《美國國家可再生能源實驗室：2024年eGridGPT：值得信賴的控制室AI助手研究報告（英文版）（26頁）.pdf》由會員分享，可在線閱讀，更多相關《美國國家可再生能源實驗室：2024年eGridGPT：值得信賴的控制室AI助手研究報告（英文版）（26頁）.pdf（26頁珍藏版）》請在三個皮匠報告上搜索。

1、NREL is a national laboratory of the U.S.Department of Energy Office of Energy Efficiency&Renewable Energy Operated by the Alliance for Sustainable Energy,LLC This report is available at no cost from the National Renewable Energy Laboratory(NREL)at www.nrel.gov/publications.Contract No.DE-AC36-08GO2

2、8308 Technical Report NREL/TP-5D00-87440 May 2024 eGridGPT:Trustworthy AI in the Control Room Seong Lok Choi,1 Rishabh Jain,1 Patrick Emami,1 Karin Wadsack,1 Fei Ding,1 Hongfei Sun,1 Kenny Gruchalla,1 Junho Hong,2 Hongming Zhang,3 Xiangqi Zhu,4 and Benjamin Kroposki1 1 National Renewable Energy Labo

3、ratory 2 University of Michigan 3 Lower Colorado River Authority 4 Oregon State University NREL is a national laboratory of the U.S.Department of Energy Office of Energy Efficiency&Renewable Energy Operated by the Alliance for Sustainable Energy,LLC This report is available at no cost from the Natio

4、nal Renewable Energy Laboratory(NREL)at www.nrel.gov/publications.Contract No.DE-AC36-08GO28308 National Renewable Energy Laboratory 15013 Denver West Parkway Golden,CO 80401 303-275-3000 www.nrel.gov Technical Report NREL/TP-5D00-87440 May 2024 eGridGPT:Trustworthy AI in the Control Room Seong Lok

5、Choi,1 Rishabh Jain,1 Patrick Emami,1 Karin Wadsack,1 Fei Ding,1 Hongfei Sun,1 Kenny Gruchalla,1 Junho Hong,2 Hongming Zhang,3 Xiangqi Zhu,4 and Benjamin Kroposki1 1 National Renewable Energy Laboratory 2 University of Michigan 3 Lower Colorado River Authority 4 Oregon State University Suggested Cit

6、ation Choi,Seong Lok,Rishabh Jain,Patrick Emami,Karin Wadsack,Fei Ding,Hongfei Sun,Kenny Gruchalla,Junho Hong,Hongming Zhang,Xiangqi Zhu,and Benjamin Kroposki.2024.eGridGPT:Trustworthy AI in the Control Room.Golden,CO:National Renewable Energy Laboratory.NREL/TP-5D00-87440.https:/www.nrel.gov/docs/f

7、y24osti/87440.pdf.NOTICE This work was authored in part by the National Renewable Energy Laboratory,operated by Alliance for Sustainable Energy,LLC,for the U.S.Department of Energy(DOE)under Contract No.DE-AC36-08GO28308.Funding provided by the U.S.Department of Energy Office of Electricity.The view

8、s expressed herein do not necessarily represent the views of the DOE or the U.S.Government.This report is available at no cost from the National Renewable Energy Laboratory(NREL)at www.nrel.gov/publications.U.S.Department of Energy(DOE)reports produced after 1991 and a growing number of pre-1991 doc

9、uments are available free via www.OSTI.gov.Cover Photos by Dennis Schroeder:(clockwise,left to right)NREL 51934,NREL 45897,NREL 42160,NREL 45891,NREL 48097,NREL 46526.NREL prints on paper that contains recycled content.iii This report is available at no cost from the National Renewable Energy Labora

10、tory at www.nrel.gov/publications.Acknowledgments The authors acknowledge Ali Ghassemian,Sandra Jenkins,and Roshanak Nateghi of the U.S.Department of Energy(DOE)for their continued support and leadership in the development of high-impact generative artificial intelligence research in the United Stat

11、es.The authors offer special thanks to our industry members,who provided valuable insights during an interview process as well as technical reviews of this report,including(listed alphabetically by last name):Brad Bouillon,California Independent System Operator(CAISO)Russell Boyer,Dell Technologies

12、Mark Daus,Dell Technologies Venkata Tirupati,Electric Reliability Council of Texas(ERCOT)Ming Wu,Microsoft.Input from the industry members and other reviewers does not constitute an endorsement of this report or its contents.Further,the views expressed herein do not necessarily represent the views o

13、f DOE or the U.S.Government.The authors take sole responsibility for the contents of this report.iv This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.List of Acronyms AI artificial intelligence CAISO California Independent System Operator

