1、Jose Valenzuela del Rio is an engineer in Siemens Energy, Orlando, Florida. Yanal Issac and Adam Coulon are graduate students in the Department of AerospaceEngineering at the Georgia Institute of Technology, Atlanta, Georgia. Scott Duncan is a research engineer in the Department of Aerospace Enginee
2、ring at theGeorgia Institute of Technology, Atlanta, Georgia. Dimitri Mavris is a professor in the Department of Aerospace Engineering at the Georgia Institute ofTechnology, Atlanta, Georgia.Chilled Water System AbnormalityDetection with Machine LearningAlgorithmsJos Valenzuela del Ro, PhD Yanal Iss
3、ac Adam CoulonScott Duncan, PhD Dimitri Mavris, PhDABSTRACTThis paper applies machine learning (ML) algorithms to detect abnormalities in chilled water systems (CWS) at building level. The identificationof two abnormal situations are pursued: inaccurate chilled water sensor readings and low thermal
4、efficiency in terms of T. The visualization of buildingchilled water data provides general trends and an initial identification of the building abnormalities; this visualization also helps to lay down therequirements for the abnormality detection algorithms, and eventually, their selection. Abnormal
5、 performance is flagged by these algorithms andabnormality indices are assessed with respect to two baselines: normal operation and the best operation found (in terms of T). Two contexts ofabnormality detection and quantification are presented. First, historical data is searched for normal and abnor
6、mal clusters. The abnormality indices ofcluster frequency, thermal efficiency and power loss, all with respect to the normal operation baselines, are calculated. The second context is theabnormality detection of real-time data, where abnormality indices thermal efficiency and power loss assessment a
7、re assessed with respect to the bestoperation found in the normal operation cluster.INTRODUCTIONTechniques in machine learning have shown promising results in automated knowledge discovery making itmore and more crucial when large data is at hand. The Georgia Tech Facilities Management group has bee
8、ncontinuously recording chilled water energy meter data for several years, resulting in a large amount of disparate dataover an extended period of time, which makes it very difficult to manage and manually analyze. Systems that employwater chillers are commonly called chilled water systems (CWS), tr
9、ansporting cooling fluid to load terminals and backto the chillers (Air 2001). Two major CWS issues are low values of T and sensor malfunction. T is the temperaturedifference between the chilled water temperature returning from a building and the chilled water temperature suppliedto a building. T th
10、at is too low leads to increased pump energy usage and either an increase in chilled energy usage ora failure to meet cooling load (Taylor 2002).Fortunately, recent advancements in the field of machine learning show promising results in detecting sensormalfunction. Moreover, research applying machin
11、e learning to different aspects of chilled water systems showpromising results. Yun and Won (2012) propose a new HVAC control strategy for energy systems using machinelearning to provide consistent comfortable working conditions based on temperature and humidity.Traditionally, machine learning can b
12、e divided into two broad groups, supervised learning and unsupervisedlearning. In supervised learning, the data is labeled; however, in unsupervised learning the data is not labeled. One ofthe primary goals of unsupervised learning is to discover any hidden structure within the data, this is known a
13、sclustering, or to determine the distribution of data within the input space, known as density estimation (Bishop 2006).Current research in machine learning for classification and outlier and novelty detection has utilized techniquesrelying on support vector machines (SVM). For example, Manevtiz and
14、 Yousef (2001) implemented one class SVMfor information retrieval and document classification with promising results. Support vector data description is usedto determine the boundary around a data set, enabling the detection of outliers and novelties (Ghahramani 2004). Anexample of algorithms for cl
15、ustering large spatial databases with minimal knowledge is DBSCAN (Bie2009).Chilled water system energy data is merely the recording of sensory measurements with no label to indicateperformance of the campus subsystems, so the data in this paper is deemed an unlabeled data set. Using unsupervisedlea
16、rning techniques, the data is clustered in to different groups. Once a cluster is identified, it is then investigated andlabeled as a cluster representing normal or abnormal operation. Using this approach, it is possible to analyze large datasets, for the many buildings on campus. This enables the F
