State Estimation of Electrical Grids via Neural Networks

Modern power systems are facing challenges due to the increasing variability and stochasticity of system states caused by the widespread use of renewable energy sources. Monitoring the grid's conditions in real-time has become crucial, and the existing power system state estimation methods are either too expensive in terms of computational time or require accurate information about the grid's topology and parameters. To overcome these difficulties, we use deep neural networks for real-time power system state estimation. An efficient Feed Forward Neural Network (FNN) is proposed for this purpose. We first generate a synthetic time series for the system load and PV generation, which is generated using real data. Further, we utilize multiple time indices as input features to the FNN (i.e., the estimation is based on measurements plus the time of estimation). Three FNN models are trained which differ in terms of the number of input measurements (i.e., available sensors). The models are then tested under three scenarios: 1) mild perturbation of measurements, 2) medium perturbation plus dropping of measurements, and 3) severe perturbation plus dropping of measurements. The IEEE 14 bus system is utilized to demonstrate the results.