W2D2S2: Machine Learning for Adaptive RAS/SPS System Settings/Constrained Physics
Machine Learning for Adaptive RAS/SPS System Settings in Power System Operation and Control Xiayuan Fan (Senior Research Engineer, PNNL)
Given the increasing penetration of renewable energy, demand response and smart controllers in today’s power grid, atypical power flow patterns such as reverse flows, loop flows, and stochastic dynamic behavior are being observed in real-time operation. These new patterns may invalidate the existing protection relay settings, especially Remedial Action Scheme (RAS), also known as Special Protection Scheme (SPS), which when invalidated can potentially cause cascading failures if the operational issues caused by the new challenges are not fully understood and addressed.
Traditionally, RAS settings have been determined using offline study, which is very time-consuming, due to a lack of automation and computational power. No automated tools exist to assist planning and protection engineers in adaptively determining RAS settings to enable better response to unknown grid conditions. Thus, these settings are typically overly conservative, causing unnecessary flow curtailment or generation tripping that can affect revenue of generator owners and economic operation of the entire power network.
Several challenges are identified in the industry preventing the RAS settings from being determined in an adaptive/online manner. One key issue is that the computational speed is not fast enough in today’s commercial tools to perform a full-scale study to calculate RAS parameters such as the arming level and validate the control performance in a preventive way.
The Pacific Northwest National Laboratory (PNNL) project team has developed innovative mathematical and advanced computing methods for adaptively setting Remedial Action Scheme/Special Protection Scheme (RAS/SPS) coefficients with the consideration of realistic and near real-time operation conditions. In this DOE-funded project, the Jim Bridger RAS in U.S. Western Interconnection served as the use case for testing and validating the proposed methodology and the corresponding prototype, Transformative Remedial Action Scheme Tool (TRAST).
TRAST uses a novel approach to generate use cases automatically, bringing in advanced statistical data analysis tools, and using machine learning algorithms to analyze, validate, and help create RAS plans. The parallel computing platform at PNNL, as well as Microsoft cloud environment, are utilized for steady state and dynamic simulations under massive contingencies and operating conditions to calculate more realistic settings of RAS systems in a real-world environment. TRAST could significantly simplify and shorten the RAS design and study process. Additionally, continuous improvement and validation can be achieved using the proposed evaluation methodology.
Both grid operators and utility planning engineers will benefit from this technology, and a better RAS modeling process will also increase interconnection-level situational awareness.
Constrained Physics-Informed Deep Learning for Stable System Identification and Continuous Control, Jan Drgona (Data Scientist, PNNL)
We present a novel data-driven method for learning deep constrained continuous control policies and dynamical models of the controlled system. By leveraging partial knowledge of system dynamics and constraint enforcing multi-objective loss functions, the method can learn from small and static datasets, handle time-varying state and input constraints and enforce the stability properties of the controlled system.
We use a continuous control design example to demonstrate the performance of the method on three distinct tasks: system identification, control policy learning, and simultaneous system identification and policy learning. We assess the system identification performance by comparing open-loop simulations of the true system and the learned models. We demonstrate the performance of the policy learning methodology on closed-loop control performance using the ground truth system model under varying levels of parametric and additive uncertainties affecting its dynamics.
We then evaluate the potential of simultaneously learning the system model and control policy. Our empirical results demonstrate the effectiveness of our unifying framework for constrained continuous optimal control to provide stability guarantees, explainable models, robustness to uncertainty, and remarkable sampling efficiency.