PowerPlay: Creating Virtual Power Meters
Contents
1. ized step feature Deltas previously assigned to other fea tures are excluded from consideration in this process Cycle Detector Unlike the detectors above the cycle detector operates on a series of labeled features from the detectors above and then i identifies each potential cyclic feature from the data and ii chooses a sequence of the fea tures that most closely matches the cycle s expected period length Figure 6 illustrates the process where the cyclic fea ture is a spike To determine the best sequence of cyclic fea tures of a particular type we chose an arbitrary cyclic fea ture of the type at time t then the next one closest to time t2 t period and so on for tg t _1 period To account for features missed by its particular feature detector we may also match fy to tk tk 1 2 period The error of the re sulting sequence of tg is computed as t tk 1 period i e the amount the sequence differs from the expected pe riod This error is computed for all sequences starting from each possible t and the detector selects as the predicted cy cle the sequence with the lowest total error After determin ing the sequence of cycle on events we filter and recon struct the feature s energy usage by filling in its correspond ing load s model starting from each on event as shown in the final step of Figure 6 E a Spike Detect b Sequence c Reconstruct Figure 6 Op2. s energy usage profile over time which is based on a small number of fundamental electrical characteristics 1 e whether a load is resistive inductive non linear or cyclical A detailed description of these load types and their corre sponding models is described in prior work 3 We select a compact set of identifiable load features from the models and then design efficient online methods for tracking loads by detecting one or more of these features in smart meter data In doing so we make the following contributions Feature Selection We describe a compact set of fea tures that loads may exhibit including power steps spikes growths decays oscillations and cycles We extract a load s features from its model and then choose a small set of iden tifiable features for tracking Using only identifiable features to track loads increases efficiency compared to using every feature while maintaining accuracy Online Load Tracking For each feature we design effi cient online methods to detect that feature in smart meter data Since a load may exhibit multiple features tracking a load may require using multiple feature detectors Hence we present an online tracking algorithm that combines multiple feature detectors to efficiently detect and track loads Implementation and Evaluation We implement our load tracking system called PowerPlay and evaluate it live us ing a 1Hz power meter We show that our approach enables e
3. LMA 20 to perform curve fitting If the fit fails or the derived de cay growth parameter is far from the expected value the de tector moves on to the next possible on event If the fit is successful then the detector identifies the off event for the device or equivalently the duration of the decay growth To do this the detector gradually extends the fitted curve while looking for an off step of the expected magnitude based on the magnitude of the on step plus the cumulative growth or decay of the fitted curve which increases with the length of the curve The detector then chooses the off step within a bounded interval most closely matching the expected value In this case bounding prevents a runaway search After se lecting the off step the detector is able to trivially recon struct the entire feature based on the identified on and off events and the fitted curve between them The process of fit ting and filtering a decay feature is illustrated in Figure 5 Spike Detector Power spikes manifest themselves across multiple seconds either due to variation in a load s exact ac tivation time i e when it activates within the one second sampling interval or due to a short ramp up period which is a On Step b Fit Reconstruct c Off Match d oe 5 Operation of the decay growth detector especially prevalent in high wattage loads Thus the spike detector collapses
4. oscillations LCD TV cyclic 700 spike 600 min fi oscillations 500 400 4 1300 4 200 100 d Heater i e Refrigerator iai d Reconstruct a Filter b Label c Cluster Figure 3 Detection of a stable oscillation feature 4 2 Feature Detection PowerPlay s tracking algorithm relies on individual fea ture detectors to identify the features described in 3 includ ing power steps spikes growth decay bounded oscillations and stable min max oscillations We detail each of these fea ture detectors below Stable Oscillation Detector This detector examines data for frequent power oscillations from a stable minimum or maximum power level such that for every negative power delta i e a power drop there is a corresponding positive power delta in the near future More formally it identifies a stable power oscillation feature by scanning a recent window of data while maintaining a stable power level p which it updates only if power deviates from p by at least T watts for at least D seconds The parameters T and D are specific to a particular device that exhibits this feature Power changes that update p are considered background activity which are excluded from the stable power oscillation feature while any other oscillations within the window are flagged for consid eration in the feature Finally we cluster nearby groups of labeled points to result in the time
5. 18 26 in two important respects Simplicity Online load tracking is a simpler problem than complete load disaggregation load tracking targets in dividual loads while complete load disaggregation focuses on disaggregating an entire building by apportioning its total energy usage across every load Clearly if complete accu rate and inexpensive disaggregation was feasible it would subsume the problem of online load tracking However tech niques for complete disaggregation continue to suffer from inaccuracy especially when disaggregating small loads or scaling up to large numbers of loads 1 Thus load tracking is better suited for scenarios where disaggregating all of a building s loads is either infeasible due to the large number of loads or simply not necessary Efficiency Prior disaggregation techniques implicitly as sume offline analysis and are often computationally expen sive In contrast load tracking explicitly targets online mon itoring in near real time This leads us to focus on perfor mance issues not addressed in prior research such as en abling tracking to either i run on the low power embedded platforms used in smart meters or ii scale to thousands of loads on server platforms To enable high performance we take a model driven approach to load tracking which focuses on detecting a small number of identifiable load features in smart meter data These features derive from a parameterized model of a load
