GridFM - 电网基础模型.pdf

1、PublicTowards Foundation Models for the Electric GridPRESENTED BY:FRANOIS MIRALLSRESEARCHERHYDRO-QUBEC RESEARCH CENTERCANADAOCTOBER 3RD2025 AT THE HYDRO-QUEBEC RESEARCH INSTITUTEPublicOpen-source foundation models for the electric gridSource:H.Hamann et al.Foundation models for the electric power gr

2、id,Joule,Cell Press,2024 https:/doi.org/10.1016/j.joule.2024.11.002Source:Optimal Power Grid Operations with Foundation Models,Alban Puech,Jonas Weiss,Thomas Brunschwiler,Hendrik F.Hamann 2024 https:/arxiv.org/abs/2409.021483H y d r o-Q u b e cPublicMany different time scales,phenomena and dataFract

3、ion of wave(EMT)Steady stateMinutes to hours(Quasi-static)A few seconds to minutes(electro-mechanical transients)Time scaleValue of AI accelerationDifficultyElectromagnetic steady state butElectromechanic and control dynamics+operating rulesElectromagnetic dynamicsElectromechanic and electromagnetic

4、 steady statePhenomenaData and pretraining taskLonger timeShorter timeSmall frequency variations hypothesis(but higher in case of HQ relative to other networks,up to 1Hz)Fixed frequency hypothesisLarge frequency variation hypothesisAI ModelsPretraining on node and edge imputation on a set of snapsho

5、tsPretraining on node and edge imputation on a time series of snapshotsPretraining on wave imputationStep:10sStep:4,16ms(1/4 cycle 60 Hz)4H y d r o-Q u b e cPublicModule 5:gridfm-federated Module 2:gridfm-datakitModule 3:gridfm-benchModule 6:gridfm-lifelong-learning.Module 4:gridfm-cybersecurityModu

6、le 1:gridfm-graphkitApache 2.0One foundation for allAll for one foundationOpen-source,open-weight,GridFM trained on open-data.can be tested more rapidly for a variety of applications including cybersecurity.GridFM finetuned behind utility firewall for further testing.GridFM finetuned behind associat