Deep Reinforcement Learning-Assisted Robust Cubature Kalman Filter for Power System Dynamic State Estimation With Multi-Rate Measurements

Abstract

The coexistence of high-frequency phasor measurement units (PMUs) and conventional SCADA systems raises the challenge of heterogenous-source and multi-rate measurements which significantly degrades the performance of power system dynamic state estimation. In this work, a deep reinforcement learning (DRL) assisted robust cubature Kalman filtering (CKF) scheme is proposed to handle measurements from hybrid sources and with different time scales. In specific, a multi-rate measurement function reconstruction approach is designed with an independent discretization mechanism to lift the present limitation of requiring an integer multiple relationship of the sampling rates from multiple sources in most of existing works. Embedded with this discretization mechanism, a deep reinforcement learning assisted two-parameter linear exponential smoothing method is proposed to reconstruct the slow measurement model with online adjustable estimation parameters. A generalized correntropy loss criterion is also included in the robust CKF to counter the non-Gaussian noise and the noise distribution variation caused by the reconstruction. Comparisons results demonstrate that the proposed DRL-based robust CKF method can achieve better accuracy and robustness under various operating scenarios. Note to Practitioners—This work aims to improve the performance of the power system state estimation function in power system control center. Most of the power system state estimation approaches handle measurements with same sampling rates and/or with a multiple relationships of sampling rates. However, in power system practice, PMU and SCADA units exist simultaneously and their sampling rates may be not with a multiple relationship. This work proposes a deep reinforcement learning assisted robust cubature Kalman filter embedded with a multi-rate sampling integration mechanism. It will be able to achieve better estimation accuracy under time-varying operation conditions in power systems such as sudden shifts of load and renewable generations. The method can be integrated into the existing GUI in power system control center where the state estimation also exists. The deep reinforcement learning may increase the need of computing capability of control center which has been upgrading by most of power utilities.