6. 2011 14 J Kelso editor 20 1 Buildings Energy Data Book Department of Energy March 2012 15 E Keogh and M Pazzani Dynamic Time Warping with Higher Order Features In SDM April 2001 16 H Kim M Marwah M Arlitt G Lyon and J Han Unsupervised Disaggregation of Low Frequency Power Measurements In SDM April 2011 17 N Klingensmith D Willis and S Banerjee A Distributed Energy Monitoring and Analytics Platform and its Use Cases In BuildSys November 2013 18 J Kolter and M Johnson REDD A Public Data Set for Energy Dis aggregation Research In SustKDD August 2011 19 J Kolter and A Ng Energy Disaggregation via Discriminative Sparse Coding In NIPS December 2010 20 K Levenberg A Method for the Solution of Certain Non Linear Prob lems in Least Squares Quarterly of Applied Mathematics 1944 21 S Patel T Robertson J Kientz M Reynolds and G Abowd At the Flick of a Switch Detecting and Classifying Unique Electrical Events on the Residential Power Line In Ubicomp September 2007 22 J Taneja D Culler and P Dutta Towards Cooperative Grids Sen sor Actuator Networks for Renewables Integration In SmartGrid Comm 2010 23 Energy Inc http www theenergydetective com 24 The Power Consumption Database http www tpcdb com 25 Smart Meter Implementation Programme https www gov uk government uploads system uploads attachment_data file 42737 1480 design requireme
7. g that include the largest changes in power to the feature set and then executing our tracking algorithm on historical data until the tracking error factor is below a pre defined threshold 4 Online Load Tracking In this section we first describe PowerPlay s online track ing algorithm and then describe the various feature detection techniques the algorithm uses to detect the features from 3 The right side of Figure 1 depicts this process 4 1 Tracking Algorithm PowerPlay s tracking algorithm takes as input a set of loads to track a set of identifiable features for each load and a continuous stream of data from a smart meter Feature detectors for each load operate over a moving window of data points of size W starting from the most recent data point in the time series of a home s power readings i e a sliding window ending with the most recent reading The window represents the minimum time period over which a feature manifests itself The output of the tracking algorithm acts as a set of virtual power meters providing device level power data for each tracked load 260 1000 r bounded oscillations p 255 980 stable step step 250 4 960 245 940 240 920 235 900 230 b i 880 i b Coffee pot Figure 2 Annotated features from representative loads a Light PowerPlay orders the list of all identifiable features across all loads into three sets from most to leas
8. inferring its average power usage p t from a home s total power usage P t recorded by its smart meter over the period t t t Due to its online nature computing each p t must complete within t for some value of Observe that tracking a load s power usage p t also indi rectly reveals when it turns on and off Load tracking targets individual loads and does not attempt a full disaggregation as is common with NILM techniques which try to infer p t for all n building loads such that 1 9 p t P t Further to the best of our knowledge no prior NILM technique ad dresses online operation with a timing constraint Of course perfectly tracking all n loads would be equiv alent to a complete and accurate disaggregation Since load tracking values system performance as well as the accuracy of a load s inferred power readings its goal is to both min imize and maximize accuracy In this case we measure accuracy based on a load s tracking error factor 5 which is simply the error between a load s actual and inferred power usage normalized by its total energy usage If p t denotes load p s actual power usage at time ft and p t denotes its inferred power usage from load tracking at time t then we define the tracking error factor over T intervals as T A 5 E Bi t pilt a X Bilt Here the numerator is the sum of the absolute errors at each data point and the denominator is the load s
9. 25 2 2 Prior Work Our focus on tracking individual loads rather than com plete disaggregation stems from a recognition that i accu rate disaggregation continues to be an elusive goal despite two decades of research and ii the simpler load tracking problem is sufficient for many sensor based applications and can be more efficient and accurate Prior disaggregation ap proaches differ widely based on t s value which ranges from gt 100 000 000 samples per second 21 to one sample per hour 19 Interestingly a recent survey 1 points out that despite t s importance prior work often does not report it In addition despite the plethora of prior work on dis aggregation the same survey 1 highlights the lack of re search that targets second level sampling To the best of our knowledge only Hart s original work 11 and two re cent papers 16 18 which both use an approach based on Factorial Hidden Markov Models FHMMs target data with second level sampling resolution albeit for full disaggrega tion Since there is no prior work on online load tracking we use a FHMM technique modified for online operation as a baseline strawman for comparing PowerPlay s perfor mance and accuracy as described in 6 2 3 Basic Approach PowerPlay employs a model driven approach for load tracking which ensures accuracy and computational effi ciency by decomposing tracking into multiple distinct sub problems Note that p
10. 3 Case Study Demand Response Capacity Lastly we consider a real application of scalable online load tracking where a utility wishes to monitor aggregate de mand response capacity across a neighborhood in real time In this case we assume the utility is only able to reduce de mand by deferring customers A Cs such that the demand response capacity at any point in time is the amount of power consumed by each active A C Thus to estimate demand re sponse capacity over time the utility must know i what percentage of its customers have active A Cs and ii how much power they are consuming We assume a utility server collects smart meter data from each home and runs PowerPlay to track the power usage of customer A Cs For our case study we consider a 10 day period of our deployment home s smart meter data in cluding a central A C To simulate many homes across a neighborhood we generate 100 virtual homes by randomly time shifting the A C s power usage within the smart meter data which results in 100 distinct homes with different time varying A C power usage PowerPlay then uses our model of the A C which includes a mix of the cycle decay and step features to track each home s A C power usage Finally we use PowerPlay s output to query the set of active A Cs 1 2 1 2 pe PowerPlay single device only e PowerPlay all except HRV mmm Pewen ari all circuits m g 1 HMM single device only e 1 HMM Gi ex
