The operation of current power systems depend on deterministic and static models, which are not suitable for analyzing smart power grids due to the increasing large-volume of data collected by the grids and sensors and the need to integrate intermittent renewable resources and dynamic load compositions. Large uncertainty in the model prediction is problematic as it does now allow careful planning, and failure to identify large fluctuations and possible instabilities could endanger the reliable operation of the power grid. Hence, it is crucial to incorporate new monitoring capabilities realized by new tools such as machine learning and predictive multi-rate modeling in modeling the smart grid. Classical methods that deal with uncertainty lead to inefficient solutions as they are too slow to converge to a solution and hence they cannot be used effectively for real-time control of power grids. This difficulty stems from the requirement of sampling the very complex power grid thousands of times in order to arrive to a reasonably accurate solution. The goal of this project is to establish significant advances in research and education in the development of machine learning and real-time predictive modeling of power systems, with particular focus on the smart grid.
Machine learning and real-time predictive modeling have received increasing attention in recent years. Extensive research effort has been devoted to these topics, and novel numerical methods have been developed to efficiently deal with sensor data and complex engineering systems. Both machine learning and real-time predictive modeling enable us to better extract the useful information from available sensor data and make critical decision in real time with the presence of uncertainties. For example, solar and wind energy will depend on the weather condition. Machine learning and real-time predictive modeling are thus critical to many important practical problems such as power system stability analysis and social cyber-network prediction, etc. For large-scale power systems, deterministic simulations can be very time-consuming, and conducting predictive simulations further increases the simulation cost and can be prohibitively expensive. One of the biggest challenges in machine learning and real-time predictive modeling is how to develop hierarchical reduced-order models and how to fuse information from such hierarchical reduced-order models. This project aims to address these critical challenges. A novel set of deep-learning based multi-fidelity algorithms (deep Gaussian processes) will be developed for real-time prediction of power systems. The approach under development in this research project is based on scalable algorithms for building deep-learning based reduced-order models for efficient power system dimension reduction. The new algorithms will be based on building multi-fidelity models via deep learning for power systems, and they will significantly advance the current state of the art of deep learning and real-time predictive modeling. The project will also integrate educational opportunities and will expand the population of modelers who use machine learning and predictive modeling tools to solve network problems. The project will expose a diverse group of undergraduates and minority students to machine learning and predictive modeling.
