An adaptive conditional triggered dual-modal graph network for photovoltaic power prediction based on feature enhancement and structure-aware clustering

Accurate photovoltaic (PV) power forecasting is essential for real-time scheduling and risk management in renewable energy systems, yet it remains challenging due to complex spatio-temporal dependencies and sudden fluctuations. To address these challenges, we propose an adaptive conditional triggered dual-modal graph network (ACT-DMGN). It enhances features by embedding past observations into node attributes through a lag matrix. A hypergraph Laplacian is built from these expanded features, and spectral hypergraph clustering yields multi-scale embeddings that capture coherent relationships. ACT-DMGN employs a sliding-window error-trigger mechanism to alternate between a stable static graph—reflecting enduring patterns—and an adaptive dynamic graph—responsive to sudden changes. An adaptive update strategy focuses graph refinements on intervals with high variance, reducing computational overhead. The fused graph representations feed into a temporal convolutional decoder that produces point forecasts. Evaluated on two real-world PV datasets, ACT-DMGN achieves mean absolute errors of 1.9590 MW and 1.7326 MW, representing reductions of 22.78% and 26.33% over benchmark models. Its prediction performance is also significantly better than that of other comparison models on both datasets. These results indicate that lag-based feature enhancement combined with spectral clustering and conditional graph adaptation yields robust and efficient PV forecasting under volatile conditions.