Thrust 1: Scale-up Decision Support

The Goal

To develop high-fidelity power grid models and data sharing mechanisms to fuel stochastic optimization and machine learning (ML) routines that are scalable to a large number of inverter-based distributed energy resource (DER) and variable renewable energy (VRE) sources.

The Theory

With the integration of new inverter-based VREs, storage, and DER resources and blurring the line between transmission and distribution grids, models underlying decision support for a 100% renewable grid power grid (T100RE) become more complex and computationally demanding. The first step will be to extend status quo computing tools to a multi-stage, stochastic optimization and control formulation for energy management across multiple time scales with a detailed representation of the distribution grids. Cutting-edge stochastic optimization and ML are capable of producing optimal or near-optimal solutions for large-scale, network-constrained scheduling and market-clearing problems. These methods also leverage massive amounts of historical data that can be complemented with data augmentation processes to capture the full range of operating conditions, including extreme weather events. Despite availability of stochastic optimization, ML, and data processing methods, their adoption has been hindered by a lack of trust in their explain-ability, preparedness and reproducibility among stakeholders, especially when the power grid is stressed.

The Methods

A fundamental challenge inhibiting efficient, reliable and resilient power management in a T100RE grid is the ability to derive both tractable and physically accurate representations of the increasingly complex power grid with an increased number of DERs and VREs. This requires enhancing the dimensionality of the proposed models in terms of (a) the number of connected resources (e.g., to integrate such DERs as small-scale photovoltaic panels and storage units), (b) spanning decision reach from transmission to distribution grids, and (c) and increasing their temporal resolution to account for shorter timescales (e.g., to account for dynamics of the T100RE power grid). Additional effort will be required to represent inverter based resources (IBRs) and their flexibility, a challenge that we will deal jointly with the team working on RD 2.1 below. Finally, in addition to temporal resolution, temporal coupling between different operational time scales (e.g., min-to-hours and hours-to-days) must be enhanced to accommodate and efficiently use capabilities of energy storage and demand response resources. In addition to modeling T100RE representations, we will develop data augmentation to simulate, jointly with Project Affiliates NREL and Berkeley, operating conditions with increasing penetrations of IBRs, VRE and electrified loads and employ weather-stressing to consider extremes.

Building trust in ML applications also requires such ML routines that respect power grid physics and provide guarantees on feasibility, optimality, and computational stability. Despite recent advances, there is a gap in fidelity (e.g., modeling alternating current power flow models) and scalability (e.g., the ability to solve stochastic management problems for larger network instances with countless DERs and multi-period couplings). Moreover, when filling these gaps, it is critical to augment these ML routines to convey an intuition behind each solution to grid operators (e.g., if a decision is made, it is important to understand what causes this decision and how this decision fares against plausible alternatives).

Our approach to resolve these challenges is to embed representation of the IBRs and power grid constraints into ML training procedures and to develop efficient “learning-to-optimize” methods. To scale to a large number of DERs, the embedded representations of DERs will leverage inference and aggregation. For tractability, ML training will be pushed from real-time to off-line, with only light optimization routines solved live. To enhance the robustness of these methods under uncertainty, we will work with Project Affiliates NREL and EPRI to leverage recent advances in end-to-end learning to integrate forecasting and decision-making (optimization) and to align their objectives. This RD will deliver novel decentralized stochastic optimization algorithms that capture a massive number of DERs, storage and demand-side management options, accelerate them using dual prediction methods, and learn them using decentralized optimization proxies as in. This decentralization will enable efficient coordination between transmission and distribution power grids, thus delivering the DER benefits across large geographical areas.

The ability to operationalize the T100RE models and ML routines described above assumes that data is available and shared among all participants. In practice, however, data sharing is obstructed by valid power grid integrity and security concerns, motivated by risks of malicious interventions (e.g., cyber and physical attacks, loss of competitive advantages (e.g., power producers, utilities, retailers) and/or invasion of privacy (e.g., households participating in demand response). Therefore, we will develop a data sharing framework that incentivizes stakeholders to allow their data to be used in power management, while allowing for control of their data risk. While such privacy-preserving analytics exists for simple problems (e.g., unconstrained optimization to estimate regression parameters), constrained, mixed-integer and ML problems still lack foundational privacy-by-design approaches. The T100RE goal is to develop privacy-preserving analytics that allows for data valuation within power management routines, thus taking into account downstream implications, and provides verifiable guarantees. We will enhance the derivation and analysis of data sharing incentives by blending ML and algorithmic game theory. Building on the incentive-based data acquisition approach, we will explore methods to assess data value from data to both operationalize and valorize digitalization of electric power grids.

The proposed data sharing with high-fidelity models in RD and trustworthy ML in RD will enhance the degree of coordination between power producers, grid operators, and consumers, and become the scientific backbone of the next-generation decision support enhancing efficiency, reliability and resiliency of power management. This will shape computational methods for future T100RE power grids, form the basis of standards for validating explainable and trustworthy forecasting methods and ML systems, and update the current decision-making pipeline (e.g., day-ahead and real-time power grid scheduling) in partnership with G-PST and FPFM. We will collaborate with the NSF AI Institute for Advances in Optimization (AI4OPT), led by EPICS-US co-PI Van Hentenryck, in the area of ML/optimization use for decentralized DER integration.

The Team

Enrique Mallada

Associate Professor of Electrical and Computer Engineering, Johns Hopkins University

Yury Dvorkin

Associate Professor of Civil and Systems Engineering and Electrical and Computer Engineering, Johns Hopkins University

Pascal Van Hentenryck

A. Russell Chandler III Chair, Georgia Tech

Taj Khandoker

Research Scientist of Energy & Economic Modeling, CSIRO

Luis (Nando) Ochoa

Professor of Smart Grids and Power Systems, University of Melbourne

Pierre Pinson

Chair in Data-centric Design Engineering, Imperial College London