As new technologies such as renewable energy resources, distributed generation, and plug-in electric vehicles penetrating the power distribution systems, intelligent monitoring and control become increasingly important for reliable and efficient operation of smart grids, and continuous delivery of high quality electricity to the customers. To accommodate these new technologies, supporting hardware including advanced two-way communication and remotely controllable devices had a significant development over the last decade. However, how to manage, coordinate, and supervise the distribution systems remains a great challenge for the electric utilities.
To address these challenges, we developed four use cases and applications for smart grid monitoring and control from both model-based and data-driven perspectives. Namely, distribution system state estimation (DSSE), distribution system anomaly detection, distribution network reconfiguration (DNR), and Volt-VAR control (VVC). For data-driven algorithms, we derive algorithms that are sample efficient, interpretable, and theoretically justifiable.
Specifically, for distribution system monitoring, we address the low observability and numerical instability issues with the unbalanced DSSE problem with a constrained maximum likelihood (CML) estimator and a sparse subspace Gauss-Newton algorithm. The uncertainty estimate is also derived within the CML framework. To address the physical interpretability of anomaly detection algorithms, we established the connection between the linear approximation of the power flow manifold and a class of modified linear regression models. Algorithmically, we estimate the model parameters and detect anomalies using smart meter data only, without detailed network parameters or confirmed anomaly cases.
For distribution system control, we propose model-based decentralized and data-driven centralized approaches to the DNR problem. For decentralized algorithms, we improve the convergence speed of alternative direction method of multipliers (ADMM) by an approximated Newton's update. For data-driven centralized algorithms, we improve the sample efficiency of existing reinforcement learning (RL) algorithms by a heuristic data augmentation approach, and a principled framework termed batch-constrained soft policy iteration theory. Both approaches improve the existing control policy in the historical datasets, and the batch-constrained RL scales up to networks with hundreds of nodes and switching devices. Lastly, a randomized communication-efficient consensus multi-agent RL (C-MARL) based VVC framework is developed.
