Applications of Smart Grid Big Data Analytics

GAO Feng , LIU Guangyi , SAUNDERS Chris ,ZHU Wendong , TAN Chin-woo , YU Yang

(1.Smart Grid Research Institute North America Inc., Santa Clara, CA 95054;2. China Electric Power Research Institute, Beijing 100192, China; 3. Stanford University，Palo Alto 94305)

?

Applications of Smart Grid Big Data Analytics

GAO Feng1, LIU Guangyi2, SAUNDERS Chris1,ZHU Wendong1, TAN Chin-woo3, YU Yang3

(1.Smart Grid Research Institute North America Inc., Santa Clara, CA 95054;2. China Electric Power Research Institute, Beijing 100192, China; 3. Stanford University，Palo Alto 94305)

With the rapid progress in information technology, a novel concept—“Energy Internet”, that concentrates on the coordination and optimization of multi-type energy flows via advanced communication and internet technology, has received a lot of attention. The result of such an inevitable trend is that a fundamental technique—Big Data Analytics—must be developed to handle massive influx of data from multiple heterogeneous sources, as well as utilize the data to swiftly derive an economic value. The paper gives an overview on research works conducted at SGRI North America big data lab with highlights on hardware configuration and software deployment of the cluster environment. The paper reviews several ongoing research topics performed in the lab with an emphasis on customer segmentation and response targeting (collaboration with Stanford University); and energy disaggregation. These works are built on an integrated power system data model that is supported by open source technology. Preliminary results show that our research will benefit both utility companies and customers.

Big Data Analytics; mixed-integer programming; customer segmentation and targeting; energy disaggregation

0 Introduction

With the rapid progress in information technology, a novel concept-“Energy Internet”, that extends “Smart Grid” to concentrate on the coordination and optimization of multi-type energy flows via advanced communication and Internet technology, has received a lot of attention. The deployment of a massive amount of distributed, intelligent, cost-effective sensors, controllers, meters, and processors within the value chain of energy system forms the foundation for “Energy Internet”. The result of such an inevitable trend is that a fundamental technique must be developed that can handle massive influx of data from multiple heterogeneous sources, as well as utilize the data to swiftly derive an economical value. Big data analytics, as a process of examining large data sets containing a variety of data types to uncover hidden patterns, unknown correlations, market trends, customer preferences and other useful business information, has consequently increased its popularity.[1]

The utility industry is facing unprecedented challenges caused by extremely high volume and high frequency measurement data. According to the Navigant Research Report, the estimated installed base of smart meters worldwide will surpass 1.1 billion by 2022.[2]Advanced Metering Infrastructure (AMI) typically collects electricity usage data in the range of 15 minutes to 1 hour. This is up to a three thousand fold increase in the amount of data utilities would have processed in the past.[3]Meanwhile, synchrophasor is being deployed around the global that collects a large volume of low-latency, real-time streaming measurement data. Phasor Measure Unit (PMU) can measure AC waveforms (voltages and currents) typically at a rate of 48 samples per cycle (2 880 samples per second for 60 Hz systems)[4]. Just one Phasor Data Concentrator (PDC) collecting data from 100 PMUs of 20 measurements each at 30 Hz sampling rate generates over 50 GB of data one day.[5]

Big data analytics provides a suite of techniques for the utility industry that are deemed to resolve these challenges. The in-memory calculation engine and parallel computing framework, Hadoop/MapReduce and Spark, are ready for handling an extremely large scale of dataset; on the other hand, the stream processing engine, Storm, Streams, and Spark Streaming, are built to analyze data in motion and act on information as it is happening.

Consequently, big data analytics could be applied to improve both power system short-term operations and long-term planning processes. The promising applications for big data analytics include detection of energy theft, strategic adoption for electric vehicle and rooftop solar integration, fine granularity load forecast and renewable generation forecast, distribution system topology identification, online asset risk assessment, distribution system voltage and var optimization, customer segmentation & targeting, and revenue protection etc.[3]

The paper gives an overview on research works conducted in SGRI North America big data lab with highlights on hardware configuration and software deployment of the cluster environment. The paper reviews several ongoing research topics performed in the lab, for example: a conceptual design of data structure and computing architecture for smart grid, customer segmentation and response targeting (collaboration with Stanford University), and energy disaggregation. These works are built on an integrated power system data model that is supported by open source technology. Preliminary results show that our research will benefit both utility companies and customers.

