large language models with retrieval-augmented generation to improve the efficiency, reliability, and cost-effectiveness of real-time CFA.The approach is intended to reflect emissions under current production structures and support timely carbon-management decisions.Source: §I · p. 2
  • RAG retrieves current reference information to enrich LLM prompts, helping capture information across standards, methods, policies, industry processes, and technical literature.The paper says this supports real-time information acquisition and accounting while reducing reliance on human expertise.Source: §I · p. 2
  • RAG dynamically retrieves relevant professional information from a preconstructed knowledge base to supplement prompts and improve the reliability and professionalism of CFA.Source: §I · p. 2
  • The method avoids frequent LLM parameter updates by combining high-frequency RAG and LLMs for automated, scenario-specific carbon-footprint information acquisition.The paper presents this as a low-cost way to balance real-time updates with cost-effectiveness across contexts.Source: §I · p. 2

A. Introduction of Carbon Footprint Accounting

Carbon footprint accounting quantifies greenhouse gas emissions, commonly using life cycle assessment, but traditional LCA-based accounting depends on manual expert work. The section also reviews LLM capabilities and limitations relevant to automating CFA.

A. Introduction of Carbon Footprint Accounting

  • CFA quantifies greenhouse gas emissions attributable to activities, products, processes, or an entity, supporting carbon-reduction targets and mitigation strategies.Source: A. Introduction of Carbon Footprint… · p. 2
  • LCA is a widely adopted CFA approach that can span a product’s life cycle, while some studies limit the boundary to cradle-to-gate.Its stages include defining the system boundary, inventory analysis, and impact assessment, with emissions converted to CO2-eq for aggregation.Source: A. Introduction of Carbon Footprint… · p. 2
  • Traditional LCA-based CFA depends on experts manually selecting and analyzing information, making scaling, automation, and real-time updates difficult.The resulting estimates may not reflect changes in production processes and environmental impacts.Source: A. Introduction of Carbon Footprint… · p. 2

B. Artificial Intelligence and Large Language Models

III. PROPOSED METHOD

LLMs-RAG-CFA automates key CFA tasks by retrieving relevant carbon-footprint fragments and combining them with user queries to guide LLM responses. It supports different data-source lengths, draws on updated professional sources, and uses vector similarity to retrieve relevant text.

A. Roadmap

  • The framework automates inventory construction and emissions acquisition, then uses rule-based calculation models to calculate product carbon footprints.It adapts prompt construction to different lengths of carbon-footprint data sources.Source: A. Roadmap · pp. 3–4
  • Users submit a data source and natural-language query; retrieved information is used to generate inventory and emissions data for a rule-based model to calculate the product footprint.The system can also generate the life-cycle inventory and impact assessment.Source: A. Roadmap · p. 4

B. Data Preprocessing and Text Segmentation

  • Carbon-footprint sources include user uploads and material retrieved from related websites, including regulations, papers, reports, standards, and policies that can be continuously updated.The documents provide a data foundation for professional coverage and real-time capabilities.Source: B. Data Preprocessing and Text… · p. 4

C. Text Vectorization and Vector Similarity Calculation

  • The vectorization model uses paired related and unrelated fragments with twin BERT encoders, bringing related embeddings closer and unrelated embeddings farther apart.Text fragments and user queries are represented in a shared semantic space to support similarity-based retrieval.Source: C. Text Vectorization and Vector… · p. 4
  • The system retrieves the Top-K fragments with the highest similarity to a user query and uses them to enrich LLM prompts.Similarity can be calculated using dot product or cosine similarity.Source: C. Text Vectorization and Vector… · pp. 4–5

D. Knowledge Fusion and Answer Generation

The section describes how relevant carbon-footprint fragments are combined with a user query to condition LLM-generated CFA information. Transformer components process the enriched input, and generated data is passed to a rule-based accounting model.

