Large Scale, Data Driven, Digital Twin Models: Outlier Detection and Imputation

Precise and comprehensive records of system performance is imperative for the efficient monitoring and prediction of Photovoltaic (PV) power generation. Nonetheless, the inevitable failure of real-world sensors and monitoring devices results in information loss. Additionally, the high variability in...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Conference record of the IEEE Photovoltaic Specialists Conference S. 0902 - 0905
Hauptverfasser: Wieser, Raymond, Fan, Yangxin, Yu, Xuanji, Braid, Jennifer, Shaton, Avishai, Hoffman, Adam, Spurgeon, Ben, Gibbons, Daniel, Bruckman, Laura S., Wu, Yinghui, French, Roger H.
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 09.06.2024
Schlagworte:
Anomaly detection
Data models
data-driven Digital Twin
Deep Learning
Digital twins
Graph Neural Network
Imputation
Inverters
Maintenance
Monitoring
Outlier Detection
Photovoltaic systems
Production
Statistical Analysis
System performance
Timeseries
Timeseries Analysis
ISSN:2995-1755
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Precise and comprehensive records of system performance is imperative for the efficient monitoring and prediction of Photovoltaic (PV) power generation. Nonetheless, the inevitable failure of real-world sensors and monitoring devices results in information loss. Additionally, the high variability in production data can obfuscate erroneous measurements as nominal values. The occurrence of such "missingness" significantly impacts the stability and precision of performance estimation. By leveraging the inherent value dependencies present in PV production data, graph data driven Digital Twin models can be created for individual sites that capture the high frequency spatial weather patterns that determine the precise performance for that specific instance in time. Through varying the graph structure, the same model architecture can target both outlier detection, and imputation of missing values. st-GAE, an existing spatio-temporal graph autoencoder, was used to detect and impute outliers for a collection of 98 inverters with 5 minute interval data for a period of two years. Outlier detection was compared against existing maintenance logs which were available for all of the systems. stGAE was shown to correctly identify 90% of maintenance events and was able to reconstruct those missing values with a MAE of less than 1.5W.
ISSN:2995-1755
DOI:10.1109/PVSC57443.2024.10748985