1 Big Data Laboratory for Power Application

The SGRI North America big data lab was funded by SGRI North America Inc., and the construction will be finished by the end of 2015. The lab will serve as an integrated development environment and testing bed for data integration, data management and advanced optimization, data mining, and data analytics. The lab is composed of Hadoop cluster environment, data extract/transform/load platform, data mining & exploration platform, and data analysis platform, with two hundred terabyte initial processing power and extensible capacity.

The lab is equipped with fourteen high performance computing facilities and state-of-the-art apache Hadoop packages. Fig. 1 gives an overview of system architecture for the lab. Specific hardware and software configurations are shown as below:

·The lab is equipped with eight 4-U 4-Socket servers, six 2-U 2-Socket servers, and one 2600T storage server.

·The lab intranet consists of one 10 G switch, two 1 G switches, and one fiber switch. Firewall and wireless access point technology are deployed to assure cyber security.

·The lab deploys enterprise-grade big data platform—BigInsights/Open Data Platform. The platform integrates Apache open source components: Hadoop/MapReduce, Yarn, Spark, Hive, HBase, and big data analysis tools: BigSQL, BigSheet, and Text Analytics.

·The lab deploys real-time big data analytic tool— Streams, to provide integration and analysis services for streaming data.

·The lab deploys large-scale optimization and decision analysis package-ILOG CPLEX to address multi-dimensional optimization, decision, pricing and resource allocation problems for “Energy Internet”.

·The lab deploys business intelligence package to serve as an integrated environment for managing structural data and performing multi-dimensional analysis.

·The lab deploys open source programming language R that provides a complete set of data processing, statistical analysis, computation, and visualization tools.

圖1 大數(shù)據(jù)實(shí)驗(yàn)室系統(tǒng)架構(gòu)Fig.1 System architecture for big data lab

In the lab, the research and development works are performed on an end-to-end platform for addressing the needs of the future “Smart Grid” and “Energy Internet”. A holistic solution including data processing, storage, analytics, and visualization technologies used in this lab enable rapid insights and informed decision making for meeting the business demands of a power system where distributed renewable generation, continuous metering and system monitoring, and frequent customer interaction are a reality. Fig. 2 illustrates software packages and workflow deployed in big data lab.

2 Demonstration of Research Works

2.1 Conceptual Design for Hierarchic Smart Grid Big Data Analytics Platform

The lack of a unified data model blocks the efficient data integration and module deployment for smart grid big data analytics platform. Our research teases out five rules and three criteria (shown in Fig. 3) for building a platform designed to comprehensively enable resilient features of smart grid big data.[6]

We propose a conceptual framework to enable situational awareness and control capabilities within a smart grid, by establishing a unified structural hierarchy which simultaneously yields an organizational pattern for the storage, processing, analytics, and control of a smart grid.[7]

·Comprehensive: store, process and communicate all available information and integrate all available tools and interfaces.

·Adaptive and dynamic: adaptively organized, dynamically innovated, and function-based structured to support all kinds of process and analytics.

圖2 大數(shù)據(jù)實(shí)驗(yàn)室軟件配置與流程圖Fig.2 Software packages and workflow for big data lab

圖3 智能電網(wǎng)大數(shù)據(jù)平臺(tái)建設(shè)標(biāo)準(zhǔn)Fig.3 Rules and criteria for smart grid big data platform

·Distributed and multi-layered: multi-layered and each component are equipped with an intelligent “brain”. Only necessary data and information are communicated in-between.