11. Oscillations Many non linear devices based on electronic controllers e g microwaves draw a seemingly random amount of power within a fixed range when on We consider bounded power oscillations between maximum and minimum power thresholds as a distinct fea ture resembling a random walk between thresholds Stable Power Oscillations Some non linear loads only have either an upper threshold or a lower threshold resulting in oscillations from a stable power state e g due to the vari able draw of a switched mode power supply Stable power oscillations are a combination of the stable power feature and power spike feature that captures frequent positive or nega tive random fluctuations from a stable power level Power Cycles Many loads include timers that operate them periodically in a repeating pattern e g a dehumidi fier may include a timer that turns it on for two hours out of every four hours A cyclic feature captures the interval and conditions at which the features repeat and potentially their duration e g the length of a stable power level Since essentially every electrical load is either an induc tion motor heating element non linear electronics or some combination thereof every load exhibits one or more of the above features Since the feature set is small we only re quire a small set of detection techniques to identify these features in smart meter data as described in 4 Note that the features above are parame
12. PowerPlay Creating Virtual Power Meters through Online Load Tracking Sean Barker Bowdoin College sbarker bowdoin edu Abstract Online load tracking is the problem of monitoring an in dividual electrical load s energy usage by analyzing a build ing s smart meter data The problem is important since many energy optimizations require fine grained per load en ergy data in real time it also differs from the well studied problem of load disaggregation in that it emphasizes effi cient online operation and per load accuracy rather than accurate disaggregation of every building load via offline analysis In essence tracking a particular load creates a vir tual power meter for it which mimics having a networked connected power meter attached to it To enable high per formance we take a model driven approach that focuses on efficiently detecting a small number of identifiable load fea tures in smart meter data Our results demonstrate that our system called PowerPlay i enables efficient online tracking on low power embedded platforms ii scales to thousands of loads across many buildings on server platforms and iii improves per load accuracy by more than a factor of two compared to a state of the art load disaggregation algorithm Categories and Subject Descriptors H 4 Information Systems Applications Miscella neous J 7 Computer Applications Computers in Other Systems Command and control General Terms Des
13. are presumably caused by other devices As an example we might parameterize a bounded oscillation feature for a par ticular microwave by dictating that at least 50 of reversals over its time window are within a 30W range Thus over an initial 15s window there must be at least 8 reversals to detect the feature at which point the detector extends the window until i the minimum reversal percentage no longer holds or ii a short period passes e g 10s without any rever sals This approach serves to extend the window as long as necessary without overly lengthening the window for long running loads To extract the feature we pair active windows of reversals with matching on and off power steps of the ap proximate expected size for the feature e g 1000W for a particular microwave as illustrated in Figure 4 Growth Decay Detector To detect a decay or growth feature we identify positive steps near a feature s expected magnitude representing possible on events Since the ex pected decay or growth rate specifies a maximum per second negative step for a decay or positive step for a growth the detector then scans forward discarding all changes that exceed the expected maximum The result of this process is a filtered time series that assuming the data actually rep resents a growth or decay should approximately fit an ex ponential or logarithmic curve The detector then performs the standard Levenberg Marquardt Algorithm
14. ative Since the FHMM approach requires a sizable amount of data e g 24 hours for complete disaggregation it cannot operate on a small window size As a result our modified FHMM executes a similar main loop as PowerPlay but al ways disaggregates the most recent 24 hours of data Our example online FHMM incurs an 86 second tracking delay to track the loads in Figure 7 for a single home In contrast PowerPlay imposes only a 5 6 second and 0 6 second delay for the 24 hour and 4 hour tracking windows respectively for the same home We also plot the scalability of each ap proach on a quad core server running at 2 4GHz in Figure 8 where the number of independent homes we track is on the x axis each home is an independent tracking process that runs in parallel We see that the FHMM approach does not operate in real time even tracking loads in a mere 10 homes imposes a tracking delay greater than 10 minutes Power Play performs much better with the same 24 hour time win dow supporting roughly 100 homes with a tracking delay of 2 5 minutes The more realistic scenario with a smaller 4 hour time window scales even better PowerPlay tracks each of the five loads in 1000 homes or 5000 total loads with a tracking delay of only 2 5 minutes Result PowerPlay scales to support online tracking of many homes in this case tracking 5000 loads across 1000 homes with a tracking delay of only 2 5 minutes Finally we also consider PowerPlay
15. cept HRV e 1 HMM all circuits WwW wW wW 5 0 8 5 0 8 5 0 8 S P Ei 0 6 2 0 6 2 0 6 3 0 4 gt 0 4 3 0 4 fos o o LQ 02 LQ 0 2 Q 0 2 I I mo E _ o 4 Toaster Fridge Freezer Dryer HRV Toaster Fridge Freezer Dryer Toaster Fridge Freezer i Device Device Device a Self Input b Aggregate Data minus HRV c Aggregate Data Figure 11 PowerPlay is more robust to noisy smart meter data than the FHMM based approach across homes over time For example at a random point in time 34 of the 100 homes had an active A C with Power Play correctly identifying the status of each A C with 96 accuracy In particular PowerPlay detected 30 out of 34 ac tive A Cs and all inactive A Cs demonstrating 88 recall and 100 precision Of the 30 detected A Cs PowerPlay s second to second inferred power readings differed from the A Cs actual power usage by an average of 104W out of its 3kW peak and 2 6kW average power PowerPlay estimated the total A C power usage across the neighborhood i e its demand response capacity to be 78 1kW which differs from the actual capacity of 87 9kW by 12 with the difference primarily due to the four undetected active A Cs Excluding the undetected A Cs the total A C power inferred by Power Play differed from the actual power by less than 1 Result PowerPlay enables new applications for online an alytics on smart meter data in this case accurate online estimation of the grid s demand re