14、DOE U.S.Department of Energy eGridGPT Electric Grid Generative Pretrained Transformer EMS energy management system ERCOT Electric Reliability Council of Texas GenAI generative artificial intelligence G-PST Global Power System Transformation Consortium GPT generative pretrained transformer Llama Larg

15、e Language Model Meta AI LLM large language model ML machine learning NERC North American Electric Reliability Corporation NIST National Institute of Standards and Technology NREL National Renewable Energy Laboratory SCADA supervisory control and data acquisition v This report is available at no cos

16、t from the National Renewable Energy Laboratory at www.nrel.gov/publications.Executive Summary This report is the first research effort to apply large language models(LLMs),a type of generative artificial intelligence(GenAI),in the power grid control room for decision making.Serving as the operation

17、al“brain”of the grid to balance supply and demand,operators decision making is crucial for maintaining grid reliability from one moment to the next.Just as the human brain processes inputs from the senses to make decisions,the National Renewable Energy Laboratory Electric Grid Generative Pretrained

18、Transformer(eGridGPT)is engineered to virtually support power grid control room operators by assisting in decision-making processes and interpreting the data and models.Thanks to the favorable economics of new technologies and their ability to deliver secure,reliable,and resilient energy,the power s

19、ector is undergoing a significant transition,characterized by three major shifts at its core.First is the increasing use of variable renewable energy,such as wind and solar,to help decarbonize the energy sector.Second is the proliferation of distributed energy resources,including rooftop solar photo

20、voltaics and distributed energy storage,and the increasing use of electrified loads,such as electric vehicles and heat pumps.Third is the digitization of power system communications and controls.As the power sector transitions,it presents substantial operational and external challenges to grid opera

21、tors.Inverter-based resourcessuch as solar,wind,and batteriesintroduce new operational challenges because they behave differently from synchronous generators and push the limits of managing increasingly complex networks.Externally,operators must also navigate the grid management challenges of the im

22、pacts of increasingly frequent extreme weather events and cyber-attack by hostile nations or other malevolent actors.Grid operators are at the forefront of this shift.These challenges test the ability of grid operators to make real-time decisions safely,efficiently,and reliably while meeting decarbo

23、nization goals and evolving customer needs.The critical question that emerges is:How can researchers assist operators decision making?One viable approach to assist operators is the broader implementation of artificial intelligence(AI).LLMs,a type of GenAI,are computational tools that excel at langua

24、ge processing and general-purpose tasks.LLMs,such as OpenAIs GPT-4 or Metas Llama 3(Large Language Model Meta AI),represent a remarkable breakthrough in AI by helping with increasingly complex tasks.This report describes the first research effort to apply GenAI in the power grid control room.It outl

25、ines the synergy between human decision making and eGridGPT,where eGridGPT supports operators by analyzing procedures,suggesting actions,simulating scenarios with physics-based digital twins,and recommending optimal decisions.The system operators can then make decisions on how to adjust the grid or

26、not based on the suggestions.A human system operator is placed in the final decision loop because eGridGPT is not legally accountable or able to automatically implement suggestions.Figure 1 shows an imagined control room of the future with an AI-based assistant to help make suggestions on system ope

27、rations.The report also presents the results of a preliminary case study showing the ability of eGridGPT to handle an equipment model mapping task between real-time operations and offline planning.vi This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/pu

28、blications.Figure 1.Imagined control room of the future with AI-based assistant Figure from Benjamin Kroposki and Seong Lok Choi,NREL,and DALL-E In addition,the report addresses the challenges and limitations of GenAI,stressing the importance of model accuracy and high-quality data,ethical AI use,tr

29、ustworthiness of the suggested actions,and cybersecurity concerns.The report also emphasizes low-budget eGridGPT solutions for smaller utilities that can be implemented on-premises with limited resources.The report concludes that the integration of eGridGPT in the power grid control room is importan

30、t to help industry transition to successful operations of a more complex clean,resilient,and affordable future energy system.vii This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Table of Contents Executive Summary.v 1 Introduction.1 2 Ove

31、rview of eGridGPT.6 2.1 Trustworthiness of eGridGPT.7 2.2 Training eGridGPT Inferences.7 2.3 Attesting eGridGPT by Digital Twin.9 2.4 Physics-Informed eGridGPT.9 2.5 Explainable eGridGPT.9 3 Preliminary Result for Equipment Model Mapping Between Real-time Operation and Offline Planning.10 4 Challeng

