Physics-Informed Deep Neural Network for Partially Observable Distribution Grid State Estimation
2:55 pm, Room St Clair 3B
Power distribution grids, which deliver electrical energy from the substation to the end-users, are transforming rapidly with the penetration of highly fluctuating renewable energy sources. For the grid operators, knowing the state of the grid is becoming more challenging, and is often addressed by the incorporation of Internet-of-Things-based meters, capable of providing synchronous measurements across the grid. However, these meters are subject to technical reliability issues and cyber attacks, which altogether can render the grid partially-observable. Motivated by this problem, we developed a deep neural network (DNN) that accurately estimates the state of the grid in the event of a sudden observability loss. Our key contribution is in the explicit incorporation of the known physical information into the training routine of the proposed DNN. We evaluate the method using data collected from a real-world distribution grid and show that it (i) reduces state estimation error by an order of magnitude when compared to a DNN trained without the physical information, and (ii) reduces state estimation error by two orders of magnitude compared to a weighted least-squares solution with pseudo measurements.
Joint work with Jonatan Ostrometzky (Tel Aviv University), Konstantin Berestizshevsky (Tel Aviv University), and Andrey Bernstein (NREL)