A Forecasting-Aided State Estimator Based on Improved Bayesian Deep LSTM Neural Network

This paper proposes a forecasting-aided state estimator (FASE) based on improved Bayesian deep long-short term memory (BDLSTM) neural network, which has significantly higher robustness against non-Gaussian distributed measurement noise and multiple types of uncertainties than existing methods. The proposed method optimizes neural network parameters using the variational autoencoder (VAE) algorithm and constructs the Variational Autoencoder-Bayesian Deep LSTM Neural Network (VAE-BDLSTM) prediction model. This model proficiently mitigates non-Gaussian noise problems and addresses the model uncertainty, stochastic uncertainty, and parameter uncertainty to an extent, thus improving the accuracy and robustness of state estimation. The unscented Kalman filter (UKF) is adopted as the nonlinear filter that combines predictions and measurements to produce the state estimate. Using the IEEE 39-bus system to simulate, and the results prove the effectiveness and practical value of the proposed method in improving prediction accuracy and estimator robustness.