32、es and Limitations.14 4.1 Fundamental Advances in GenAI.14 4.2 Cybersecurity.14 4.3 Ethical AI.14 4.4 Low-Budget On-Premises Solutions for Small Utilities.14 5 Concluding Remarks.15 viii This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Li

33、st of Figures Figure 1.Imagined control room of the future with AI-based assistant.vi Figure 2.Power grid control room as the operational“brain”of the grid to balance supply and demand.1 Figure 3.NERC Energy Emergency Alerts Level 3.3 Figure 4.eGridGPT architecture,an AI-based decision support syste

34、m for grid control rooms.4 Figure 5.Three-step training process for eGridGPT assets:(1)Train on general power engineering knowledge.(2)Train on control room operational procedures and standards.(3)Fine-tune using utility-specific data.8 Figure 6.Example prompt instructing an LLM(Mistral 7B)on how to

35、 map between EMS operational models and planning models by identifying similar name fields.11 Figure 7.Comparison of power flow on a major path/flow gate interface and 500-kV bus frequency for a specific system event,simulated using the eGridGPT converted base case versus the EMS state estimation ca

36、se.The close match validates the accuracy of eGridGPT model mapping.13 List of Tables Table 1.Sample of Mismatched Equipment Names and Bus Numbers Between EMS Operational Models and Planning Models Illustrating the Model Mapping Challenge.10 Table 2.LLM-Generated Equipment Model Mapping Results With

37、 Substation Names Masked for Security.12 1 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.1 Introduction The primary goal of the power industry is to operate a safe and reliable electrical power grid.This“keeping the lights on around th

38、e clock”approach has been the bedrock of modern society.Power system reliability requires three major steps:First is to know how much load there will be moment by moment,second is to schedule the power generation to meet the load,and third is to identify which lines and associate equipment will be u

39、sed for the power to flow from the generation to the load.All these steps are managed inside the power grid control room.The control room,therefore,is the brain of grid operation;see Figure 2.It is similar to how the human brain processes information from the eyes and ears and makes decisions on how

40、 to act.The control room receives grid operational data from the generation,transmission,and distribution systems.It also receives look-ahead planning data,such as outage,load,interchange schedule,generation,and weather forecast.Inside the control room,system operators or dispatchers play a unique r

41、ole in decision making,while information technology personnel and engineers support the operators by maintaining critical hardware infrastructure,analytical software tools,and data and ensuring they are providing essential information.Figure 2.Power grid control room as the operational“brain”of the

42、grid to balance supply and demand Image from Christopher Schwing,NREL As a decision maker,transmission operators are well prepared,having passed the North American Electric Reliability Corporation(NERC)system operator exam and undergone extensive training specific to their grid network.Additionally,

43、they are certified by fellow operators,and they collaborate to reach consensus on operational decisions(NERC 2023b).In 2 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.short,they are well trained to keep the lights on by balancing gener

44、ation and load from one moment to the next and proactively looking ahead for potential grid problems.Over the years,the power sector has seen significant changes due to the cost-effectiveness of new technologies and their capacity to provide secure,reliable,and resilient energy.These changes are dri

45、ven by three primary trends.First is the increasing use of low-cost,variable renewable energy,such as wind and solar,to help decarbonize the energy sector.Second is the proliferation of distributed energy resources,including rooftop solar photovoltaics and distributed energy storage,and the increasi

46、ng use of electrified loads,such as electric vehicles and heat pumps.Third is the digitization of power system communications and controls such as smart grid or wide area monitoring systems.These changes are all aimed at improving system operational efficiency and reducing the carbon footprint.As th

47、e power sector transitions,it presents substantial operational and external challenges to grid operators.They are at the forefront of this shift,managing the transition from traditional power operationscharacterized by one-way flow from large-scale generation to the distribution level,fixed power di

48、spatch capacity,and a predictable dispatch scheduleto more dynamic and complex system.The future grid will feature more bidirectional flow from customers back to substations,variable generation output that fluctuates with weather conditions,and less visibility of behind-the-meter customer generation

49、 resources,making the resources non-dispatchable from the control room.In addition to adapting to these operational changes,grid operators in the control rooms also face nonoperational but externally critical emerging threats,such as extreme weather conditions and cyberattacks.Decision-making tool c

50、hallenges:The responsibility for decision making in grid operations still predominantly rests with system operators.This is largely because the technology for helping decision making,whether software or hardware,is not yet sufficiently advanced for deployment in control rooms.Consequently,operators

51、must depend on their experience,memory,and available tools when making decisions.What tools do they have?Presently,the decision-making tools are mainly supervisory control and data acquisition(SCADA)systems and energy management systems(EMS).SCADA systems are used to monitor and control the grid,whe

