Architecture and Distributed Management for Reliable Mega-scale Smart Grids

Abstract:

A primary objective of this research is to establish a foundational framework for smart grids that enables significant penetration of renewable DERs and facilitates flexible deployments of plug-and-play applications.  Under this common theme, the PIs have taken  a  data  analytics  perspective  to  explore  rigorous  approaches  in  modeling, optimization, and control of wind generation integration. Short-term forecast of wind farm generation is studied by applying spatio-temporal analysis to extensive measurement data collected (over two consecutive years)  from a large wind farm. Specifically, graph- learning based spatial analysis  is carried out  to characterize the statistical distribution of the overall wind farm generation and time series analysis is used to quantify the level crossing rate. Built on these characterizations, finite-state Markov chains are constructed for each epoch of three hours, which account for the diurnal non-stationarity and the seasonality of wind generation. Exploiting the Markovian property of the forecast model, the joint optimization of economic dispatch (ED) and interruptible load management  is cast as a Markov decision process (MDP) problem.  Numerical studies, via using the IEEE Reliability Test System – 1996 and realistic wind measurement data from an actual wind  farm,  demonstrate  the  significant  benefits  obtained  by  integrating  the  above forecast model and the interruptible load management, compared with conventional wind-speed-based forecast methods. In another research thrust, PMU measurements from multiple locations, are used for learning,  characterizing  and  classifying  event-specific  spatial  signatures,  and probabilistic models are developed to subsume measurement data.   Both the decision tree approach and dimensionality reduction approach are applied to identify impending signatures of catastrophic events to provide early warning to power system operators. To handle missing PMU data, randomized attribute subsets are used: 1) to reduce the impact of missing PMU data and 2) to reduce the complexity of training small DTs. Further, to attain high reliability, a trustworthy middleware tailored towards smart grid design, is devised to shield the grid design from the complexities of the underlying software world, using automatic generation of invariants for software validation.