17、acilities department to quickly assess and react tosubsystems performing poorly, therefore saving energy and cost. By using outlier and novelty detection based ondensity estimation it is possible to detect sensory malfunction and by clustering the data it is possible to explore anyhidden structure w
18、ithin the data revealing different operation modes of the campus subsystems.METHODOLOGYThe purpose of this paper is to learn insights and automatically identify abnormality scenarios in the GeorgiaTech (GT) campus at the building level. The targeted abnormalities are sensor malfunctioning detection
19、and low T.First, manual data visualization is used to understand the general trends, detect chilled water system (CWS)abnormalities of several campus buildings, and define technology requirements to detect these abnormalities in anautomatic fashion. Next, machine learning (ML) algorithms are present
20、ed. It is followed by the application of the MLalgorithms to CWS abnormality detection in GT buildings and discussion of results. Finally, the main conclusions ofthe work and future work are drawn.Visualization of Chilled Water DataThe first step in detecting abnormalities at building level is to vi
21、sualize the data to provide general trends,preliminary visual classification, and abnormal situations in the CWS at the building level. Data visualizations areperformed by correlations between building CWS variables. It enables learning about the relationship betweenvariables, the identification of
22、clusters within the whole data, and abnormality detection. The historical data collectedin this study ranges from 01-10-2013 to 09-24-2013. Erroneous sensory data has been removed from the data set.Once the erroneous data is removed, the correlation between chilled water variables in Buildings A and
23、 C areshown in Figs. 1 and 2, respectively, which provide evidence that there is a strong positive relation between the energyflow and T. As the building energy demand increases, a greater amount of heat is removed by the CWS. Anothergeneral trend depicted in Figs. 1 and 2 is the positive correlatio
24、n between the cooling flow and the supplied energy. Itimplies that more cooling flow is the consequence of an increasing demand of energy in the building, which is anintuitive and reasonable behavior. However, building C does not experience this correlation, see Fig. 2.Inspection of Figs. 1 and 2 al
25、so indicates that differences in the outside temperature, Tout, provoke operationalchanges and different kinds of CWS abnormalities, as discussed in subsequent sections.Influence of Outside Temperature. Three ranges of Tout are studied: Tout,1 =, 281.15 K (, 46.4 F);Tout,2 = (281.15, 297.15) K (46.4
26、, 75.2) F); and Tout,3 = 297.15, K (75.2, F). A color code is used in Figs. 1and 2 to indicate that a point membership: red for Tout,3, blue for Tout,1, and grey Tout,2. Figures 1 and 2 indicate adifferent data pattern for each Tout range, thus there is a different operation depending on the Tout, a
27、nd therefore theseason.Chilled Water System Abnormalities. Two main abnormalities are found in the correlation plots: high and lowT operation and possible sensor misreadings.Figure 1 Visualization of meter data history showing correlations between three variables for Building AFigure 2 Visualization
28、 of meter data history showing correlations between three variables for Building CContexts, Baselines, and Abnormality IndicesOne of the objectives of this paper is to detect abnormalities (particularly, low T at building level and sensormisreadings). These abnormalities are desired to be quantified
29、 in terms of abnormality indices such as frequency ofoccurrence with respect to the normal operation, thermal efficiency and power loss. In order to quantify the last twoaspects, baselines must be defined. These baselines are assessed by finding the normal and/or best operation ofavailable building
30、chilled water data. Thus, only inefficient system behavior with respect to its normal and/or bestoperation can be inferred.An average T of the building normal operation can provide the building thermal efficiency. Then, systemarchitectural changes and new control approaches can increase the overall
31、system efficiency. However, this paper isonly interested in the detection of abnormalities with respect to the building current normal or best operation.The abnormality detection and quantification is done in terms of two contexts:Historical Data. The identification of abnormal building behavior is
32、pursued. This context can provide buildingmaintenance prioritization in terms of the abnormality indices. Preliminary building chilled water observations showthat clusters of normal and abnormal operation exist, see Figs. 1 and 2.Real-time Data. The real-time operation of the building is monitored.
