Research On Distribution System State Estimation Method Based On Deep Learnin

As China’s social economy develops rapidly,so does the construction of smart grid.The distribution network,which directly faces the end users and is closely related to the production and life of the people,is a key component of the power system.The operation and control of the smart distribution network are more complex than those of the traditional distribution network,and the amount of data also increases sharply.Bad data caused by transmission errors and other reasons may result in incomplete measurement data,which can lead to erroneous judgment of the distribution network operation situation and affect the safe and stable operation of the power grid if used directly for distribution network state estimation.To address this issue,this paper proposes a data-driven bad data cleaning method and a distribution network state estimation method.The main contributions are as follows:1.Bad data cleaning method.Proposed a two-stage data cleansing method that combines Stacking and Isolation Forest.In the training phase,a bad data detection model based on Stacking ensemble learning and a power load prediction model are trained using historical data.Appropriate machine learning models are selected as base learners and meta learners for Stacking ensemble learning.In online application phase,the current time measurement sample is first input to the bad data detection model to judge whether it contains bad data.If yes,it is further input to the Isolation Forest anomaly detection model for bad data localization,and the bad data are replaced with pseudo-measurement data generated by load prediction and power flow calculation to complete bad data cleaning.Simulation experiments show that this method has high accuracy of bad data identification and good cleaning effect.2.Distribution network state estimation method.A spatio-temporal graph convolutional network based on attention mechanism for distribution network state estimation is proposed,which considers the temporal-spatial correlation between nodes in distribution network.This method takes the measurement data at the current time point and the previous three time points as input,adaptively extracts and fuses features from temporal-spatial dimensions in measurement data,and obtains the final state estimation result.Experiments show that the method of combining bad data cleaning with state estimation can effectively deal with incomplete measurement data,thus improving estimation accuracy and robustness.