LLM-based renewable energy generation forecasting via multi-scale hyper-connection feature fusion

Most deep learning-based renewable energy forecasting methods are primarily designed for single-source data, failing to fully exploit the inherent correlations between different types of energy sources. Recently, Large Language Models (LLMs) have shown strong potential in representation transfer and cross-domain modeling, making them ideal for constructing generalizable forecasting systems. Nevertheless, renewable energy outputs often exhibit complex temporal variations, characterized by slowly evolving trends and abrupt fluctuations. To tackle these challenges, we propose MHFF-LLM, a LLM-based renewable energy generation forecasting via multi-scale hyper-connection feature fusion. Our model integrates hierarchical convolutional encoders to extract temporal patterns and employs a learnable hyper-connection strategy to effectively capture both fine-grained and coarse-grained temporal relationships. Additionally, a cross-modal attention mechanism is introduced to align meteorological textual cues with multi-scale sequential features, enhancing semantic–temporal coherence for improved predictive accuracy. Extensive experiments on several real-world datasets demonstrate that MHFF-LLM consistently surpasses leading baselines, achieving notable improvements in forecasting performance and robustness across heterogeneous renewable energy scenarios.