We propose a Spark Streaming-based Real-Time Complex Event Processing framework for the smart grid (shown in Fig. 4). A number of commercial and open source events processing software are available for building complex event processing applications. Apache Spark is a fast and general engine for large-scale data processing[8]. It is a unified platform combining Spark SQL, Spark Streaming, and ML-Lib for machine learning and GraphX. Spark now boasts the ability to not only process streams of data at scale, but to “query” that data at scale using SQL-like syntax.[9]

圖4 復(fù)雜事件處理架構(gòu)Fig.4 Complex event processing system architecture

2.2 Customer Segmentation and Response Targeting (Collaboration with Stanford University)

圖5 大數(shù)據(jù)解析應(yīng)用流程圖Fig.5 Advanced big data analytics workflow for power application

We design a workflow for potential Big Data Applications that include segmentation, targeting, and disaggregation etc. (shown in Fig. 5). Segmentation based on consumption behavior benefits both utilities and customers. The primary benefit for utilities is to achieve higher returns in demand response programs as well as equipping decision makers with information to advance resource allocation, pricing, and program development. Segmentation also provides time of day consumption, daily usage pattern stability over time, as well as actual volume of energy use that would potentially drive customers to improve home energy efficiency.

Traditional segmentation method is based on customer self-report and survey data, without leveraging real electricity consumption measurement. The widespread deployment of smart meters creates opportunities for segmentation strategies based on 15 min, 30 min or hourly household energy use. Collaborating with Stanford University, we integrate segmentation package with commercial database and implement a novel browser-server based analysis platform. The segmentation method is called adaptiveK-means with a customized threshold to construct a shape dictionary[10]. The algorithm starts by a set of initialized cluster centers utilizing a standardK-means algorithm, with an initialK=k0. AdaptiveK-means then adds additional cluster centers, whenever a load shapes(t) in the dataset violates the mean squared error threshold condition as shown in (1):

(1)

where:

s(t):loadshape;

Ci*(t)(t):representativeloadshape;

θ:clusterthreshold.

Thestepsofthealgorithmarehighlightedasbelow.Thedetailedexplanationisinreference[10].

·Calculatingdailytotalconsumptioncharacterizationanditsprobabilitydistribution;

·Encodingsystembasedonapre-processedloadshapedictionary;

·ApplyingadaptiveK-means on normalized data;

·Performing hierarchical clustering.

The drive towards more green energy has enabled signicant growth of renewable generation. Demand response (DR) has become an efficient practice to address the intermittence issues caused by deep penetration of renewable generation. It becomes important to be able to target the right customers among a large population to keep DR enrollment cost low. The availability of high resolution smart meter information can signicantly reshape such a targeting schema.

We collaborate with Stanford University and integrate targeting package into the analysis platform. Since an enrollment decision is made in advance of the actual consumption period, only a prediction of the DR potential is available. The prediction can be estimated by analyzing the historical high resolution consumption data for each customer. Given a prediction of DR potential being a random variable, the targeting package chooses sufficient number of customers to balance the magnitude of demand response potential, and the uncertainty in the prediction.[11]

Assume the utility desires to enroll up toNcustomers from a population ofKindividuals, aiming to achieve at leastTkWh of energy savings with high probability. The targeting problem can be stated as:

The goal is to maximize the likelihood of saving at leastTkWh, given that we are limited to selecting at mostNcustomers amongKcandidates. In general, this is a stochastic knapsack formation and belongs to the family ofNP-hard problems.

(2)

where:

rk:energysavingresponseofcustomerk;

xk:decisionvariableforselectionofcustomerk.

We re-visit the problem from a different angle. Assume that the utility desires to achieve at leastTkWh of energy savings with probabilityP0, aiming to enroll the least number of consumers to keep the cost low.

We figure out a schema to transform the problem into a deterministic formulation with its relaxation being a Second Order Conic Programming (SOCP), a special case of Convex Programming. In this case, the problem becomes a little easier to conquer. Several commercial and academic optimization packages would provide SOCP solvers, for example: CPLEX and GUROBI etc.

(3)

where:

rk:energysavingresponseofcustomerk;

xk:decisionvariableforselectionofcustomerk;

Σ:covariancematrixofenergysavingresponse;

P0:probabilitythresholdoftargetingproblem;

Φ:cumulativedistributionfunctionofstandardnormaldistribution.