52、reas EMS provides advanced computations and visualizations of the current and contingent states of the system.Simply monitoring and measuring data is insufficient for effective grid control.It is crucial to transform this data into actionable information,enhancing situational awareness for operators

53、,particularly in energy emergency scenarios.Additionally,SCADA and EMS were designed for conventional generators that provide stable,dispatchable,and weather-independent output.As a result,the current tools do not adequately support operators when dealing with high levels of integration of variable

54、renewable energy,which is variable,non-dispatchable,and dependent on weather conditions.Another challenge has been that earlier versions of SCADA and EMS did not include battery modeling within the generator elements.With the growing increase in the use of batteries and other energy storage systems

55、to provide grid support,especially during periods when solar PV generation declines(like at sunset),vendors of SCADA/EMS have updated their software to account for the state of charging and discharging of batteries.To take advantage of this new feature,the control room must upgrade their SCADA/EMS,w

56、hich can be years of lengthy and complex process.3 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Decision-making regulatory challenges:Even with a new tool,the operators decisions need to comply with massive volumes of protocols and mu

57、ltiple governance requirements.For instance,the latest NERC standard is 1,926 pages(NERC 2023a).Combined with regional electric reliability organization standards,independent system operator technical manuals,and utility standards and procedures,grid operators need to review several thousand pages o

58、f documentation and be able to quickly retrieve this information.This increases the margin of error for making decisions,especially under time-critical events.Motivation:Numerous challenges,from operational to regulatory,make it hard to prepare grid operators for the power system transformation duri

59、ng the clean energy transition.Although grid operators build knowledge over time to improve decision making,the evolving grid complexity hinders this.If there is an unpredictable load pattern due to the weather,it requires an uncertainty quantification of the imbalance risk between the generation an

60、d load.For example,on July 26,2021,the California Independent System Operator(CAISO)issued energy emergency alerts from 6 p.m.10 p.m.asking the public to reduce their energy demand against a heat wave(CAISO 2021)and there are increasing numbers of critical energy emergency alerts(EEA)over the last 5

61、 years(NERC 2023c).Figure 3 shows EEA level 3 issuance,which represents actual energy deficiency,has risen from 6 in 2017 to 25 in 2023.Moreover,with the rapid growth of variable renewable energy sources,control rooms urgently need advanced decision-making solutions to manage the variability and unc

62、ertainty of the power output introduced by weather while still operating traditional synchronous generators.The crucial issue at hand is:How can researchers assist operators with these challenges?Figure 3.NERC Energy Emergency Alerts Level 3 4 This report is available at no cost from the National Re

63、newable Energy Laboratory at www.nrel.gov/publications.Generative AI:To address these challenges,we reviewed the use of generative artificial intelligence(GenAI).GenAI,most notably through large language models(LLMs),has revolutionized the field of natural language processing and machine learning(ML

64、)in general,especially with the introduction of transformer models(Vaswani et al.2017).The generative pretrained transformer(GPT)(Brown et al.2020)family of models has demonstrated remarkable achievements across various domains,including passing medical exams,scoring high on standardized tests,compo

65、sing music,and creating art(OpenAI 2023).Exploring how next-generation GenAI can increase community preparedness for climate-related risks,enable clean energy deployment,and enhance grid reliability and resilience is a considerable yet critical endeavor(The White House 2023).Decision-making platform

66、:With GenAI and classical analytical systems used in the power industry today,we envision the development of a trustworthy Electric Grid Generative Pretrained Transformer(eGridGPT).eGridGPT builds on the ability of ML and LLMs.Understanding natural language and preserving the context is key to enhan

67、cing situational awareness and tailoring the response to the needs of grid operators.eGridGPT is aimed as an ever-evolving ecosystem for control room operators that can adapt the available tools to their needs and grid operation requirements.It is the underlying platform being developed at the Natio

68、nal Renewable Energy Laboratory(NREL)that integrates LLMs,digital twins,and advanced visualizations to provide holistic decision support for grid operators and improve their situational awareness;see Figure 4.There are two reasons for introducing eGridGPT:(1)the trustworthiness of GenAI response and

69、(2)the advanced display capability by adopting the multimodality of GenAI from the operators queries and prompts.As defined by the National Institute of Standards and Technology(NIST),“trustworthy AI is valid and reliable,safe,fair,and bias is managed,secure and resilient,accountable and transparent

70、,explainable and interpretable,and privacy-enhanced”(NIST 2022).In this report,we discuss how eGridGPT addresses several characteristics of AI trustworthy,such as validation and explainable.Figure 4.eGridGPT architecture,an AI-based decision support system for grid control rooms eGridGPT integrates

