Real-Time Estimation of Smart Grids: A Neural Network-Based Unscented Kalman Filter Approach

Dynamic state estimations in the control centres hold an important role in formulating the current operating scenario of the grid while the states may fluctuate rapidly due to varying load scenarios. Unscented Kalman filters (UKFs) are a class of such aforesaid estimators that are based on the state prediction model, thus depending vastly on the convergence and the accuracy of the estimated states for the previous time step. To effectively develop an accurate state prediction model under varying load scenarios along with improving the estimation accuracy of the UKF, this work undertakes recurrent neural network structures incorporating convolution neural networks (CNNs) and gated recurrent units (GRUs). The CNNs are effectively deployed for feature extraction from time series data of state variables while the GRUs are implemented for an enhanced prediction of the estimated states in real-time. With efficient model training followed by error minimization for the UKF, the estimation accuracy of such dynamic state estimators has been enhanced. The simulation results when carried out over the IEEE 14 bus system promote the efficacy of the developed approach against traditional state estimation techniques under varying load scenarios. The developed model also showcases a real-time power system state estimation under varying noise margins as well.