2.3EnergyDisaggregation

Segmentationandtargetingaretypicallybasedonhouseholdaggregateconsumptiondata.However,energydisaggregation,alsoknownasnonintrusiveloadmonitoring(NILM),isthetaskofseparatingaggregatedataforacustomerintotheenergydataforindividualappliances.Studieshaveshownthatsimplyprovidingdisaggregateddatatotheconsumerimprovesenergyconsumptionbehavior.[12]

Energydisaggregationhasbeenstudiedmorethan30years.Theliteraturecanbeclassifiedintotwomainareas:supervisedandunsupervisedmethods.Superviseddisaggregationmethodsrequireadisaggregateddatasetfortraining.Unsupervisedmethodsdonotrequireadisaggregateddatasettobecollected.However,theydorequirehandtuningofparameters.Theexistingsupervisedmethodsincludesparsecoding[ 1 3],changedetectionandclusteringbasedapproaches[14-15]andpatternrecognition[ 16].TheexistingunsupervisedmethodsincludefactorialhiddenMarkovmodels,differenceMarkovmodelsandvariantsandtemporalmotifmining.Weproposeasupervisedmethodbuiltonamixed-integerprogramming(MIP)formulationthatcanachievehighperformancewithstate-of-the-artbranch-and-cutalgorithm.

Typically,powerconsumptionforaparticularappliancecanbemodelledasmonotonicnon-decreasingcurvewithrespecttoitscontrolinputsthatcouldbeeithercontinuousordiscrete.Forexample,stoveorheatertypicallycanoperateatcontinuestemperaturesettings;andwasherordryercanoperateatdiscreteloadinglevels.Fig. 6illustratesanexemplarconsumptioncurveforanappliancewithrespecttoitscontrolinputs.Apiecewiselinearcurveisusedtoapproximatetheconsumptioncharacteristics.

WeproposeaMIPmodeltooptimizetheconsumptionerrorresidualsoverthetimehorizon.Meanwhile,eachapplianceisconstrainedbyitscharacteristics.Additionally,wecanmodelmore“l(fā)ogic”constraintsfordevices,forexample,typicallyitisnottruetoturnonorturnoffastoveorheateratahighfrequentrate.Thedeviceswouldratherstayat“on”or“off”statusduringapre-definedtimeperiodinreality.Theseadditional“l(fā)ogic”constraintscanbeeasilyimplementedwithinaMIPmodel.

圖6 設(shè)備用電特性曲線Fig.6 An example of device consumption curve

The objective function of the MIP model is to minimize error residues for the power consumption over the time. Equation (4) describes device’s power consumption must be within high and low limits. Equation (5)～(7) are an “efficient” implementation for “l(fā)ogic” requirement that device has to stay at the same state for a certain time period after switching status. These constraints define the most compact relaxation area called “Convex Hull” for a MIP model that will speed up the overall performance. Fig. 7 illustrates several different MIP formulationsv.s. Convex Hull formulation. More discussion on efficient formulation for MIP model can be found in reference.[17]Equation (8) defines devicei’s power consumption characteristics.

圖7 優(yōu)化約束的Convex Hull模型Fig.7 Alternative formulation and convex gull in MIP min ‖Y·1-E‖1,2,

s.t.

(4)

(5)

(6)

vti≥uti-u(t-1),i

(7)

yti=Ci(xti)

(8)

wheredecisionvariables:

xti:devicei’s control input at timet;

uti:devicei’s status at timet, 1: on, 0: off;

vti:devicei’s start-up status at timet, 1: turn on att;

yti:devicei’s control output at timet,i.e. consumption att;

Y:aT-by-Imatrix of device control output variables;

parameters:

oni:devicei’s minimum stay-on time;

offi:devicei’s minimum stay-off time;

T:number of time intervals under study;

I: number of devices;

E:aT-by-1 vector of power measurement input;

1:aI-by-1 vector with all elements equal to 1;

functions:

Ci:devicei’s power consumption characteristics with respect to control inputs.