71、large language models,digital twin simulations,and advanced visualizations to provide holistic recommendations to grid operators.5 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.In this report:eGridGPT,a GenAI-based virtual operator ass

72、istant,can enable to(1)understand and respond to operator queries/prompts,(2)empower the analysis of grid conditions by augmenting digital twins and analytical tools based on the request,(3)orchestrate recommendations from the tools and relevant actions into an understandable display format,and(4)bu

73、ild operators trust with a human-in-the-loop framework where LLM recommendations are vetted by operators.The essence of eGridGPT lies in its ability to act as an interface between a screen in front of the operator and the orchestrator for the comprehensive processing of large volumes of data,scenari

74、os,and digital twin simulations.6 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.2 Overview of eGridGPT We developed eGridGPT to aid bulk energy system operators by providing a decision-making platform with advanced real-time grid analy

75、tics,rigorous simulations,and visualization tools(Figure 4).eGridGPT enables system operators to analyze events,find similarities to historical events,and assess mitigation options.Our vision is that eGridGPT will serve as the larger decision-making tool that will integrate and coordinate a host of

76、additional tools to verify and validate its outputs.Below are the steps in the eGridGPT workflow as shown in Figure 4.At step 1,the system operator initiates a query while,concurrently,real-time data from SCADA,state estimation and contingency analysis is being streamed from step 2.Step 3 involves t

77、he AI system analyzing the operators query alongside the real-time data to understand the context of the question.Since the AIs training is not real time,it uses various techniques to generate responses or suggestions appropriate for the situation,though these are not yet visible.In step 4,the AIs r

78、ecommendations are tested through simulation.This process,called reinforcement learning,requires simulation to verify the accuracy of the AIs suggestions.One major challenge with generative AI is ensuring accuracy where the absence of simulation can lead to a lack of“ground truth.”This lack of verif

79、ication through power flow models can result in inaccurate recommendations,highlighting the need for a fifth step involving simulation using a digital twin.During the sixth step,recommendations generated by the AI are refined through digital twin simulations,with only the most viable options being f

80、orwarded.The operator then assesses these options,determining their effectiveness.Step 7 allows operators to introduce alternative scenarios,such as a failure in communication equipment leading to a disruption in data communication.Simulations including these hypothetical situations are then conduct

81、ed.The outcomes of these simulations are documented in step 8 and subsequently evaluated in step 9 to identify any potential adverse effects on power operations.This involves an additional filtering process to exclude recommendations that could cause significant disruptions,like a major power outage

82、 due to the intentional disconnection of critical transmission lines.By step 10,only the recommendations that are considered safe are presented to the operator.Finally,in step 11,the operator reviews these vetted suggestions to make the final decision.The role of eGridGPT is to coordinate and refine

83、 the execution of the approach to process a given request and to guide system operators through challenging grid conditions.With additional insights,grid operators can make informed decisions to manage the grid and ensure reliable operations with high integration levels of renewable generation.Above

84、 all,the resulting system must be trustworthy.A key focus in developing eGridGPT is ensuring it embodies the principles of trustworthy AI development.As previously noted,NIST characterizes trustworthy AI as being“valid and reliable,safe,secure and resilient,accountable and transparent,explainable an

85、d interpretable,privacy-enhanced,and fair with managed harmful bias.”The following section of the report will detail how eGridGPT has implemented these principles in its decision-making framework.7 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publ

86、ications.2.1 Trustworthiness of eGridGPT eGridGPT incorporates several safeguards to be trustworthy and is being designed to ensure:Validity and reliability:The three-step training process(Section 2.2,Figure 5)and benchmarking against NERC system operator exams can ensure that eGridGPT outputs are a

87、ccurate and consistent.Digital twin verification(Section 2.3)provides further validation.Safety,security,and privacy:Restricting eGridGPT to open-source models that can run locally without public network access enables compliance with NERCs Critical Infrastructure Protection(CIP)standards.Additional

88、ly,following the best practices in NIST Cybersecurity framework will further mitigate any risks(NIST 2024).Utility-specific fine-tuning is performed on-premises with data access restricted to authorized personnel.Further measures to protect sensitive data used in training will also be explored and i

89、mplemented.Fairness and bias management:We are curating the training datasets to consciously detect and remove sources of bias throughout the process.Human oversight will also allow monitoring for unintended biases.Accountability and transparency:Human-in-the-loop decision making will ensure that gr

90、id operators vet the systems recommendations.Audit logs enable tracing actions back to the source.The outcomes will be marked with a timestamp and source(e.g.,eGridGPT vs.operator).Explainability and interpretability:The physics-informed approach(Section 2.4)and explainability(Section 2.5)will allow

