R studio Machine learning

project/project/Big Data Applications in Power Systems.html

Big Data Applications in Power Systems

Wenyuan Tang

Lecture 22

In this lecture, we will go over some of the chapters of Big Data Application in Power Systems that are most suitable to be turned into project ideas. You are not restricted to reproduce or improve their results; you can work on the same topic while considering a different objective, a different method, or a different data set.

Deep learning for power system data analysis

Chapter 7 presents a deep learning framework to automatically extract knowledge and use it to improve grid operation. The chapter starts with a moderate introduction to the most well-known deep learning concepts, such as deep belief networks and high-order restricted Boltzmann machines, followed by a discussion on their theoretical advantages and limitations, such as computational requirements, convergence, and stability. As a concrete application, two case studies involving building energy prediction using supervised and unsupervised deep learning methods are presented. The chapter concludes with a glimpse into future trends highlighting some open questions as well as new possible applications.

The Individual Household Electric Power Consumption Data Set from the UCI Machine Learning Repository is used.

Time-series classification methods

The rapid advancement of sensing and measurement technology in power systems has given researchers access to real-time records of system dynamic states. In particular, development of phasor measurement unit (PMU) technology has allowed the continuous monitoring of the transmission line and the connected power systems, and can be complemented with utility monitoring devices, smart meters, and insulation monitoring units to build a thorough picture of the whole grid structure, health, and dynamic behavior. The data collected from these real-time measuring procedures is usually in the form of time series (TS). Hence, Chapter 9 presents an overview about the most recent ML techniques used for TS pattern recognition. The chapter first summarizes existing methods of TS classification and highlights the issue of computational complexity, and then provides discourse on the various dimension reduction and numerosity reduction techniques for a more parsimonious and informative representation of TS data. The chapter concludes with a comprehensive comparison of diverse classification methods in terms of their underlying assumption, performance, computational complexity, flexibility for decentralized execution, and other categories.

One application is fault detection and isolation, e.g., Support Vector Machine Based Fault Detection & Classification in Smart Grids and Transmission Line Fault Classification Based on Dynamic State Estimation and Support Vector Machine. Due to the limited actual data, the simulation of synthetic data is often adopted (e.g., using the relevant toolboxes in MATLAB/Simulink).

Future trends for big data application in power systems

Chapter 10 provides a comprehensive overview of data-driven trends such as feature extraction/reduction and distributed learning to extract knowledge from the power system and market data. Furthermore, it describes the data-driven techniques for dynamic and steady-state analysis and control of distribution and transmission systems.

Some applications include: grid topology identification, e.g., Grid Topology Identification Using Electricity Prices, in which synthetic data are generated for an IEEE test case; price spike forecasting, e.g., A Framework for Electricity Price Spike Analysis with Advanced Data Mining Methods, in which actual market data from the Australian National Electricity Market are used; load forecasting given demand response, e.g., Models and Techniques for Electric Load Forecasting in the Presence of Demand Response, in which demand-side flexibility is modeled and simulated on top of actual load data.

Grid topology identification

Chapters 12 and 13 are devoted to grid topology identification. Knowledge of the exact topology, the open or closed status of switches and circuit breakers throughout the network, is essential for all aspects of the power system operation. Chapter 12 presents a greedy algorithm to learn the grid topology using voltage measurements collected at a subset of the buses in power distribution networks, in which the data in the numerical experiments are generated through simulation. Chapter 13 proposes an algorithm based on the identification of Markov random fields (graphical models) and conditional correlation properties that characterize voltage measurements in power distribution networks. It shows the correlation of voltage magnitude measurements in a radial distribution feeder with the topology of the grid. The numerical experiments are based on an IEEE test feeder.

Predictive analytics for situational awareness of power systems

Being able to forecast energy resources, load patterns, and system state are key features of next-generation smart grid technology. An accurate predictive platform would greatly benefit the planning, scheduling, and unit commitment in terms of both efficiency and security. Chapter 16 provides an overview and a thorough discussion on predictive ML methods for wind, solar energy forecasting, load prediction, power system state estimation, etc. The ML tools included in the chapter range from classical regression, TS analysis, to kernel method such as support vector regression and Gaussian process.

For wind and solar forecasting, the data sets and tools developed by NREL are useful.

Data analytics for energy disaggregation

Chapters 17 and 18 are devoted to a particular yet important application of big data analytics method to smart grid, namely energy disaggregation or nonintrusive load monitoring (NILM). In essence, the goal is to estimate the power usage of individual appliances from an aggregate electricity consumption measurement. Provided with more precise information including itemized energy consumption profiles, both end users and grid managers can improve their utility in terms of energy consumption prediction, demand side management, and user segmentation. Chapter 17 surveys the existing literature for background, ML methods, and possible applications of energy disaggregation, while Chapter 18 discussed the issue of privacy in the energy disaggregation framework. Both chapters are witness to the combination of cutting edge ML methods and a deep understanding of the system characteristics for the advancement of smart grid technologies.

Pecan Street’s Dataport is the world’s largest residential energy and water research database. As a university member of Dataport, Pecan Street has provided access to static time-series datasets (1-second energy, 1-minute energy, and 15-minute energy) for 25 homes from three of their research regions (New York, California, Austin). The Disaggregation Toolkit is a simple python script that evaluates the performance of energy disaggregation machine learning models. Researchers can see how well their training model works by using Pecan Street’s electricity data available on Dataport to compare their predicted data against actual data.

A open source NILM toolkit called NILMTK has been developed to enable the comparison of energy disaggregation algorithms in a reproducible manner.

Data resources

Some other useful data resources are listed in the following:

• Smart* Data Set.
• Open Data Sets.
• Open Power System Data.
• Energy Data Resources.
• Texas Synchrophasor Network and a guide.
• NOAA Climate Data.