D. Knowledge Fusion and Answer Generation

  • Top-K relevant carbon-footprint fragments are fused with the user query into an enhanced prompt for LLM generation of CFA-related information.The system selects relevant fragments as reference information before combining them with the query.Source: D. Knowledge Fusion and Answer… · p. 5
  • The enriched input is processed by a Transformer built from encoder and decoder layers with multi-head self-attention and feed-forward neural networks.The section identifies these components as enabling text processing and generation.Source: D. Knowledge Fusion and Answer… · p. 5, Figure 4 · p. 6
  • Multi-head attention lets the model consider different input aspects simultaneously and capture a range of semantic relationships among carbon-footprint fragments.Each attention head uses learned weight matrices applied to queries, keys, and values.Source: D. Knowledge Fusion and Answer… · p. 6
  • After attention, the output passes through a ReLU feed-forward network with residual connections and layer normalization.The section describes the feed-forward network as introducing non-linearity and enhancing representational capacity.Source: D. Knowledge Fusion and Answer… · p. 6
  • The generated data corresponding to the user query is used in a rule-based accounting model to obtain the CFA result.Source: D. Knowledge Fusion and Answer… · p. 6

E. Economic and Real-Time CFA based on LLM and RAG

LLMs-RAG-CFA uses RAG to incorporate relevant carbon-footprint information in real time rather than relying on frequent parameter updates. Its handling of source texts varies by length, and the paper reports improved retrieval efficiency and support for lifecycle inventory analysis.

E. Economic and Real-Time CFA based on LLM and RAG

  • RAG incorporates real-time relevant carbon-footprint information, reducing the need for frequent LLM parameter updates and their computational and financial burden.The paper links this design to lower costs while maintaining model effectiveness and improving real-time capabilities.Source: E. Economic and Real-Time CFA based on… · p. 6
  • For long-text sources, the process preprocesses, segments, vectorizes, and similarity-searches text, then combines selected fragments with the user query in an enhanced prompt.The LLM uses that prompt to generate more comprehensive and professionally accurate CFA-related information.Source: E. Economic and Real-Time CFA based on… · p. 7, Figure 5 · p. 7
  • For short-text sources, the text directly serves as reference information; without a source, the system queries the LLM using only the user query.Source: E. Economic and Real-Time CFA based on… · p. 7
  • Overall, LLMs-RAG-CFA improves CFA-related information retrieval efficiency and supports lifecycle inventory analysis with real-time datasource.The paper states that this reflects the carbon footprint of production scenarios.Source: E. Economic and Real-Time CFA based on… · p. 7

A. Scenarios and Benchmark

The case study evaluates carbon-footprint information retrieval and accounting across five industries. It compares LLMs-RAG-CFA, using GPT-4, with several other models against expert-compiled answers from a shared datasource.

A. Scenarios and Benchmark

  • The case study covers primary aluminum, lithium batteries, photovoltaics, new energy vehicles, and transformers.Source: A. Scenarios and Benchmark · p. 7
  • Benchmarks include ERNIE Bot, GPT-4, ChatGLM 3, Qwen-14B-Chat, and iFlytekSpark-13B-base; GPT-4 serves as the LLM within LLMs-RAG-CFA.Source: A. Scenarios and Benchmark · p. 7
  • Each method receives the same literature-based industry datasource, and its generated information and results are compared with expert-extracted answers from that datasource.Source: A. Scenarios and Benchmark · p. 7

B. Evaluation Index

The evaluation uses information retrieval rate (IRR), information deviation (ID), and accounting deviation (AD) to assess CFA methods. Together, they cover retrieval, information accuracy, and accounting accuracy.

B. Evaluation Index

  • IRR measures relevant carbon-footprint data retrieved relative to the relevant data available, with higher rates indicating broader retrieval.The paper notes that IRR also reflects reduced reliance on manual information extraction.Source: B. Evaluation Index · pp. 7–8
  • ID measures the discrepancy between retrieved information and true data, using mean absolute percentage error (MAPE).The true data are based on aggregated expert opinions, with outlier opinions discarded and the average taken.Source: B. Evaluation Index · p. 8
  • AD measures the difference between the CFA result calculated from retrieved data and the true carbon footprint; lower AD means the result is closer to the true value.Source: B. Evaluation Index · p. 8

V. RESULT ANALYSIS

Across five industries, LLMs-RAG-CFA retrieved carbon-footprint information at higher rates than other methodologies. Its retrieval performance was attributed to combining LLM capabilities with RAG-based domain-specific information retrieval.

