Research on Missing Data Estimation Method for UPFC Submodules Based on Bayesian Multiple Imputation and Support Vector Machines

With the increasing complexity of power systems, the monitoring data of UPFC submodules suffers from high missing rates due to sensor failures and environmental interference, significantly limiting equipment condition assessment and fault warning capabilities. To overcome the computational complexity, poor real-time performance, and limited generalization of existing methods like GRU-GAN and SOM-LSTM, this study proposes a hybrid framework combining Bayesian multiple imputation with a Support Vector Machine (SVM) for data repair. The framework first employs an adaptive Kalman filter to denoise raw data and remove outliers, followed by Bayesian multiple imputation that constructs posterior distributions using normal linear correlations between historical and operational data, generating optimized imputed values through arithmetic averaging. A kernel-based SVM with RBF and soft margin optimization is then applied for nonlinear calibration to enhance robustness and consistency in high-dimensional scenarios. Experimental validation focusing on capacitor voltage, current, and temperature parameters of UPFC submodules under a 50% missing data scenario demonstrates that the proposed method achieves an 18.7% average error reduction and approximately 30% computational efficiency improvement compared to single imputation and traditional multiple imputation approaches, significantly outperforming neural network models. This study confirms the effectiveness of integrating Bayesian statistics with machine learning for power data restoration, providing a high-precision and low-complexity solution for equipment condition monitoring in complex operational environments. Future research will explore dynamic weight optimization and extend the framework to multi-source heterogeneous data applications.