Electrolysis is expected to play an essential role in future energy systems that rely on renewable energy sources,especially in achieving climate neutrality in sectors that are challenging to electrify directly.The ec...Electrolysis is expected to play an essential role in future energy systems that rely on renewable energy sources,especially in achieving climate neutrality in sectors that are challenging to electrify directly.The economic success of this technology is largely dependent on effective predictive maintenance,which requires a clear understanding of the systems’current and future state-of-health to ensure adequate replacement planning with decreasing efficiency and prevent undesired aging-related failures.Given the incomplete physical understanding and mathematical description of degradation processes,while more and more data is becoming available,data-driven machine learning models are increasingly moving into focus.These models can learn underlying relationships from data without necessitating prior knowledge.Therefore,this study concentrates on predicting the degradation trend of a proton exchange membrane water electrolysis cell using a data-driven machine learning approach.To this end,a comprehensive data-driven modeling matrix is proposed and evaluated through selected practically relevant modeling concepts,which are characterized by different combinations of available training data and desired model outputs.Experimentally,this is facilitated by a targeted accelerated stress test consisting of operating and characterization phases.The applied machine learning pipeline,covering the hierarchical sequence of necessary data preprocessing and modeling steps,is presented in detail to ensure the traceability and reproducibility of the methodology from data collection to model testing and evaluation.As a major finding,it was demonstrated that the degradation trend prediction for the entire polarization curve can be realized using only typical operating data.This represents an initial step toward predicting the complete cell characteristic without interrupting ongoing operation for corresponding measurements.展开更多
基金financial support by the Federal Ministry of Education and Research of Germany in the frame-work of SEGIWA(grant no.03HY121G)HyThroughGen(grant no.3HY108C)as part of the hydrogen flagship project H2Giga.
文摘Electrolysis is expected to play an essential role in future energy systems that rely on renewable energy sources,especially in achieving climate neutrality in sectors that are challenging to electrify directly.The economic success of this technology is largely dependent on effective predictive maintenance,which requires a clear understanding of the systems’current and future state-of-health to ensure adequate replacement planning with decreasing efficiency and prevent undesired aging-related failures.Given the incomplete physical understanding and mathematical description of degradation processes,while more and more data is becoming available,data-driven machine learning models are increasingly moving into focus.These models can learn underlying relationships from data without necessitating prior knowledge.Therefore,this study concentrates on predicting the degradation trend of a proton exchange membrane water electrolysis cell using a data-driven machine learning approach.To this end,a comprehensive data-driven modeling matrix is proposed and evaluated through selected practically relevant modeling concepts,which are characterized by different combinations of available training data and desired model outputs.Experimentally,this is facilitated by a targeted accelerated stress test consisting of operating and characterization phases.The applied machine learning pipeline,covering the hierarchical sequence of necessary data preprocessing and modeling steps,is presented in detail to ensure the traceability and reproducibility of the methodology from data collection to model testing and evaluation.As a major finding,it was demonstrated that the degradation trend prediction for the entire polarization curve can be realized using only typical operating data.This represents an initial step toward predicting the complete cell characteristic without interrupting ongoing operation for corresponding measurements.