3 Numerical Examples

We now illustrate some selected numerical examples for the applications of customer segmentation, targeting and energy disaggregation at SGRI North America big data lab. All numerical examples are running on a computer with 8 G RAM and Intel Dual Core CPU @ 2.40 GHz.

3.1 Customer Segmentation

We create about six million records for customer meter data by Gridlab-D simulation tool and store them into Oracle and PostgreSQL database. Our software platform built on browser-server architecture is developed by R-shiny that is an open source tool. R script is fully supported as a complete set of data processing, statistical analysis, computation, and visualization tools.

Fig.8 shows multidimensional segmentation results based on smart meter data. We describe customers by two aspects: quantity and variability. The left graph uses Entropy and logarithm of average daily consumption. The right graph adds one more dimensionality that is maximum daily consumption. As shown in both graphs, those households with heavy and stable consumption behaviors would become good candidates for demand response during peak days.

圖8 用戶分類分組示意圖Fig.8 Multidimensional customers segmentation examples

We are performing more business intelligence analysis based on meter data. Fig. 9 shows the software would select top usage customers based on minimum and maximum daily consumption criteria set by uses.

Fig. 10 shows the most popular load shapes in the sample data. This is again based on adaptiveK-mean clustering algorithm. The top rated shape (56%) is of twin-peak: morning peak and evening peak, which are typically caused by more activities occurring during the time in a household.

圖9 用戶最小最大日用電量Fig.9 Customer sorted by minimum and maximum daily consumption

圖10 典型負(fù)荷曲線示意圖Fig.10 Top 9 typical load shapes

3.2 Customer Targeting

Fig. 11 draws several plots to illustrate customer targeting results. The left shows a graph of “ReliabilityvsNumber of Selected Customers” given target energy saving equal to 100 kWh. The right one shows a graph of “ReliabilityvsAvailable Demand Response Energy” given the number of customers to recruit equal to 200. Based on the same principle, one can draw a plot of “Available Demand Response EnergyvsNumber of Selected Customers” given a reliability level.

圖11 定量需求響應(yīng)示意圖Fig.11 Response targeting examples

3.3 Energy Disaggregation

We implement the MIP model for Energy Disaggregation described in Section 2.3 by IBM Optimization Programing Language (OPL) that is a type of high-level scripting language seamlessly integrated with CPLEX optimization engine. We use OPL to create a moving window for energy disaggregation on an appliance level with an interval equal to 30 s. The total time interval for disaggregation is 600 s. All of the parameters are configurable in real time. Table 1 lists consumption characteristics for five types of appliances represented by piecewise linear curves.

表1 用電設(shè)備特性參數(shù)

Table 1 Parameters for five devices’ consumption characteristics

In this example, we selectL1-norm as the objective function to minimize the aggregated absolute error residues of consumption data，however, our software does supportL1,L2, and L-infinite norms in the optimization model. We use a random term with normal distribution to simulate the electricity consumption over time produced by smart meters. Fig. 12 shows the total power consumption from both measurement and disaggregation result. During most of the time points, the error residue is small except at two time points the relative error is beyond 25%. The root cause is that all of the data are from a stochastic simulation process and between time point 10 and 15 there is a significant change in simulation data. In reality, the case is having a low probability to occur.

圖12 總用電量與擬合誤差Fig.12 Total power consumption and relative error

Fig. 13 shows device’s on and off status from disaggregation results. During the two valley spots, the Washer and HAVC got shutdown accordingly.

Table 2 gives each device’s power consumption based on disaggregation results. The result is shown in Fig. 14. Our energy disaggregation tool can operate on a continuous rolling window mode. All of the simulations perform very efficiently, typically less than 1 s.

