Prediction of Damaged Spatial Distribution of Towers in Different Distribution Network Areas Based on Gradient Boosting Regression Tree

In towers in different distribution network areas, only the preliminarily screened feature parameters can be used, resulting in low spatial prediction accuracy. Therefore, a prediction method for damaged spatial distribution of towers in different distribution network areas based on gradient boosting regression tree is proposed. Multi-source data such as geographic information, power grid operation parameters and environmental monitoring data are widely collected. For problems such as missing data and different dimensions, mode filling and linear interpolation methods are used for preprocessing, and a comprehensive impact index is constructed. Based on the gradient boosting regression tree, cosine similarity is calculated to preliminarily screen variable combinations with strong correlation, and a fusion feature matrix is constructed to further screen feature parameters. With iterative training as the core, the performance of weak learners is evaluated with the help of negative gradients, and a prediction model for the spatial distribution of damaged towers in distribution network areas is established. In the construction of the training sample set, the Manhattan distance between samples is defined, the nearest neighbors are found and weighted, the sample information volume is defined, and the threshold is set to screen samples to form a new sample set training model, so as to realize the prediction of damaged spatial distribution of towers in different distribution network areas. The experimental results show that the prediction accuracy of the design method is above $95 \%$ in all regions, and the consistency index increases from 0.67 to 0.80, proving that it has high spatial prediction accuracy.