Power system state forecasting using machine learning techniques

Modern power sector requires grid observability under all scenarios for its ideal functioning. This enforces the operator to incorporate state estimation solutions based on a priori measurements to deduce the corresponding operating states of the grid. The key principle for such aforementioned algorithms lies on an occurrence of an over determined class of system having an ample redundancy in the measurements. Operators employ state forecasting solutions to counter the loss of real-time measurements. This work encompasses a critical comparison between several machine learning models along with ARIMA and time delayed neural network architecture for proper forecasting of operating states under normal as well as contingency scenarios. To showcase the efficacy of the proposed approach, this work incorporates a comprehensive comparison between them based on RMSE, MSE and MAE index. Copula-based synthetic data generation based on Gaussian multivariate distribution of the a priori measurements and operating states along with optimal hyper-parameter tuning of the models have shown the effectiveness of such algorithms in predicting future state estimates. The proposed machine learning models can be also seen to showcase an effective forecasting strategy under varying noise scenarios. This work also showcases the implementation of the models for real-time state forecasting strategy having computational times in the order of micro seconds. All the simulations have been carried out on the standard IEEE 14 bus test bench to support the former proposals.