16. consecutive power steps in the same di rection e g up or down into a single aggregate power step Once collapsed we identify spikes by a large positive step followed immediately by a smaller but still significant neg ative step currently at least 30 of the positive step Im portantly the spike detector separates the spike itself from its load s standard power step feature For example PowerPlay considers the series of changes in power 0 0 500 400 0 0 both a 100W power step feature with a 500W power spike Although the naive step only approach would output a 500W step and a 400W step the spike detector recognizes that this time series most likely represents a 100W inductive load such as a 100W refrigerator Since the magnitude of a spike is highly influenced by when a load turns on within the sampling interval we represent the spike as a binary flag as sociated with the regular power step feature e g the 100W step in our refrigerator example Step Detector While power steps are the simplest fea ture the trivial approach to identifying them detecting second to second deltas of a certain magnitude is often in accurate due to the fact that loads turn on at different points within the sampling interval Thus similar to the spike de tector above we collapse multi second power deltas in the same direction into a single aggregate delta before compar ing the step s magnitude against a specific i e parameter
17. ediment to improving building energy efficiency is that despite much prior research 12 accurate fine grained online monitoring of electrical loads at large scales remains problematic deploying and maintaining large numbers of embedded networked sensors in every building is prohibitively expensive invasive and unreliable Unfortunately timely and accurate knowledge of per load energy usage is a prerequisite for implementing many energy optimization techniques 5 7 22 Rather than rely on expensive instrumentation via em bedded sensors to monitor loads an alternative approach is to analyze electricity data from smart meters to infer a load s energy usage This approach is becoming increasingly attractive since smart meters which monitor an entire build ing s energy usage at small intervals e g minutes to sec onds are now being widely deployed by electrical utilities and consumers 10 In this paper we propose a new analy sis technique which we call online load tracking that mon itors the operation of individual building loads i e when they turn on or off and their fine grained energy usage by analyzing smart meter data In essence tracking a partic ular load creates a virtual power meter for it which mimics having a network connected energy meter attached to it Tracking loads online i e in real time as a smart me ter generates new data is critical since many higher level energy optimization tec
18. eration of the cycle detector As an example consider a refrigerator with a 30 minute period and a magnitude range between 80W and 120W for its spikes at startup Now suppose the detector extracts all spikes due to the refrigerator s compressor from the data and of those spikes each one with a step between 80W and 120W occur at times 0m 20m 30m 55m In this case the detector labels events at 0m 30m and 55m as the on events of the refrigerator while excluding the the event at 20m as it is does not match the expected period While this is a brute force approach the relatively small number of cyclic loads ensures the process is not computationally expensive 5 Implementation We implement PowerPlay s feature detectors and tracking algorithm as a library in Perl The input to the tracking algo rithm is a continuous stream of new smart meter data which PowerPlay buffers while executing its main loop Thus if each iteration of the main loop takes time then the next iteration will consider the set of data points that arrive and are buffered over the previous The tracking algorithm also has as input the set of loads to detect and the corresponding set of identifiable features parameterized separately for each load extracted offline The algorithm then outputs for each load its inferred per second power usage over for each it eration of the main loop resulting in a separate time series of powe
19. fficient online load tracking on a 2 4 GHz single core server PowerPlay is able to track loads in smart meter data comprised of nearly 100 loads in real time each second the same resolution of the building s power meter We also show that PowerPlay improves per load accuracy by more than a factor of two compared to a state of the art disaggregation algorithm based on Factorial Hidden Markov Models FH MMs 16 18 designed for offline analysis 2 Background and Approach PowerPlay assumes a building equipped with a networked power meter that monitors its aggregate electricity usage over time We refer to this building power meter as a smart meter We assume smart homes employ automated energy management techniques which require real time operational knowledge of particular loads energy usage e g air condi tioners A Cs furnaces or other appliances amenable to au tomated energy management Rather than directly monitor ing such loads using sensors our goal is to provide a virtual power meter abstraction that tracks a load s energy usage and when it turns on and off from the home s smart meter data Load tracking is useful in scheduling home loads or pushing alerts to users e g to indicate that a laundry cycle is com plete or when exercising control over large loads such as A Cs across many homes to smooth grid demand 2 1 Problem Statement Formally we define the problem of online tracking for load p as
20. g delay We also observe that the tracking delay effectively varies linearly with the tracking window size As a result shortening the window size linearly decreases the tracking delay In practice most features require significantly less than a 24 hour window to reliably detect HRV x B Dryer messem we s eT Freezer x eo Refrigerator ae O 15 Toaster Q a a D g w K c 1 O wo 0 5 F _ H 5 10 15 20 25 Tracking Window hours Figure 7 PowerPlay s tracking algorithm is efficient with tracking delays of at most a few seconds Result PowerPlay is able to track multiple loads in real time or near real time on commodity servers We also compare PowerPlay s scalability with a com plete disaggregation algorithm based on FHMMs Here we assume a server must track loads across many homes not just a single home We quantify both PowerPlay s perfor mance with 24 hour and 4 hour tracking windows and an FHMM approach following 18 Since disaggregation using the FHMM is exponential in the number of building power states which is based on the number of loads and the num ber of power states per load the FHMM approach models each load as having only four power states and disaggregates at the level of circuits rather than individual loads Since our home has only 25 circuits but operates 92 individual loads our FHMM performance numbers for a complete disaggre gation are conserv
