Pre-Trained Large Model as Distributed PV Power Forecaster: Integrating Scenario-Consistent Retrieval-Augmented Generation and Cross-Modal Semantic Alignment

The rapid proliferation of distributed photovoltaic (PV) systems presents significant challenges for accurate power generation forecasting due to their inherent intermittency and spatial-temporal variability. This paper proposes a novel framework that leverages pre-trained large models for distributed PV power forecasting by integrating retrieval-augmented generation with cross-modal semantic alignment. To unlock the reasoning capabilities of pre-trained large models, we design a scenario-consistent RAG module that retrieves relevant historical PV sequences guided by long-term dependencies and temporal correlations, and integrates them through a Chain-of-Thought prompting paradigm. Furthermore, we develop a cross-modal semantic alignment mechanism to bridge the representational gap between numerical PV series data and the language semantic space of PLMs, enabling effective pattern recognition across modalities. To ensure physically meaningful forecasting within the training and inference processes, we also establish a constrained activation function that explicitly considers the inherent characteristics of PV power generation. Extensive experiments on nine real-world distributed PV datasets demonstrate that our framework achieves state-of-the-art performance, significantly outperforming six benchmark models.