Distributed Agent-Based Management of Agile Microgrids

US Department of Defense, Army Research Laboratory

This project plan (APP) describes the third year of the four year program for distributed agent-based management of agile microgrids. In year 1, the team has evaluated modeling and forecasting techniques for renewable energy sources as well as developed relevant case studies. In year 2 the further developed the models and forecasting techniques as well as begin implementation of simulations and hardware test cases.

The existing simulation models a user-definable, network ofmicrogrids and the Autonomous Agile Microgrid (AAM) control system. The AAM has three main components – (1) a low-level, asset control system (Decentralized Closed Loop Controller agent- DCLC), (2) a mid-level, optimal, grid state-change solver (Decentralized Model Based Control agent – DMBC) and the highest level reasoning layer, (Distributed Grid Management agent- DGM).
The entire system is “driven” by a user-configurable, time-history of prioritized loads and events based on field data.

The focus of the year three plan is to (1) increase the reasoning capability of the DGM, (2) develop an optimal power flow strategy at the DMBC level and (3) design a human-in-the-loop interface that permits real-time interaction with the simulation.

Deliverable 1. The AAM uses a command line approach to execute the simulation and observe the grid’s evolution based on a pre-defined time history scenario of events and loads. While the process for designing rich scenarios has a well-defined workflow, the system currently lacks the ability to respond to real-time inputs from a user. The deliverable is demonstration of a new human-in-the-loop capability for the AAM simulator. It will permit one user to “actuate” the power grid manually, or in an AAM-assist mode where the user can optionally decide to implement the AAM’s recommended actions. A second user will be able to trigger events and load changes in real-time, including policy and scenario changes in the DGM. The intent for year three is to create an environment for more complete testing of the AAM and demonstrate its capabilities. This feature would then be available for future studies to increase the reasoning functions of the DGM using human-in-the-loop training.

Deliverable 2. The DMBC currently computes an optimal solution to transition the power grid from its current state to a new state as requested by the DGM. These requests are based on load, generation and storage forecast agent calculations. The DMBC also triggers a new solution based on high-tempo changes to the bus voltage, independent of the DGM, due to unforecasted changes in loads or generation. All DMBC solutions are based on the assumption of a fully functioning, well-defined set of loads, generation, and storage assets. The DMBC does not compute optimal redirection of power flow based on catastrophic generation or load failures. The deliverable is the development and demonstration of a scheme to optimally redistribute power flow for contingency and catastrophic events including equipment faults and attack damage. This redistribution strategy control may be at the level of the DMBC, DCLC or both. The human-in-the-loop capability, described in Deliverable I, will be used to demonstrate this new feature by instantaneously removing generation assets and loads.

Deliverable 3. The DGM relies on data-driven load and generation forecasts to compute grid state change requests for the DMBC. The forecasts will be improved with the inclusion of additional knowledge of inventory, asset models, and situational information. While load prioritization is accommodated, there is not functionality for addressing situations where there is not enough power to accommodate all of the highest priority loads. Policies and negotiation protocols for the DGM multi-agent system that enable power sharing among microgrids will be explored. Additional policies for fine control of load shedding will be examined and simulated. The ERDC-CERL VFOB project can potentially provide a rich source of data and models to the DGM design that support more elaborate forecasts and reasoning under conflict. In addition, the data can drive methods for scenario classification (prediction of the current and future events of the base, e.g., patrols, heightened alerts, etc.). The deliverable is a report and demonstration of improved forecasting agents and conflict resolution handling, through the power scheduling agent, based on knowledge-based reasoning mechanisms and statistical risk analysis metrics. The report will document the agent models and reasoning strategies along with a description of opportunities and gaps for implementing a fully autonomous, resilient power grid.

Investigators: Gordon Parker, Laura Brown, Wayne Weaver, Steven Goldsmith