Representative Photovoltaic Scenario Extraction Using Graph Clustering Models

Abstract

This chapter addresses the challenge of extracting representative scenarios from multiple photovoltaic (PV) stations, where strong uncertainty and spatiotemporal correlations complicate power system optimization. A novel adaptive weighted two-layer graph clustering model is presented. At the local level, an adaptive nearest-neighbor graph with sparsity-promoting feature weights is learned for each station to emphasize critical daily periods and suppress irrelevant temporal variations. At the global level, these local graphs are fused into a consensus graph via learnable fusion weights, and a Laplacian rank constraint is imposed to automatically determine the optimal number of clusters without manual tuning. An alternating direction optimization algorithm is derived, with theoretically proven monotonic convergence guarantees. As a complementary pathway, this chapter presents an end-to-end deep adaptive graph neural network that jointly learns node embeddings and iteratively refines a learnable adjacency matrix through a clustering-oriented multi-objective loss function. The proposed methods are validated on three real-world datasets, namely the Australian PV dataset, ERA5 reanalysis data, and the Open Power System Data (OPSD), demonstrating superior cluster compactness, separation, and spatial correlation preservation over state-of-the-art benchmarks. Probabilistic power flow analysis on the IEEE 69-bus system further confirms the physical feasibility of the extracted scenarios under realistic network constraints.