Algorithm for predicting cost data of substation engineering based on fused XGBoost

ZHOU Bo; LIU Yun; LI Weijia; QI Yanxun; WANG Ligong

doi:10.7688/j.issn.1000-1646.2025.03.07

Abstract：[Objective]Traditional cost prediction methods for power grid substation engineering often rely on single influencing factors or linear assumption models,which fail to comprehensively capture the complex non-linear relationships among multiple factors,resulting in low prediction accuracy.Furthermore,existing methods face challenges such as dimensionality explosion or information loss when handling high-dimensional categorical variables,and especially,overfitting is prone to occur in small-sample datasets.Therefore,this study aims to develop a robust cost prediction model for substation engineering that effectively integrates multi-source influencing factors,adapts to non-linear relationships,and performs well in small-sample scenarios,thereby providing more accurate technical support for investment decisions in power grid enterprises.[Methods]To address these issues,a substation engineering cost prediction model(ME-XGB)based on the fusion of mean encoding(ME)and the extreme gradient boosting(XGBoost)framework was proposed.First,13 key influencing factors were extracted from dimensions such as equipmentand materials,construction techniques,construction scale,geographical environment,and design standards,covering both categorical and continuous variables.For categorical variables exhibiting non-linear relationships with cost,ME was applied for feature engineering.This method converted categorical variables into continuous features by calculating the mean of the target variable(cost per unit capacity)within each category and combining with a smoothing factor to retain category information while avoiding dimensionality explosion.Second,the XGBoost algorithm was utilized to construct the prediction model.The generalization ability of the model was enhanced by integrating multiple decision trees to iteratively correct residuals and incorporating regularization terms and hyperparameter tuning.Experiments were conducted using 200 substation engineering samples from a power grid company,which were randomly divided into a training set(80％)and a test set(20％).The performance of ME-XGB was compared with MK-TESM based on a Mann-Kendall(MK)trend test method and a three exponential smoothing method(TESM),backpropagation(BP)neural network,and the original XGBoost model by using mean absolute error(MAE)and goodness of fit(R2)as evaluation metrics.[Results]Experimental results demonstrate that the ME-XGB model significantly outperforms comparative models in prediction accuracy on the test set.Specifically,the median and mean MAE values of ME-XGB are 5 and 6.875,where are lower than those of MK-TESM,BP neural network,and the original XGBoost.Additionally,the R2 value of ME-XGB reaches 0.857 9,significantly higher than those of the other models,indicating stronger explanatory power for data variations.Boxplot analysis further reveals that ME-XGB has the narrowest distribution range of prediction errors,confirming its greater stability.Hyperparameter tuning results show that settings of hyperparameters such as tree depth and learning rate effectively balance model complexity and overfitting risks.[Conclusion]The proposed ME-XGB model addresses the challenges of non-linear representation and dimensionality control for categorical variables through ME,while leveraging the ensemble learning capability of XGBoost to significantly enhance prediction performance in small-sample scenarios.ME-XGB outperforms traditional models in terms of MAE,R2,and error stability,providing a more reliable cost prediction tool for power grid enterprises.Future research can further explore the modeling of dynamic influencing factors and extend the application of the model to cross-regional projects through transfer learning.