DRIG-Net: A Dynamic Renewable Interaction Graph Framework for Latent Operating State Analysis

This paper presents a framework to construct a Dynamic Renewable Interaction Graph (DRIG) to characterize renewable operating dynamics. We used autoencoder-based representation learning to extract latent embeddings from the data, and we assessed the learned latent representations using reconstruction loss, variance analysis, and downstream performance on an Extreme Gradient Boosting (XGBoost) model. The Principal Component Analysis(PCA) of the latent representations reveals the structure of the renewable operating state (ROS). K-means clustering identifies the operating regimes among the ROS. Subsequently, utilizing the latent embedding and ROS, the DRIG was constructed which models the temporal dynamics within Markov modeling technique to investigate regime persistence and seasonal dynamics analysis. Finally, an explainable AI technique was applied to XGBoost for SHAP analysis, which identifies the dominant factors driving changes in the latent ROS. Experimental results demonstrated that the proposed autoencoder-based extracted latent embeddings learns most informative representations with low reconstruction error. The proposed framework offers a new perspective on renew- able–demand interactions and interprets meaningful seasonal variation in renewable availability and renewable operating-state dynamics.