Handling the Data Explosion in Tomorrow’s Power Systems
- Written by Aranya Chakrabortty
With grid monitoring systems in various areas already taking billions of data samples each day, and with the numbers sure to become even more gigantic in the near future, development of more sophisticated analytical methods for power system monitoring is imperative. Such work is proceeding apace in the context of the North American Synchrophasor Initiative, as well as at universities and research institutions.
As we gradually move from today’s data-generating power system to tomorrow’s data-driven power network, data analytics and data security are going to play tremendous roles in the research, development and operational evolution of the North American grid. Rapidly progressing instrumentation technologies such as the Wide-Area Measurement System (WAMS) for transmission-level systems, and smart metering for distribution-level systems, will require highly intricate, scalable and decentralized data analysis techniques for postmortem as well as real-time analysis of the grid dynamics.
Analytical methods, based on the advanced ideas of statistical signal processing, pattern recognition and intelligent controls, will become imperative as the number of functional phasor measurement units (PMUs) in the United States grows to over a thousand in the next few years under the Department of Energy’s smart grid demonstration initiative. In the Eastern Interconnect (EI), for example, about 40 PMUs are currently streaming data via the Internet into a super-PDC (phasor data concentrator) located at the Tennessee Valley Authority, processing nearly 2 billion data samples per day.
As volumes of such data become even more gigantic, the primary challenge for system operators will be to deploy fast and robust algorithms that can help them extract important patterns and information in such data. Applications of these algorithms include filtering and predictive modeling for wide-area monitoring, congestion impact analyses of renewable power injection and plug-in-hybrid vehicles at various temporal and spatial scales, and statistical methods for accurate prediction of processing delays and communication latencies from PMU data.
The Operations Implementation Task Team (OITT) of the North American Synchrophasor Initiative (NASPI) is currently carrying out preliminary investigation on the use of pattern recognition of PMU data for several of such applications.
A key feature of these data analyses and processing algorithms, specifically for PMU data, will be reflected in their transition from state-of-the-art centralized approach to a distributed approach. For example, the centralized super-PDC architecture of the Eastern Interconnect will become untenable as the system scales up to 300-400 PMUs required to implement high-voltage transmission PDCs in every control region. Therefore, the challenge will be to develop a distributed architecture that enables wide-area phasor analytics to be divided among hierarchical levels of regional, intra-regional and inter-regional computational clusters that are connected to each other via a resilient communication network.
We cite two practical examples to illustrate the research challenges that may arise with such distributed data processing. The first is modal decomposition of power system dynamics, which is used as a common tool to understand the different types of oscillations stirred in the system by any specific event. The fundamental idea behind modal decomposition is to accumulate PMU data from different parts of a power system, use them to construct dynamic models specific to any particular post-event operating point, and finally to employ the identified models to compute the oscillations incurred in the system at various temporal scales.
The current practice is to apply modal decomposition algorithms to data streams coming from various locations in the system all taken together in one centralized computing node. Given adequate spatial coverage of PMUs, the accuracy of such estimation is optimal because measurements from all locations are jointly evaluated to take maximum advantage of the system observability. However, if the data explosion forces us to depart from this centralized architecture in favor of a distributed computing framework, then subsets of data will need to be processed by individual local computers communicating with each other over a network. Then the obvious question would be – how can the various degrees of freedom in this distributed computation be optimized to guarantee that the final result is as close to the centralized result as possible?
In collaboration with colleagues from the Renaissance Computing Institute (RENCI) at the University of North Carolina Chapel Hill, we are currently investigating such optimization algorithms for identifying the best sets of data from the best sets of locations to be sent to a given set of computers, with constraints posed by the distance between the PMU locations and the computing nodes, and the associated communication delays.
Another example is the construction of performance indices for transient stability monitoring of power systems by using distributed processing of phase angle measurements in real time. Energy functions, quantifying the collective oscillations in kinetic and potential energies of different groups of synchronous generators in the system following a disturbance, are commonly used as one such metric in traditional power system stability studies. These metrics indicate the regions of stability of the system corresponding to the particular disturbance, and how quickly the oscillations at different parts of the system synchronize with each other.
While kinetic energy of a given cluster of generators can be computed from local frequency measurements available from the cluster, potential energy is a 'relative' quantity, and, hence, its computation depends on global analyses of phase angle measurements from all clusters taken together. In terms of our anticipated distributed computing framework, as communication between PMUs and PDCs become decentralized, local computers in each cluster must generate estimates of their 'local' potential energy, and share these estimates with their neighboring clusters in order to develop the global potential energy of the entire interconnection in an iterative fashion. Similarly, if the objective is to develop metrics that indicate specific patterns in spatially dispersed data sets in a distributed fashion, then such algorithms will need to be reformulated as well with the goal of guaranteeing a close-to-optimal result.
At the end of the day, we need to find out if pattern recognition, analytics and data mining of PMU data can indicate the level of immunity of the grid to various kinds of disturbances and failures. Modeling immunity of next-generation power networks will, in fact, be a grand challenge for us. From the security point of view, if such immunity exists, how does it help the system in resisting the impact of an incoming disturbance, both temporally and spatially? Can efficient control designs increase this immunity and pave the way to a new paradigm of wide-area control?
With the recent outburst of smart-grid related research interests in the computing, communications and control communities in the United States, and their gradual unification via common research platforms such as IEEE, answers to such important questions do not seem far away.