33、Thus, campus diagnosis is provided.Abnormal detection and quantification of real-time data is assessed by comparing with baselines obtained from thehistorical data. Two baselines are employed, one for each of the contexts for abnormal detection:Normal Operation. It reflects the most common operation
34、 of the building CWS. Therefore, this baseline helpsto identify historical abnormal operation in the building CWS and give awareness and quantify the disadvantages ofthe corresponding abnormal historical data group. This baseline is obtained from the historical data.Best Operation. It is characteriz
35、ed by the most thermally efficient operations of the normal operation baselineof the building CWS. Therefore, building operations with efficient Ts are employed to obtain this best operationbaseline. However, preliminary visualization showed that best seen Ts vary with the Tout and the building dema
36、nd.Thus, for a given Tout and building energetic needs, the building chilled water data points of the normal operationbaseline with T the closest to 6.672 K (12F) (note it is a temperature difference) are selected as the best operationbaseline. Since only normal operation baseline with low T are cho
37、sen, the best operation baseline can be written as:(1)where Best and Norm are the baselines for the best and normal operation, respectively. This baseline is employedfor the identification of real-time abnormal data.Once the baselines are available for comparison, abnormality indices can be evaluate
38、d. The followingabnormality indices are explored:Frequency of abnormality with respect to normal operation. Since the campus acquisition system providesdata every 15 minutes, the number of points of each abnormal cluster indicates the frequency of the abnormaloperation related to the cluster.(2)wher
39、e Hist is the historical data-set, ClAb represents the abnormal cluster, and 1ClAb the indicator function of theset related to the abnormal cluster ClAb. This index is used only in the historical data because real-time data onlysnaps the current operation.Thermal efficiency index. This index provide
40、s information about how well the heat is transferred at buildinglevel. T is employed to obtain this index. In the historical context, this index is assessed in terms of the differencebetween the average T of the normal operation baseline and the average T of the abnormal clusters ClAb.(3)where NA is
41、 the number of points in the set or cluster A. Notice that since this index is targeted to low Toperation, it is expected that the normal cluster has a higher average T than that of the abnormal operation clusterClAb.Regarding the real-time operation this index is assessed by comparing the real-time
42、 T with the best T found inthe normal operation baseline of the historical dataset for the equivalent operation(4)where is the best T found at an equivalent operation to the real-time state xrt; i.e. same suppliedenergy Eprov and range of T0. A negative value of TEI implies better performance than t
43、he best operation baseline andthe large positive values imply much worse thermal efficient than the best operation baseline.Power loss index. The extra pump power consumption can be also utilized as an indicator of a poor use of theCWS when T is low. The identification of devices with high power con
44、sumption is interesting in both contexts. In afirst approximation, the pump power consumption is related to the cubic power of the mass flow rate, this power losscan be eventually translated into money loss. For the historical context it could be written as the difference betweenthe average mass flo
45、w rate of the normal operation baseline and the average mass flow rate of the abnormal clusters(5)Notice that since this index is targeted to low T operation, it is expected that the normal cluster has a smallermass flow rate than that of the abnormal operation cluster.For the real-time context, thi
46、s power loss index could be written as(6)where is the best found at an equivalent operation to the real-time state xrt; i.e. same suppliedenergy Eprov and range of T0. A negative value of PLI implies better performance than the best operation baseline andthe large positive values imply much higher p
47、ower loss than the best operation baseline.Machine Learning AlgorithmsThis investigation pursues the identification of regular operations (normal and best operation baselines) andabnormal operations. These operations are found through clustering and classification techniques.Selection of Relevant Fe
48、atures. In Figures 1 and 2 exhibit CWS operations for two studied buildings; theiroperations clearly depend on the Tout. Due to these different normal operations at different Tout, the data is dividedinto groups with several ranges of Tout. Now, the purpose is to automatically identify seasonal norm
49、al operation (notmanual clustering), best operation and abnormal operation (unsupervised learning), and based on the normaloperation found classify real-time points (supervised learning). In order to do it, the selection of the features forclustering operational groups is crucial.The main key variables to detect low T operation it are mass flow rate (mdot), Eprov, and T. T is highlycorrelated to the energy supply (measured in tons of cooling). Besides, it is well known the relationship between thesetwo variables with the cooling flow:(7)Becau