91、 operators to understand the rationale behind eGridGPT outputs.Inconsistencies with physical laws can be identified.2.2 Training eGridGPT Inferences eGridGPT AI model is trained in three steps;see Figure 5:1.Train state-of-the-art LLM models on general power engineering knowledge,transmission bus sy

92、stem,or publicly available grid data,2.Train using control room operational procedures from NERC,electric reliability organizations,independent system operators,state public utility commissions,the Institute of Electrical and Electronics Engineers,and other U.S.grid standards,and,3.Train using super

93、vised fine-tuning on system operator/utility operational and management procedures,power system data,field settings,and infrastructure information.8 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Figure 5.Three-step training process for

94、 eGridGPT assets:(1)Train on general power engineering knowledge.(2)Train on control room operational procedures and standards.(3)Fine-tune using utility-specific data Illustration from Besiki Kazaishvili,NREL During the third step,information specific to each implementation is used by a system oper

95、ator to create a unique instance of eGridGPT.Utility-specific knowledge obtained from operational logs,tagging notes,historical and operational analysis reports,training documents,and others will be implemented via further fine-tuning or retrieval-augmented generation(Lewis et al.2020).The base LLMs

96、 for eGridGPT include only open-source models(e.g.,Llama 3,Mistral)(Rudolph,Tan,and Tan 2023),which can be run locally without public network access,to comply with NERC CIP standards.After the training,eGridGPT is benchmarked with the NERC system operator exams(NERC 2023b),which operators must also

97、take,to ensure its effectiveness.9 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.2.3 Attesting eGridGPT by Digital Twin One of the primary challenges with GenAI is its accuracy.Instances often arise during complex scenariosreferred to

98、as“hallucinations”where the absence of simulations can result in a lack of“ground truth”producing a faulty recommendation.With large quantities of high-quality training data,GenAI can achieve satisfactory levels of accuracy.In the power industry,data are not typically publicly shared for the purpose

99、s of research;thus,the datasets used to train eGridGPT will be a mixture of carefully curated sources,including both publicly and privately shared data and synthetic data generated by physics-based digital twin simulations.Without the ability to verify these scenarios through power flow models,the r

100、ecommendations could lack adequate creditability.To improve the accuracy of the eGridGPT response,digital twins can be used.Digital twins can evaluate the accuracy and applicability of the eGridGPT response after simulations with the latest model and data.2.4 Physics-Informed eGridGPT When making de

101、cisions,operators are required to document and justify their chosen actions.NERC CIP standards mandate the disclosure of details like measurement data,study findings,or network models,alongside a description of how consensus was reached among operators.This requirement for transparency,interpretabil

102、ity,and accountability should extend to GenAI-based decisions as well.Therefore,the predictions made by GenAI tools for power systems should be consistent with the fundamental laws of physics.Physics-informed eGridGPT,such as those that accurately simulate the flow of electricity through the grid,ca

103、n lead to more efficient,explainable,and reliable solutions.This can be achieved in a few different ways,such as by adopting physics-based digital twin simulations.2.5 Explainable eGridGPT As GenAI becomes more capable,it also becomes more complex,limiting the ability of users to understand how it a

104、rrives at a particular decision.This gap in understanding hinders trust.For these systems to be practically helpful,they should be able to make decisions that are physically grounded and human interpretable,and they should be able to provide explanations detailing the intermediate decisions that wer

105、e made along the way.eGridGPT interactive,dynamic display transforms how operators trust monitoring systems.It responds to the operators prompts,adapting continuously to the most current data and emphasizing critical information based on the alarms severity and the situations urgency powered by digi

106、tal twin.This feature streamlines the workflow for operators,enhancing their decision-making process by clearly presenting explainable recommendations.10 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.3 Preliminary Result for Equipment

107、Model Mapping Between Real-time Operation and Offline Planning This section describes the preliminary results for the equipment model mapping between the simulated real-time operations(node-breaker)and the offline planning(bus-branch).The offline planning cases are developed by regional entities wit

108、h data provided by the transmission system operators.A major issue in developing the offline planning cases is that the names and bus numbers in the real-time operation model usually do not match the planning names and bus numbers,as shown in Table 1.The worst case is when the name does not match wi

109、thin the utility department.Finally,offline planning cases that assume maximum power output from all generators and assume all transmission lines are energized face significant challenges in developing accurate scenarios for upcoming extreme weather events,such as hurricanes or ice storms,based sole

