Hybrid Data-Driven and Physical Model for Distribution System State Estimation

Realizing accurate Distribution System State Estimation (DSSE) faces challenges including a lack of observability and high complexity of the distribution system. A data-driven approach based on machine learning models may be an option to address this issue, but due to the lack of data labels in practical applications, their application in DSSE is limited. In fact, measurements in distribution systems often contain noise, are damaged or missing, and are unusable. In order to better address these issues, a distribution system state estimation solver is proposed, which is a deep learning model based on Residual Network (ResNet) that study the relationship between measurement values and state variables. A learning method combining traditional machine learning and power flow guidance was proposed to train ResNet, which respects the physical laws of the distribution system by inputting the output of ResNet into the power flow equation. This strategy can learn from noisy measurements, reducing the need for ideal data labels. Experiments on IEEE 33 bus distribution networks have shown that the model proposed in this paper outperforms traditional weighted least squares algorithms in accuracy.