21. hniques require such real time data For example an automated load scheduling policy that re duces a building s peak power demand by deferring one or more background loads must know the energy usage of each background load to determine which of them to defer and for how long 5 As another example a recommendation engine may monitor the energy usage of a building s interac tive loads to push energy efficiency recommendations to oc cupants smartphones in real time directing them to take an immediate action to better optimize their energy usage e g We use the term electrical load or simply load to refer to any distinct appliance or device that consumes electricity such as turning off an idle coffee pot 2 Essentially on line load tracking is useful for any application that requires attaching a power meter to a load that transmits its average power usage every pre specified time interval in real time Our work builds on prior work which has already de veloped a variety of analysis techniques for smart meter data including load disaggregation 1 11 18 26 and occu pancy detection 8 Many startup companies are now com bining such energy based analytics with cloud based big data platforms 6 to mine building smart meter data en masse However we argue that online load tracking dif fers from the well studied problem of complete load dis aggregation often termed Non Intrusive Load Monitoring NILM 1 11
22. ign Measurement Performance Keywords Load Monitoring Load Modeling Smart Grid This work was done while the author was at the University of Mas sachusetts Amherst Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed or profit or commercial advantage and that copies bear this notice and the full citation on the first page Copyrights for components of this work owned by others than ACM must be honored Abstracting with credit is permitted To copy otherwise or republish to post on servers or to redistribute to lists requires prior specific permission and or a fee Request permissions from Permissions acm org BuildSys 14 November 5 6 2014 Memphis TN USA Copyright is held by the owner author s Publication rights licensed to ACM ACM 978 1 4503 3143 2 14 11 15 00 http dx doi org 10 1145 2674061 2674068 Sandeep Kalra David Irwin and Prashant Shenoy University of Massachusetts Amherst skalra shenoy cs umass edu irwin ecs umass edu 1 Introduction Collectively buildings consume significantly more energy 41 than society s other broad sectors of consumption industry 30 and transportation 29 14 As a result the design of smart buildings that are capable of automatically regulating their energy usage has become an important research area However one continuing imp
23. ly receiving power readings each second and executing its main loop to perform feature detection on the most recent window of data Since PowerPlay stores recent data in memory I O overhead is negligible and efficiency is solely a result of the computa tional overhead of the feature detectors The tracking efficiency of PowerPlay is determined by the computation overhead of the feature detectors in processing the most recent window of data This overhead determines both a the tracking delay from 2 of the system where e second is perfect 1 Hz real time tracking and b the number of loads and homes that a platform can effectively track Note that since PowerPlay s main loop detects fea tures across all loads increasing the number of loads ig noring parallelism increases tracking delay across all loads Thus we measure the aggregate number of loads PowerPlay can track while maintaining a low tracking delay We perform the following experiments on a single core server running Ubuntu Linux kernel version 3 2 0 with a 2 4GHz Xeon processor We vary a common window size across all features then observe the tracking delay achieved by PowerPlay As seen in Figure 7 the tracking de lay is modest across every load For example with an exces sively long tracking window of 24 hours PowerPlay com pletes in less than 3 seconds per load As expected loads with more features e g the dryer result in a longer track in
24. nt sliding window of time series data from a smart meter to determine whether it is embedded in the data Matching typically involves computing a time series distance function such as Euclidean distance or Dynamic Time Warping 15 between the load s raw power usage and the most recent set of smart meter readings of equal size a match then occurs when the distance is less than a pre defined threshold Low level time series matching is more expensive and less robust than using higher level features for load tracking 3 Offline Feature Identification and Selection We first describe the three offline steps in PowerPlay s ap proach namely modeling a load extracting a load s features and then selecting a subset of identifiable features to track As this process is a one time step we envision manufactur ers profiling each load and supplying its model and features as part of its technical user manual The information could also be crowd sourced such as in The Power Consumption Database which already provides crowd sourced informa tion on maximum and idle power for a wide range of loads indexed by type manufacturer and model number 24 3 1 Modeling and Feature Extraction Electrical loads in an alternating current AC system fall into one of four basic types tesistive inductive capacitive or non linear Informally resistive loads include heating el ements such as a toaster inductive loads include AC mo tors s
25. nt annex pdf 26 M Zeifman and K Roth Nonintrusive Appliance Load Monitoring Review and Outlook IEEE Transactions on Consumer Electronics 57 1 February 2011
26. o higher when including its full set of identifiable features compared to re stricting it to only step features However beyond a com plexity of 1000 power deltas the error factors stay roughly constant with the exception of the refrigerator even when the complexity goes to 50 000 power deltas The refriger ator s accuracy decreases significantly when adding a com plex load e g in this case a heat recover ventilator that ex hibits stable power oscillations The reason is that its cycle detector is unable to select spikes that correspond to the re frigerator due to the heat recovery ventilator generating a large number of similarly sized spikes at various intervals Figure 11 then examines three specific points from the previous graph and compares them with the FHMM ap proach In Figure 11 a we use both PowerPlay and the FHMM approach to track a load from data that only includes that load As shown the FHMM approach is nearly perfect since its model is trained on the actual data we disaggregate in this case By comparison PowerPlay shows some error due to the fact that our models while accurate only include offline features and not attributes based on when and how long the load operates However Figure 11 b and c shows the error factor for the same loads if we include every cir cuit both with b and without c the complex heat recovery ventilator Prior work on load disaggregation has generally evaluated their algori
27. of complete disaggregation via a FHMM Unfortunately inferring energy usage over a long period is not appropriate for online operation and does not take into account when a load uses energy We use the tracking error factor 5 from 2 to quantify per load accuracy over time In Figure 10 we first quantify accuracy as we scale up the num ber of non tracked loads in a home since more loads result in more and less visible features In this case the x axis is a rough measure of the data s complexity i e the number of power deltas gt 15W By gradually adding circuits from our home deployment to the smart meter data For example the far left side of the graph includes only one circuit the one in cluding the corresponding tracked load and each data point to the right represents a dataset with one more circuit added to it For each new circuit we track the loads and compute the error factor per load on the new dataset Figure 10 plots the results for our representative loads Note that the x axis is on a log scale since a small number of loads contribute the majority of the power deltas For comparison we also include a second model of the freezer that only uses step features to illustrate the effect of removing all but the most trivial features present in PowerPlay As expected the error factors increase as we add more cir cuits and more complexity to a home s data We also see that the freezer s accuracy is nearly a factor of tw
