Reproduce “Temporal Graph Neural Networks for Early Anomaly Detection and Performance Prediction via PV System Monitoring Data” on OpenResearch

About this paper

Srijani Mukherjee, Laurent Vuillon, Liliane Bou Nassif, Stéphanie Giroux-Julien, Hervé Pabiou, Denys Dutykh, Ionnasis Tsanakas

A Temporal Graph Neural Network (T-GNN) is introduced to accurately predict solar photovoltaic (PV) system power output and detect anomalies by concurrently modeling spatiotemporal dependencies. The T-GNN achieved a Mean Absolute Error (MAE) of 0.0707 for normalized power output prediction and flagged 5.21% of the dataset as anomalous, outperforming several traditional and deep learning baselines.

Key insights

• Integrating both spatial and temporal dependencies via a T-GNN architecture offers superior predictive accuracy and robust anomaly detection capabilities for PV systems compared to methods focusing solely on temporal or isolated parameter analysis.
• Representing PV system parameters as a graph effectively models the complex interdependencies and causal influences within the system, leading to a more comprehensive understanding of its behavior.
• The model's ability to learn intricate spatiotemporal patterns allows it to capture rapid changes in power output and identify subtle inconsistencies, providing actionable insights for PV system maintenance and optimization.

Summary by alphaXiv.