圖13 基于電能分解的設(shè)備開關(guān)狀態(tài)Fig.13 Device status based on energy disaggregation

圖14 基于電能分解的設(shè)備用電量Fig.14 Device consumption based on energy disaggregation

4 Conclusion

The paper gives an overview on research works conducted at SGRI North America big data lab with highlights on hardware configuration and software deployment of the cluster environment. The paper reviews several ongoing research topics performed in the lab, for example: a conceptual design of data structure and computing architecture for smart grid, customer segmentation and response targeting (collaboration with Stanford University), and energy disaggregation. These works are built on an integrated power system data model that is supported by open source technology. Preliminary results show that our research will benefit both utility companies and customers. The future directions would include integrating the existing applications with parallel computing framework at big data lab to accelerate the overall performance and making the software platform mature eventually.

表2 基于電能分解的用電量

Table 2 Device power consumption

Acknowledgement

The authors would like to thank Dr. J. Kwac and Prof. R. Rajagopal for their contributions. The authors also express their thanks to Prof. C. Kang for the review and comments.

5 References

[1]What is big data analytics?-Definition from WhatIs.com[EB/OL]. http://searchbusinessanalytics.techtarget.com/definition/big-data-analytics.

[2]Smart electric meters, Advanced metering infrastructure, and meter communications: Global market analysis and forecasts[EB/OL] Navigant Research, 2013. http://www.navigantresearch.com/research/smart-meters.

[3]Yu N, Shah S, Johnson R, et al. Big data analytics in power distribution systems[J]. IEEE ISGT 2015.

[4]Wikipedia[EB/OL]. https://en.wikipedia.org/wiki/Phasor_measurement_unit

[5]Xie L, Chen Y, Kumar P R. Dimensionality reduction of synchrophasor data for early anomaly detection: linearized analysis[J]. IEEE Transactions on Power Systems, 2014, 29(6): 2784-2794.

[6]Liu G, Yu Y, Gao F,et al. Research of smart gird big data model[C]. 23rd International Conference on Electricity Distribution Lyon, 2015.

[7]Saunders C, Liu G, Yu Y, et al. Data-driven distributed analytics and control platform for smart grid situational awareness[C]. CSEE, 2015.

[8]Spark official website[EB/OL]. https://spark.apache.org/.

[9]Liu G, Zhu W, Saunders C, et al. Real-Time Complex Event Processing and Analytics for Smart Grid[C]//Complex Adaptive Systems, San Jose: Conference Organized by Missouri University of Science and Technology, 2015.

[10]Kwac J, Tan C, Sintov N, et al. Utility customer segmentation based on Smart Meter Data: Empirical Study[J]. IEEE Smart Grid Community, 2013.

[11]Kwac J, Rajagopal R, Demand response targeting using big data analytics[C]// 2013 IEEE International Conference on Big Data.

[12]Dong R, Ratliff L, Ohlsson H, et al. A dynamical systems approach to energy disaggregation[C]//2013 IEEE 52nd Annual Conference on Decision and Control (CDC).

[13]Kolter J Z, Batra S, Ng A Y. Energy disaggregation via discriminative sparse coding[C]//Advances in Neural Information Processing Systems, 2010.

[14]Drenker S, Kader A. Nonintrusive monitoring of electric loads[J]. IEEE Computer Applications in Power, 1999, 12(4): 47-51.

[15]Rahayu D, Narayanaswamy B, Krishnaswamy S. Learning to be energy-wise: discriminative methods for load disaggregation[C]//Third International Conference on Future Energy Systems: Where Energy, Computing and Communication Meet (e-Energy), 2012. [16]Farinaccio L, Zmeureanu R. Using a pattern recognition approach to disaggregate the total electricity consumption in a house into the major end-uses[J]. Energy and Buildings, 1999, 30(3): 245-259.

[17]Rajan D, Takriti S. Minimum up/down polytopes of the unit commitment problem with start-up costs[R].IBM Research Report, 2005.