28. ore easily and accu rately detect the remaining features as the residual filtered data has less noise after filtering After filtering PowerPlay applies the remaining basic feature detectors e g spikes growth decays and steps to identify and label those features in the data Finally PowerPlay runs the cycle feature detector over the list of labeled features to identify repeating patterns of features the cycle feature detector is unique in that its input is a set of labeled features rather than raw time series data and as such is run last For each desired virtual power meter i e load in the tracking set PowerPlay then examines the list of labeled but unassigned features found in the recent past over a win dow W If the identifiable features of the load are found in the window it assigns these features to the load and declares a load match Upon assigning features to a load PowerPlay removes them from the list of unassigned features For com posite loads the set of features over window W may need to occur in a certain order or within a certain time inter val to infer a load s presence Finally whenever PowerPlay detects a load based on its features it updates the load s in ferred power usage p t using the filtered feature data and the load s model which captures the load s full power usage behavior j c Microwave 3 300 p 250 a 200 pe 180 160 140 120 100 80 y 60 40
29. r data for each load in the tracking set We deploy PowerPlay in a real home We describe the home its loads and our instrumentation in prior work 4 Briefly the home includes a Internet enabled power me ter installed in its electrical panel to monitor the second to second power usage of the home and each of its cir cuits There are multitude of such meters now available both commercially 23 and in recent research 17 that record home level and circuit level data at 1Hz sampling resolution We also record ground truth power data or on off events which we correlate with the power meter data for individ ual loads not connected to dedicated circuits using either Z Wave Smart Energy Switches Insteon iMeters or Insteon SwitchLincs In total our deployment includes 92 sensors producing roughly four million data points per day Such an extensive deployment is necessary to compare our results based the home s power data with ground truth power data from each individual load Of course since our offline modeling and feature extrac tion methodology is new to this paper we must manually model each load we track and extract its important features ourselves However our hope is that by demonstrating the usefulness of our models in analysis we will motivate man ufacturers to use our methodology to derive models and ex tract features as part of a load s design and publicly release them We also plan to release the models and ex
30. range and flagged deltas comprising the feature To filter the feature from the raw data we remove from the data any oscillations that do not result in an update to p and then use them to reconstruct the feature s second to second energy usage due to its stable oscillation behavior as illustrated in Figure 3 In determining the D parameter for each load the goal is to set it long enough to ensure changes in power are not random oscillations due to some other load but short enough to prevent filtering short lived loads For T the goal is to select a value large enough to capture the expected oscillations without attributing the power usage of unrelated background loads to the feature Bounded Oscillation Detector The bounded oscillation detector examines data for groups of deltas within a cer tain range that reverse themselves change from positive to negative frequently within a given minimum window size e g 60 seconds In particular the detector looks for a minimum proportion of reversals within the window e g 50 extending the window size until the minimum pro portion is not met or several seconds have passed without a a c Reconstruct a Reversals b On off Pairing Figure 4 Example of bounded oscillation detector reversal i e power use has stabilized indicating the device is off Within the resulting window power deltas exceeding the bounded power range are filtered out as these changes
31. rior work on complete load disaggre gation typically conflates these subproblems The subprob lems include i empirically modeling a load ii extracting features from the model iii selecting the most identifiable features and finally iv detecting and tracking a load based on these features Figure 1 depicts the basic workflow of each subproblem which we in turn outline briefly below Smart Meter Offline Processing Online Processing pE A i od E i o Menn l i e Tol l l i z Basic Load Parameterized I Feature i i Models Appliance Models Detectors d L Feature Identifiable Load Extraction ER L Appliance Features Tracking Activity Figure 1 PowerPlay uses offline modeling and feature extraction for online load tracking 1 Empirical Modeling We first empirically model each load s energy usage based on properties of the four basic types of electrical loads i e resistive inductive capaci tive and non linear Prior work describes how to derive such models and shows that such empirical models accu rately capture the behavior of nearly every common house hold load 3 We assume a load s model accurately de scribes its energy usage when on 2 Feature Extraction After empirically modeling a load we decompose it into a set of features Each feature captures a subset of the load s pattern of energy u
32. rker S Kalra D Irwin and P Shenoy Empirical Characteri zation and Modeling of Electrical Loads in Smart Homes In JGCC June 2013 4 S Barker A Mishra D Irwin E Cecchet P Shenoy and J Albrecht Smart An Open Data Set and Tools for Enabling Research in Sus tainable Homes In SustKDD August 2012 5 S Barker A Mishra D Irwin P Shenoy and J Albrecht Smart Cap Flattening Peak Electricity Demand in Smart Homes In Per Com March 2012 6 Bidgely http bidgely com 7 T Carpenter S Singla P Azimzadeh and S Keshav The Impact of Electricity Pricing Schemes on Storage Adoption in Ontario In e Energy May 2012 8 D Chen S Barker A Subbaswamy D Irwin and P Shenoy Non Intrusive Occupancy Monitoring using Smart Meters In BuildSys November 2013 9 eGauge Energy Monitoring Solutions http egauge net 10 U S Energy Information Administration Frequently Asked Ques tions How Many Smart Meters are Installed in the U S and who has them http www eia gov tools faqs faq cfm id 108 6t 3 11 G Hart Nonintrusive Appliance Load Monitoring IEEE 80 12 December 1992 12 T Hnat V Srinivasan J Lu T Sookoor R Dawson J Stankovic and K Whitehouse The Hitchhiker s Guide to Successful Residential Sensing Deployments In SenSys November 2011 13 D Kelly Disaggregating Smart Meter Readings using Device Signa tures In Masters Thesis Imperial College London
