Learning-based Line Impedance Estimation for Partially Observable Distribution Systems

Abstract: Accurate distribution line impedances are vital to distribution system operation and control, while the recorded parameters are likely to be biased from the true impedances. The lack of real-time measurement data in partially observable distribution systems (PODS) increases the difficulty of distribution line impedance estimation. To tackle this challenge, we propose a learning-based distribution line impedance estimation method via mining the dependence between line parameters and the accessible measurements of the whole system. A reinforcement learning (RL) model is established for the line impedance estimation with limited measurements, and an RL reward mechanism is designed to evaluate the estimation accuracy among various operating scenarios. To explore the various scenarios in the RL training, we propose a generative adversarial network-based method to generate numerous nodal power injection samples. Besides, we design a data denoising method that combines the empirical mode decomposition with the Kalman filter, which provides denoised state information for the RL agent. Numerical case studies on the IEEE test systems validate the effectiveness and superiority of the proposed RL-based method in distribution line impedance estimation only with limited SCADA measurements.