110、ly on the planning model;therefore,the offline planning cases must adjust their network topology with the EMS operational measurement and switch statuses.Table 1.Sample of Mismatched Equipment Names and Bus Numbers Between EMS Operational Models and Planning Models Illustrating the Model Mapping Cha

111、llenge EMS Operational Model Planning Model Integrating the offline planning with the real-time operating conditions and system responses after the equipment model mapping becomes crucial to managing the reliability and resilience of the power network.This integration,through the conversion of model

112、s between the real-time operations and the offline planning,enables a more realistic analysis of power systems.It ensures the accuracy of the initial operating conditions for downstream applications.This mapping conversion from the real-time operations to the offline planning process usually takes m

113、ultiple weeks by regional entity staff.NREL has developed a technique to generate large numbers of simulations of operational data using planning models that can provide valuable datasets for AI training.With pretrained datasets,eGridGPT is able to fine-tune mappings of the EMS operational models an

114、d the planning models,as shown Figure 6 and Table 2.11 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Figure 6 shows the prompt instruction on how to map between the EMS operational models and the planning models by identifying similar

115、name identification fields in the planning fields and name fields in the EMS,as shown in Table 2.Once the similar models name is identified,it will check to ensure that the area/ID_CO name is the same between the EMS operational models and the planning models.Figure 6.Example prompt instructing an L

116、LM(Mistral 7B)on how to map between EMS operational models and planning models by identifying similar name fields Table 2 shows the results from the LLM that are generated within a minutewhereas a manual mapping effort takes multiple weeks.The results are masked to protect the sensitivity of the sub

117、station name.12 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Table 2.LLM-Generated Equipment Model Mapping Results With Substation Names Masked for Security eGridGPT can perform this mapping within minutes compared to weeks required f

118、or manual efforts.(Masked owner-substation model name)To describe the value of using eGridGPT for real-time operations,Figure 7 shows the power flow on a major path/flow gate and the 500-kV bus frequency at a specific bus for two simulations.Based on the mapping result from eGridGPT,the planning cas

119、e adjusts the measurement or switch status from the EMS State Estimation Case.Both the Converted Case generated by eGridGPT and the State Estimation Case from the actual EMS model yield very similar results,which can satisfy NERC MOD-033(Steady-State and Dynamic System Model Validation)standards(NER

120、C n.d.b).13 This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Figure 7.Comparison of power flow on a major path/flow gate interface and 500-kV bus frequency for a specific system event,simulated using the eGridGPT converted base case versu

121、s the EMS state estimation case.The close match validates the accuracy of eGridGPT model mapping.With this capability,eGridGPT can bridge historical EMS datasets,which are solved every 5 minutes,with planning model and production cost model economic dispatch files to create 8,760 hourly cases to ass

122、ess grid reliability and resilience for future years.By leveraging the model mapping feature,the system operator can evaluate the impact of the resource mix changes,the distributed energy resource aggregations,the battery energy storage systems,and the electrification loads to future grid operations

123、 with expanded planning cases with higher granularity and fidelity.Once trained,eGridGPT can become a transferable resource that can significantly reduce the time to map distinct data sources and help the utility run analyses more efficiently and at scale.14 This report is available at no cost from

124、the National Renewable Energy Laboratory at www.nrel.gov/publications.4 Challenges and Limitations 4.1 Fundamental Advances in GenAI GenAI models for language and image data are trained to create realistic-looking outputs using large datasets mined from the internet.This is inadequate for achieving

125、goals of trustworthy GenAI for power system operations;thus,fundamental AI advances and alternative training data paradigms are needed to gain trust.For example,research is needed to enable eGridGPT to make factual predictions grounded in knowledge and to reduce uncertainty across the multiple spati

126、al and temporal scales at which the power grid operates.It would also be useful for the trained AI to be able to explain how it developed its suggestions,similar to the way an engineer can explain the process for getting a certain result.4.2 Cybersecurity Adopting eGridGPT in power system operations

127、 may introduce potential cybersecurity vulnerabilities that need to be carefully addressed.Any AI in power systems must comply with robust cybersecurity policies and standards due to the critical nature of this infrastructure.Adherence to frameworks such as the NERCs CIP standards and NIST Cybersecu

128、rity is necessary to mitigate cyber threats and maintain the integrity of grid operations.4.3 Ethical AI GenAI systems are powerful tools that come with both risks and opportunities to positively impact many real-world problems.The models and datasets selected to develop eGridGPT will be carefully v

129、alidated to detect sources of bias.In general,eGridGPT will be used in a human-in-the-loop framework where it generates recommendations vetted by system operators.Human oversight will be an integral part of keeping AI-driven processes ethical,practical,and safe.4.4 Low-Budget On-Premises Solutions f