33. s performance on embedded platforms that track a set of loads within a home such as in an embedded energy monitoring and analytics Q e fo HMM disaggregation PowerPlay 24 hrs PowerPlay 4 hr a e A w e Dey i Tracking Delay s e 0 200 400 600 800 1000 Number of Homes Figure 8 PowerPlay efficiency enables it to scale to many homes while maintaining a low tracking delay 14 Ground Truth m 12 p PowerPlay 10 HMM 8 L i B Toaster Fridge Freezer Dryer Device Figure 9 Both PowerPlay and the FHMM approach ac curately assign the energy used by loads each day O N AO Assigned Energy kWh platform 17 To evaluate this case we deploy PowerPlay on a low power DreamPlug computer with a 1 2GHz ARM pro cessor and 512MB memory costing less than 100 Tracking the same five loads as above in our deployment home with a 4 hour tracking window PowerPlay achieves a tracking de lay of just 18 seconds with individual load tracking times ranging from less than a second for the refrigerator to four seconds for the toaster Result PowerPlay is capable of online tracking of loads within a home on low power embedded platforms 6 2 Tracking Accuracy In addition to efficiency load tracking must also be ac curate to be useful As before we compare PowerPlay s accuracy in tracking multiple loads real time power usage wi
34. sage within the model the set of features collectively represents a con cise description of how the load s operation manifests itself in power data Intuitively a load tracking algorithm must search for these features within a home s aggregate smart meter data to detect the presence of the load and track it 3 Identifiable Feature Selection PowerPlay optimizes load tracking efficiency by distilling a load s full feature set into a subset of its most identifiable features Identifiable fea tures are a load s most prominent and unique features such that a tracking algorithm need only search for these identifi able features rather than the full feature set to detect and track a load with high confidence Clearly the smaller the set of identifiable features the more efficient online detec tion 4 Online Load Tracking The final step is to design a tracking algorithm that detects a load s identifiable features in the smart meter data in an online fashion The first three steps above namely empirical modeling feature extraction and identifiable feature selection are one time tasks performed offline while PowerPlay s final detec tion and tracking step is continuous and online PowerPlay s model based feature driven tracking differs from low level time series matching 13 In essence the time series approach takes either a trace or model of a load s raw power usage when on and matches it against a rece
35. sponse capacity 7 Conclusions This paper presents PowerPlay a system for online load tracking that emphasizes both efficiency and accuracy In essence tracking a particular load creates a virtual power meter for it which mimics having a network connected en ergy meter attached to it PowerPlay takes a model driven approach to online load tracking which focuses on detect ing a small number of identifiable load features in smart me ter data This paper enumerates an identifiable set of fea tures common across loads and then designs methods to ef ficiently detect them in smart meter data By using a high level feature abstraction PowerPlay enhances computational tractability enabling efficient and accurate online load track ing Our results show that PowerPlay is able to track loads in near real time even on low power embedded platforms and improves per load accuracy by a factor of two compared to a FHMM based disaggregation algorithm 8 Acknowledgments This research was supported by NSF grants CNS 1405826 CNS 1253063 CNS 1143655 CNS 0916577 and a grant from the Massachusetts Department of Energy Re sources 9 References 1 K Armel A Gupta G Shrimali and A Albert Is Disaggregation the Holy Grail of Energy Efficiency the Case of Electricity Energy Policy 52 1 January 2013 2 N Banerjee S Rollins and K Moran Automating Energy Manage ment in Green Homes In HomeNets August 2011 3 S Ba
36. t distinctive The first set contains noisy features namely all stable and bounded power oscillation features across all loads in the tracking set The second set contains the remaining basic features steps spikes and decay growth features across all loads The final set contains any cycle features for loads in the tracking set Given these ordered sets the tracking algo rithm then repeatedly executes its main loop which applies every feature detector from all loads in order as described in 4 2 Note that PowerPlay buffers any smart meter data that arrives while executing its main loop and reads and ap pends it to the home s power data time series on the loop s next iteration The time taken to complete the main loop de fines PowerPlay s online performance i e the minimum it can support For example if the main loop takes 30 sec onds to complete then the tracking algorithm can only out put each load s inferred power usage every 30 seconds The exact value of depends on available hardware resources as well as the number of virtual power meters to simulate i e twice as many tracked loads will increase by roughly 2X PowerPlay first detects the noisy features i e those that contain significant power fluctuations These features are detected labeled and filtered from the home s power data as described in 4 2 Detection and filtering of noisy features first enables PowerPlay to m
37. terized for each specific load e g the magnitude of a step or the rate of a decay and may differ across two loads of the same type e g two A Cs from different manufacturers may require different features and parameters Thus PowerPlay s offline component not only extracts the features of a load but also determines the parameters for each feature Figure 2 includes annotated fea tures in power usage data for a variety of common loads 3 2 Selecting Identifiable Features Since basic loads only include a few features an online load tracking algorithm can use all of their features to detect their presence However complex loads such as a wash ing machine may exhibit an excessively large number of features Fortunately searching for every feature is gener ally not necessary for accurate detection it is often sufficient to select a subset of prominent features to uniquely identify the load PowerPlay leverages this insight to only search for a small set of identifiable features to match complex loads which improves both efficiency and scale Selecting identifiable features for a load is a one time of fline task and presents a tradeoff between accuracy and per formance A smaller set of identifiable features improves the efficiency of detection but decreases tracking s accuracy At present we construct a complex load s set of identifiable features experimentally by iteratively adding the next high est magnitude features e
