Knowledge-Fusion Graph Transformer network for wind farm assessment with sparse data

This study presents an innovative Knowledge-Fusion Graph Transformer (KFGT) network, which seamlessly integrates a physics-guided attention mechanism with a well-established analytical wake model, offering a groundbreaking approach to wind farm assessment, including power generation forecasting and critical component load prediction across a range of wind turbine layouts and operating conditions. By harnessing domain-specific knowledge within a graph-based learning framework, KFGT not only enhances predictive accuracy by an impressive 28.3% but also achieves a remarkable 39.7% reduction in trainable model size compared to the baseline Graph Transformer network. To address the challenge of sparse training data, KFGT achieves state-of-the-art performance with an overall prediction error of 5.2% across 11 evaluation indices under conditions of data sparsity. Notably, key metrics such as power output and blade root load exhibit errors below 3%. Despite operating in a high-dimensional space exceeding 75 variables and utilizing only around 2250 training samples, KFGT maintains exceptional accuracy with significantly fewer parameters. Even with a 75% reduction in training data, it retains 89.3% of baseline performance. By circumventing computational fluid dynamics modeling complexities, KFGT offers a scalable, robust, and computationally efficient solution for wind farm optimization, establishing itself as a transformative tool for advancing wind energy applications.