130、or Small Utilities The base eGridGPT model will be computationally expensive to create.To reduce the burden on utility resources,the computing power for the initial training steps can be satisfied using a high-performance computer.NREL has used its Kestrel high-performance computer with 138 graphica

131、l processing units and NVIDIA H100 nodes for the examples in this report.The base model can then be delivered to utilities for utility-specific fine-tuning.Avoiding the costs of preparing a base model can save utilities months of time and millions of dollars in computing resources(NREL 2023).The uti

132、lity-specific fine-tuning of eGridGPT will ensure that the resulting models are locally owned and operated per the utilitys protocolsand accessible only from within the utility network to comply with NERC CIP standards(NERC n.d.a).15 This report is available at no cost from the National Renewable En

133、ergy Laboratory at www.nrel.gov/publications.5 Concluding Remarks This report introduces the eGridGPT concept as the first research effort to virtually support system operators in the power grid control room of the future.It outlines interactions between system operators and eGridGPT.It also explain

134、s how eGridGPT becomes trustworthy by describing training process,validating by digital twin,and recommending on a dynamic display based on operators prompt.The short example shows that eGridGPT was able to be trained to produce offline planning case results that were similar to the actual model sta

135、te estimation.The eGridGPT concept addresses the shortages of the current control room tools and technologies,enhancing the grids secure and reliable operation.eGridGPT is a step forward in responding to a request from the system operators who are founding members of the Global Power System Transfor

136、mation Consortium(G-PST)for the control room of the future to encompass intuitive visualization,AI/ML tools,and trustworthy decision support capabilities(G-PST n.d.).eGridGPT will be an essential tool for next-generation control room solutions.16 This report is available at no cost from the National

137、 Renewable Energy Laboratory at www.nrel.gov/publications.References Brown,T.,B.Mann,N.Ryder,M.Subbiah,J.D.Kaplan,P.Dhariwal,Neelakantan,P.Shyam,G.Sastry,A.Askell et al.2020.“Language Models are Few-Shot Learners.”Advances in Neural Information Processing Systems 33:1,8771,901.California Independent

138、 System Operator(CAISO).2021.“Energy Emergency Alert(EEA)Watch Declared.”July 26,2023.https:/ PST Consortium(G-PST).n.d.“Vision for the Control Room of the Future.”https:/globalpst.org/wp-content/uploads/G-PST-Vision-for-the-Control-Room-of-the-Future-V0.5-Final.pdf.Lewis,P.,E.Perez,A.Piktus,F.Petro

139、ni,V.Karpukhin,N.Goyal,H.Kttler,M.Lewis,W.-t.Yih,T.Rocktschel et al.2020.“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.”Advances in Neural Information Processing Systems 33:9,4599,474.National Institute of Standards and Technology(NIST).2022.“Trustworthy AI:Managing the Risks of

140、Artificial Intelligence.”Sept.29,2022.https:/www.nist.gov/speech-testimony/trustworthy-ai-managing-risks-artificial-intelligence.2024.“The NIST Cybersecurity Framework(CSF)2.0”https:/nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.29.pdf National Renewable Energy Laboratory(NREL).2023.“Kestrel:Why Speed Ma

141、tters.”Kestrel Blog,Aug.3,2023.https:/www.nrel.gov/hpc/announcements/posts/kestrel-blog-july-update.html.North American Electric Reliability Corporation(NERC).2023a.Reliability Standards for the Bulk Electric Systems of North America.Atlanta,GA.https:/ Operator Certification&Credential Maintenance P

142、rogram.Atlanta,GA.https:/ American Electric Reliability Corporation(NERC).n.d.a.“Cyber SecurityInformation Protection.”Atlanta,GA.https:/ and Dynamic System Model Validation.”Atlanta,GA.https:/ Energy Emergency Alerts.”Atlanta,GA.https:/ OpenAI.(2023),GPT-4.https:/ 17 This report is available at no

143、cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.Rudolph,J.,S.Tan,and S.Tan.2023.“War of the Chatbots:Bard,Bing Chat,ChatGPT,Ernie and BeyondThe New AI Gold Rush and Its Impact on Higher Education.”Journal of Applied Learning and Teaching 6(1).Vaswani,N.Shazeer,N.Parma

144、r,J.Uszkoreit,L.Jones,A.N.Gomez,.Kaiser,and I.Polosukhin.2017.“Attention Is All You Need.”Advances in Neural Information Processing Systems 30.The White House.2023.“Executive Order on the Safe,Secure,and Trustworthy Development and Use of Artificial Intelligence.”Oct.30,2023.https:/www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/.