38. th the FHMM approach which performs a complete dis aggregation We take the conservative approach of training the FHMM on per load data from the home that we disag gregate although doing so is often not possible in practice since disaggregation is typically only useful in homes where such training data is not available As disaggregation often focuses on inferring a breakdown of per load energy usage for a building over a long time period e g an entire day or week Figure 9 shows the actual energy usage over an en tire day for five loads as well as the inferred energy usage from both PowerPlay and the FHMM disaggregation We see that both PowerPlay and FHMM accurately predict each load s energy usage over long periods of time although the FHMM approach is less accurate for the heat recovery ven tilator due to its stable power oscillations Our results are consistent with prior work on the FHMM approach which performs as well or better than other prior approaches to freezer edges only fridge u 08 freezer O HRV dryer W o6 toaster D amp S 047 AE S Tua E oz ne pamete 0 a Se eer j 100 1000 10000 100000 Trace Complexity of steps Figure 10 PowerPlay error factors when scaling up to highly noisy and complex smart meter data disaggregation 16 18 Result The accuracy of PowerPlay s inferred energy usage for loads in the tracking set over long periods is comparable to that
39. thms at small scales e g 5 10 indi vidual loads that are not representative of the multitude of small and complex loads present in a modern home Our re sults demonstrate that PowerPlay performs well even as the number and complexity of loads scales up The result shows that PowerPlay is significantly more ac curate than the FHMM approach for each load with the ex ception of the clothes dryer While PowerPlay is not more accurate than the FHMM approach at small scales as in a with less noisy data it is significantly more accurate as complexity increases For example PowerPlay is nearly per fect at detecting the second to second power usage of the toaster even within a highly complex trace largely due to PowerPlay s highly accurate model of the toaster as shown in Figure 11 a In general the improvement in error fac tor for each load over the FHMM approach is greater than 2X and over 100X in the case of the toaster Both Power Play and the FHMM approach perform well on the clothes dryer because it is large compared to the other loads 6kW peak power versus 1kW peak power such that the added complexity does not affect detection Result PowerPlay maintains a low per load tracking er ror factor as the number of loads and their complexity in creases ina home For the loads in our tracking set the error factor is generally a factor of two less than a state of the art disaggregation algorithm based on FHMMs 6
40. total en 1 ergy usage over T Lower values of 6 are better an error fac tor of zero indicates perfect tracking While there is no upper bound on the tracking error factor an error factor of one indi cates that the reading to reading errors are equal to the load s energy usage In general a tracking error factor near one is not considered good since simply inferring a load s energy usage to be zero at each time results in 1 Note that this metric is a load specific variant of the total energy correctly assigned metric from prior work 18 We denote the meter s data resolution using the sampling time interval t A coarser or longer sampling interval averages out features in P t eliminating identifiable at tributes while a finer or shorter interval reveals more at tributes but also more data to process as well as more noise Our work specifically targets consumer grade power meters such as the TED 23 eGauge 9 and BrulTech which com monly provide a sampling resolution of one reading per sec ond e g T 1 second While today s utility grade smart me ters provide at most minute level sampling e g a reading once every five to fifteen minutes is common there are in dications the next generation of meters will provide second level sampling For example a U K subcommittee defining future smart meter specifications recently released a report advocating a five second sampling resolution
41. tracted fea tures for the loads that we track in 6 6 Evaluation We evaluate the accuracy and efficiency of PowerPlay s online load tracking algorithm in our home deployment We first measure the computational overhead of load tracking to quantify PowerPlay s efficiency which enables it to either track loads on low power embedded platforms or scale to thousands of loads across many homes on server platforms We then evaluate PowerPlay s accuracy by quantifying the tracking error factor 6 for various loads In both cases since there is no prior work on load tracking we compare Power Play to a complete disaggregation algorithm based on FH MMs modified for online operation In this case we use the same approach as Kolter and Johnson 18 to evaluate their Reference Energy Disaggregation Dataset REDD which is similar to the technique by Kim et al 16 Since PowerPlay relies on load models computed offline we manually model a representative set of loads in our de ployment home that collectively cover each feature type The set includes a toaster oven steps decays a refrigerator and freezer steps spikes cycles a heat recovery ventilator or HRV stable oscillations and a dryer bounded oscillations cycles steps decays PowerPlay then tracks these loads in real time using per second power data for the entire home which operates 92 distinct loads 6 1 Tracking Efficiency PowerPlay operates online by continuous
42. uch as fans or compressors and non linear loads in clude any type of electronic device such as TVs or comput ers Loads behaves differently based on their load type but devices of the same type exhibit many common behaviors Complex appliances that operate multiple internal loads e g a refrigerator with a motor based compressor and interior light bulb exhibit a composition of these behaviors Further details of how the four basic types map to real world devices are provided in 3 Below we enumerate the identifiable features that PowerPlay tracks Stable Power Steps The simplest feature is a discrete change in average power from one stable value to another stable value Most disaggregation algorithms that analyze real power data e g at sampling resolutions coarser than 60Hz in the U S consider stable power steps as the only identifiable feature In reality only a few low power resistive loads such as incandescent lights exhibit only these simple steps when on Power Growth Decay and Spikes Many loads expe rience smooth increases or decreases in power when turned on e g due to decreasing resistance as a heating element warms or abrupt and sudden spikes in power e g when starting an induction motor We consider power growths decays and spikes as distinct features spikes capture an ini tial power surge while logarithmic growths and exponential decays capture gradual increases or decreases in power Bounded Power
Download Pdf Manuals
Related Search