关键词: 深度学习, 图神经网络, 风电功率, 风电场, 数值模拟

Abstract: Aiming at the problems of low accuracy and poor interpretability of the existing power prediction methods of non-uniform wind farm layout, we propose an improved physics-induced graph neural network （IPGNN） model to predict the power output of non-uniform wind farms. Firstly, IPGNN constructs a physical information basis function based on three-dimensional Gaussian wake model, which can more accurately reflect the wake interaction between wind turbines in non-uniform wind farms; Secondly, IPGNN designs a set of graph neural network update strategies based on the message passing framework. In the graph neural network edge update strategy, combined with the attention mechanism, the physical information basis function is used as the weight update function, which enhances the interpretability of the model. The simulation experiments of non-uniform wind farms with different numbers of wind turbines show that IPGNN model has a good prediction effect compared with PGNN model based on one-dimensional wake model and typical data-driven model. For non-uniform wind farms with 20 wind turbines, the MAPE of power generation prediction is 3.92%. It is an effective method for power generation prediction of non-uniform wind farm layout.

Key words: deep learning, graph neural network, wind power, wind farm layout, numerical simulation