A. Information Retrieval Rate

  • Across industries, LLMs-RAG-CFA consistently outperformed other methodologies in integrating and retrieving CFA-related carbon-footprint information.Source: A. Information Retrieval Rate · p. 8
  • LLMs-RAG-CFA achieved 100% IRR in aluminum, lithium battery, photovoltaic, and transformer industries, and 96.36% in new energy vehicles.It outperformed other methods in the first four industries; transformer GPT-4 also reached 100%.Source: A. Information Retrieval Rate · p. 8
  • The authors attribute the broader specialized-data coverage to RAG adding targeted, domain-specific information retrieval to LLMs’ generalist capabilities.Source: A. Information Retrieval Rate · p. 8

B. Information Deviation Analysis

Information deviation (ID) is presented as a distribution across methods within each industry, with retrieved ranges assessed using their upper and lower boundaries. The reported ranges for LLMs-RAG-CFA were lower than those for other methods across the industries described.

B. Information Deviation Analysis

C. Carbon Footprint Accounting Deviation

LLMs-RAG-CFA showed lower carbon-footprint accounting deviations than other methods across five industries. The conclusion describes its LLM-and-RAG design as supporting real-time CFA and reports efficiency and cost-effectiveness gains.

C. Carbon Footprint Accounting Deviation

  • Primary aluminum AD was 2.35% to 19.07% (other methods: -206.10% to 269.36%); lithium battery AD was -0.07% to 0.07% (others: -80.65% to 559%).These are the reported ranges for LLMs-RAG-CFA and other methods, respectively.Source: C. Carbon Footprint Accounting Deviation · p. 10
  • Photovoltaic AD was -1.11% to 1.11% (other methods: -92.45% to 165.80%); new energy vehicle AD was 13.77% to -13.77% (others: 48.06% to 127.94%).The passage reports these as the respective ranges for LLMs-RAG-CFA and other methods.Source: C. Carbon Footprint Accounting Deviation · p. 10
  • Transformer AD was -0.01% to 0.01% for LLMs-RAG-CFA, compared with -38.85% to 60.94% for other methods.The paper highlights this result in the context of energy management and loss calculations.Source: C. Carbon Footprint Accounting Deviation · p. 10

VI. CONCLUSION

  • The system combines LLM language and reasoning capabilities with RAG retrieval to dynamically obtain professional CFA information and form enhanced prompts.The conclusion describes this design as expanding professional coverage while supporting real-time processing and cost efficiency.Source: §VI · p. 10
  • Across five industries, the system closely matched manual-method reliability while improving efficiency and cost-effectiveness; it also achieved higher information retrieval rates and lower information and CFA deviations than other LLMs.The paper attributes lower accounting errors to high information retrieval rates and low information deviations, and reports reduced human effort and time.Source: §VI · p. 10, C. Carbon Footprint Accounting Deviation · p. 10

C. Carbon Footprint Accounting Deviation

  • Across five industries, LLMs-RAG-CFA showed lower accounting deviations than other methods, indicating more accurate and reliable CFA.Figure 12 compares accounting deviations; the reported ranges cover primary aluminum, lithium batteries, photovoltaics, new energy vehicles, and transformers.Source: Figure 12 · p. 10, C. Carbon Footprint Accounting Deviation · p. 10

VI. CONCLUSION

  • Future research could explore multi-agent systems to enhance the system’s intelligence and adaptability for CFA.The paper proposes that such systems could coordinate CFA tasks and improve real-time data processing and decision-making.Source: §VI · p. 11
Loading 2408.09713v2…