基于异构图学习的新能源出力预测模型推理方法

Inference Method for New Energy Power Prediction Models Based on Heterogeneous Graph Learning

双碳目标背景下,具有区域特征和性能特点的新能源出力预测模型层出不穷,如何甄别利用当前存在的海量预测模型是当前实际应用中预测人员关注的问题。因此,提出了基于异构图学习的新能源出力预测模型推理方法,将新能源出力预测分为基础模型层和模型推理层。在基础模型层中,基于不同区域特征数据集对具有不同特点的预测模型进行训练。而在模型推理层,设计了“输入信息节点-输入边-模型节点-潜在联系边-预测结果”的异构图表示方法以完成异构信息的融合,并通过异构图注意力网络实现最优模型推理,从而获得精准的新能源预测结果。以中国西南某省风电场为算例对所提方法进行验证,结果表明所提方法风电预测误差低于9%,且直接迁移于无数据新建场站上的预测表现优于其他方法。

With the background of dual-carbon targets,countless new energy power prediction models with regional characteristics and performance characteristics have emerged.Distinguishing and utilizing the numerous existing prediction models are challenges for prediction personnel in practical applications.Therefore,this paper proposes a new energy power prediction model inference method based on heterogeneous graph learning,which divides the new energy power prediction into a basic model layer and model inference layer.In the basic model layer,prediction models with different characteristics are trained using datasets with different regional characteristics.In the model inference layer,a heterogeneous graph representation method of“input information node–input edge–model node–potential connection edge–prediction result”is used to fuse the heterogeneous information,and the optimal model inference is achieved through a heterogeneous graph attention network.Thus,accurate new energy prediction results are obtained.For a case study involving a wind farm in southwest China,the wind power prediction error of the proposed method was<9%,and the prediction performance for direct migration to new sites without data was superior to that of other methods.