Research On Power Branch Parameter Identification Based On Graph Neural Networ

The branch parameter is an important variable basis of power system,which involves key issues such as power flow calculation,state estimation,fault diagnosis and short circuit calculation.Effective identification of transmission line parameters is of great significance for the safe and stable operation of modern power systems.However,the existing transmission line parameter identification methods have many shortcomings,such as over-reliance on measurement equipment,line repair scenarios,and the identification effect is not obvious in extreme weather.There are problems such as poor stability,easy divergence,and vulnerability to residual pollution.The main reason is that the existing measurement methods cannot meet the deviation caused by the on-site operating environment of the power system : ignoring the actual operating state of the power system and the contextual semantic information contained in the historical data of the power grid;it is neglected that the actual operating power grid is a global system,and the operating state of a single branch is closely coupled with the operating state of other associated branches.Based on the above requirements and scenarios,this paper proposes a new intelligent identification model of power grid branch parameters based on graph neural network considering the topology constraints of power grid transmission lines.The graph neural network learns the hidden information between branches through the preset branch topology,and uses the aggregation rules of spectral domain convolution to generate the hidden layer features of branches.Combined with different practical scenarios,a model optimization method based on the theory of self-weighing loss function is proposed.By transforming the original loss function of averaging each branch,an anti-noise factor is introduced for each branch that can learn the characteristics of the branch,and the peak clipping of outliers is realized.A time + space two-scale graph neural network time series anomaly location method based on GRU is proposed.The GRU module is aimed at the practical requirements of the global branch parameter identification of the power grid-not the time series identification of a single branch,but the time series identification of the global topology.Finally,the proposed algorithm is compared with the traditional regression algorithm and the deep learning algorithm by using the preprocessed power grid measured data.The effectiveness of the proposed method is proved by visualizing the model training results and the intermediate feature layer.