RFI.doc

Brookhaven National Laboratory Request for information on a project “A Deep Learning Based Online Platform for Critical Anomaly Detection and Emergency Control to Enhance Grid Reliability and Resiliency” BNL Technical POC: Meng Yue yuemeng@bnl.gov November 21, 2020 Accurate load modeling is a key to characterizing the dynamic behaviors of power systems. In the past few years, the WECC Load Modeling Work Group (LMWG) under WECC Modeling and Validation Subcommittee (MVS) has been continuously developing a composite load model (CLM) representing the aggregation of various end-user load devices. In the WECC CLM model, load devices are categorized based on the features and each category of the loads is represented by an aggregated model, which is expected to be sufficient to duplicate the aggregated behaviors by properly choosing the parameters.

There exist a large number of disparate load devices connected to load buses in utility networks. Since the loads may spread along with feeders of different structures connected to a load bus, the voltage levels seen by individual motors during a fault event are different from those at the load bus. In addition, the control and protection settings of individual dynamic loads such as motors can be different. Therefore, it is difficult to parameterize the CLM in different motor load categories to accurately emulate the dynamic behaviors of end-user loads under various disturbances. The existing issues mainly include:

· There is a lack of a systematic approach to tuning the CLM parameters for reproducing the behaviors of the aggregated loads and the entire system under all disturbances and operating conditions.

· Current practices rely on engineers’ knowledge and experience with the system and are exclusively trial-and-error efforts based on only a limited number of disturbance and operating condition scenarios.

· Experience has shown that the CLM parameters carefully selected for a load bus during one fault event usually cannot achieve satisfactory performance in another fault or for a different load bus but with similar load composition.

In this study, we propose to develop an optimal parameterization scheme for the WECC CLM based on a simulation approach assisted by a unique combination of state-of-the-art machine learning (ML) techniques including imitation learning (IL) and reinforcement learning (RL) or hybrid-imitation-reinforcement learning (h-IRL). The parameterization scheme will be used for a thorough validation and benchmarking of the WECC CLM model and for developing optimal parameters for the CLM to duplicate dynamic behaviors of the aggregated loads and the entire system under various disturbances and operating conditions of the grid. The same parameterization scheme will be extended for cases with incomplete information about the loads, i.e., some parameters of the loads are unknown while some other information is becoming more and more available based on DOE’s efforts. Additionally, we will propose an alternative approach to develop a library characterizing the basic load features that can be used to better build composite load models.

To facilitate the development of the project, we have the following tasks that need help

· Model uncertainties due to renewable generation using historical renewable generation data and demand variation.

· Provide technical support for BNL to develop the h-IRL architecture.

· Provide technical support to BNL in developing ML algorithms for identifying the structures and composition of different types of load devices based on the post-contingency responses of the distribution feeder.

Qualifications include experiences in applications of graph neural networks and deep operator learning to power system applications, especially in the prediction of networked system dynamics.