Fast Grid State Estimation for Power Networks: An Ensemble Machine Learning Approach

For better managing modern power distribution networks with continuously and dynamically changing grid characteristic induced by e.g. adding new renewable generation or vehicle charging stations, state estimation solutions must be capable of dynamically adapting to such topology changes. However, these solutions use a hand-made static topology model as calculation basis which cannot be easily dynamically adapted to topology changes without editing the underlying network topology model. This process is time-consuming and computationally expensive especially for large grids. In the present paper, we address the problem of grid state estimation by proposing a new machine learning-based solution. It paves the road for generic, accurate and simple state estimation independent of the grid topology and with minimal number of grid parameters. To achieve this, four simulated datasets, namely Balanced Generation-Consumption, Unbalanced Generation, Unbalanced Consumption and Mixed datasets are created representing the most common cases existing in the grid. Unlike other research works in the field of grid state estimation, this work applies ensemble learning, namely boosting, bagging and stacking for predicting the grid state variables. The obtained results are very promising and show that our ensemble-based solution exhibits very good results in terms of time efficiency and accuracy.