高峰(1977)，博士，IEEE與INFORMS會(huì)員，高級(jí)工程師。研究方向?yàn)殡娏ο到y(tǒng)經(jīng)濟(jì)運(yùn)行、電力市場(chǎng)交易、優(yōu)化與決策分析、電力大數(shù)據(jù)、電力系統(tǒng)規(guī)劃，人工智能技術(shù)；

劉廣一(1963)，博士，教授級(jí)高級(jí)工程師，研究方向?yàn)殡娏ο到y(tǒng)經(jīng)濟(jì)調(diào)度、網(wǎng)絡(luò)分析、 EMS/DMS、主動(dòng)配電網(wǎng)、電力大數(shù)據(jù)和電力市場(chǎng)；

Chris Saunders(1980)，博士，高級(jí)工程師，研究方向?yàn)閿?shù)據(jù)庫架構(gòu)、電力系統(tǒng)數(shù)據(jù)分析、電力系統(tǒng)數(shù)據(jù)挖掘、機(jī)器學(xué)習(xí)、電力系統(tǒng)運(yùn)行與規(guī)劃；

朱文東(1969)，碩士，高級(jí)工程師，研究方向?yàn)殛P(guān)系型數(shù)據(jù)庫、NoSQL & NewSQL、分布式計(jì)算、實(shí)時(shí)流數(shù)據(jù)計(jì)算、基于內(nèi)存計(jì)算的開源集群計(jì)算平臺(tái)Spark及其在電力系統(tǒng)中的應(yīng)用；

陳振宇(1961)，博士，斯坦福大學(xué)智能電網(wǎng)研究所主任。研究方向?yàn)橹悄茈娋W(wǎng)大數(shù)據(jù)，分布式能源規(guī)劃及運(yùn)作，電力系統(tǒng)動(dòng)態(tài)分析；

于洋，博士，研究方向?yàn)殡娏κ袌?chǎng)設(shè)計(jì)、新能源接入、配電側(cè)市場(chǎng)設(shè)計(jì)、電力消費(fèi)分析、電力系統(tǒng)機(jī)制設(shè)計(jì)的環(huán)境影響、電力市場(chǎng)與碳市場(chǎng)等。

(編輯：劉文瑩)

國家自然科學(xué)基金資助項(xiàng)目(51261130472)； 國家電網(wǎng)公司科技項(xiàng)目 (DZB51201403772)。

TM 76

A

1000-7229(2015)10-0011-09

智能電網(wǎng)大數(shù)據(jù)的分析與應(yīng)用

高峰1，劉廣一2，Chris Saunders1， 朱文東1， 陳振宇3,于洋3

(1. 國網(wǎng)智研院美國研究院，Santa Clara, CA 95054；2. 中國電力科學(xué)研究院，北京市 100192；3. 斯坦福大學(xué)，Palo Alto 94305)

能源互聯(lián)網(wǎng)的興起是全球能源環(huán)境與經(jīng)濟(jì)發(fā)展雙重壓力導(dǎo)致的結(jié)果，這種趨勢(shì)不可避免地促進(jìn)大數(shù)據(jù)分析技術(shù)的快速發(fā)展。大數(shù)據(jù)技術(shù)是指從各種各樣類型的海量數(shù)據(jù)中，快速獲得有價(jià)值信息的技術(shù)。簡(jiǎn)要介紹了國網(wǎng)智研院美國研究院大數(shù)據(jù)團(tuán)隊(duì)的研究工作和實(shí)驗(yàn)室軟硬件配置，介紹了智能電網(wǎng)數(shù)據(jù)結(jié)構(gòu)和計(jì)算架構(gòu)等概念性的設(shè)計(jì)。重點(diǎn)介紹了與斯坦福大學(xué)合作進(jìn)行的用戶分組分類和需求響應(yīng)定量計(jì)算方法，以及非侵入式電能分解。這些工作基于集成的數(shù)據(jù)模型和開源軟件技術(shù)，將為電網(wǎng)公司和客戶同時(shí)帶來收益。

大數(shù)據(jù)解析；混合整數(shù)規(guī)劃；用電行為分析；電能分解

2015-08-03

2015-09-14

10.3969/j.issn.1000-7229.2015.10.002

Project supported by National Natural Science Foundation of China (51261130472);Science and Technology Program of the State Grid Corporation of China (DZB51201403772).