Managing Coplex Network Operation with Predictive Analytics Zhenyu Huang, Pak Chung Wong, Patrick Mackey, Yousu Chen, Jian Ma, Kevin Schneider, and Frank L. Greitzer Pacific Northwest National Laboratory 90 Battelle Blvd, Richland, Washington, USA 995 {zhenyu.huang pak.wong patrick.ackey yousu.chen jian.a kevin.schneider frank.greitzer}@pnl.gov Abstract Coplex networks play an iportant role in odern societies. Their failures, such as power grid blackouts, would lead to significant disruption of people s lives, industry and coercial activities, and result in assive econoic losses. Reliable operation of these coplex networks is an extreely challenging task. None of the coplex network operations are fully autoated; huan-inthe-loop operation is critical. Given the coplexity involved, there ay be thousands of possible topological configurations at any given tie. During an eergency, it is not uncoon for huan operators to consider thousands of possible configurations in near real-tie to choose the best option and operate the network effectively. In today s practice, network operation is largely based on experience with very liited real-tie decision support, resulting in inadequate anageent of coplex predictions and the inability to anticipate, recognize, and respond to situations caused by huan errors, natural disasters, or cyber attacks. A systeatic approach is needed to anage the coplex operational paradigs and choose the best option in a nearreal-tie anner. This paper applies predictive analytics techniques to establish a decision support syste for coplex network operation anageent and help operators predict potential network failures and adapt the network in response to adverse situations. The resultant decision support syste enables continuous onitoring of network perforance and turns large aounts of data into actionable inforation. This paper presents exaples with actual power grid data to deonstrate the capability of a proposed decision support syste. Introduction Electric power grids, gas pipeline systes, telecounication systes, and aviation networks are just a few exaples of coplex networks that are iportant in odern society. Their failure, such as power grid blackouts, would lead to significant disruption of peoples lives, industry and coercial activities, and result in assive econoic losses (DOE 00). Operation of these coplex networks is an extreely challenging task as they all have coplex structures, wide geographical coverage, Copyright 009, Association for the Advanceent of Artificial Intelligence (www.aaai.org). All rights reserved. and coplex data/inforation technology systes. The coplex networks also exhibit highly dynaic and nonlinear behaviors with nuerous network configurations and are affected by a nuber of external factors. The external factors include physical attacks, cyber threats, huan errors, and natural disasters. None of the coplex network operations are fully autoated; huan-in-the-loop operation is critical. During an eergency, it is not uncoon for huan operators to consider thousands of possible configurations in near real-tie to choose the best option and operate the network effectively. In today s practice, network operation is largely based on operator s experience with very liited real-tie decision support. Because of the coplex nature of large networks (e.g. coplex structure and wide geographical coverage), large aounts of data and inforation have to be processed to gain adequate situational awareness and the ability to adapt to eergency situations. Managing this coplexity is eerging as a critical issue in coplex network operation. Lack of coplexity anageent often results in the inability to anticipate, recognize, and respond to situations caused by huan errors, natural disasters, and cyber attacks and inadequacy in predicting the effect of operational decisions. In this paper, we apply visual analytics techniques to enhance the processing of large aounts of operational data. Previously, visual analytics has been successfully applied to process assive aounts of data and extract useful inforation fro this data (NVAC 008). In this paper, we adapt the visual analytics techniques to convert assive aounts of operational data into actionable inforation. The resultant application enables prediction of network status, enhances the response to network operational requireents and provides real-tie decision support to network operators. The application has been successfully deonstrated with actual power grid odels and data. In the next section, a brief overview of the power grid operation is presented to address the needs for realtie decision support. The following section defines the probles in power grid operations and the process of applying visual analytics techniques. The application ais to convert data into inforation and present the inforation in an operator-friendly anner as a contoured geographic ap. With the contoured ap as the basis, the 59
next section develops a ethod to predict network security trends based on graph analysis. In both sections, actual power grid exaples are presented to deonstrate the contoured apping and graph trending analysis. The paper is then concluded with closing rearks and recoendations for future work. Overview of Power Grid Operation Recent power grid blackouts, such as the west coast blackouts of 996 (Kosterev, Taylor, Mittelstandt 999) and the east coast blackout of 00(DOE 00), brought significant attention to the reliability of power grids. How to predict and prevent or itigate such blackouts has been a central topic in the area of power syste research, and has also becoe one of the priary focuses of the DOE Office of Electricity Delivery & Energy Reliability (DOE- OE) (Congress 005). Power grid operation involves coplex coputational processes with advanced power grid odels. Figure shows a functional structure of realtie power grid operation (Huang et al. 007). The processes investigated in this paper are State Estiation and Contingency Analysis. The State Estiator typically receives teleetered data fro the supervisory control and data acquisition (SCADA) syste every few seconds and extrapolates a full set of grid conditions for operators based on the grid s current configuration and a theoretically based engineering power flow solution. The output of the State Estiator drives other operation functions including Contingency Analysis. Contingency Analysis studies what-if conditions in anticipation of potential power grid failures. Contingency Analysis identifies operation violations if one or ore eleents fail. The violation results are then presented to operators for review and decision-aking to deterine reedial actions if necessary. The burden of decision-aking all falls on the shoulders of the operators. power grid should not cause syste instabilities; this is referred to as the N- reliability criteria (NERC 005). Contingency analysis is continually run in an interval of seconds to inutes to deterine the ipact of equipent failures. If the loss of one or ore eleents does not result in any liit violations, then the syste is said to be secure for that contingency. The contingencies that result in violations of operating liits are flagged and placed in a list for the operators to inspect. NERC andates that operators take actions to itigate the situation in a tiely anner when there are contingency violations. Because it is not uncoon for several hundred contingencies to be exained, conveying this inforation to syste operators in a eaningful and easy-tounderstand way is a fundaental challenge. Because of the size of odern power grids, the nuber of contingencies to be studied can be very large. For exaple, the western North Aerican high-voltage power grid has about 0,000 eleents. Failure of any one eleent, i.e. N- contingencies, would constitute 0,000 contingency cases. N- contingencies would be in the order of 0 8. Actual grid blackouts often involve the failure of ultiple eleents (N-x contingencies). State-of-the-art coercial tools use a tabular for to display each of the contingency violations, as shown in Figure. Operator Review Figure. Functional structure of power syste operations The operator s decision-aking process is the key step in ensuring power grid reliability. The operating standards of the North Aerican Electric Reliability Corporation (NERC) require that the loss of any single eleent in the Figure. Tabular representation of violation data in the state-ofthe-art power grid operation tool Each violation of operating liits is a row in this tabular for, without showing geographical inforation and the degree of severity. When there are only a few contingencies where the syste is not N- secure, the ethod of tabular display is adequate. But when the syste is heavily stressed, and there are significantly ore contingencies violations, the tabular ethod of display is rapidly overloaded. It is then ipossible for an operator to sift through the large aounts of violation data and understand the syste situation within several seconds or inutes. However, it is in these situations that the 60
operators would ost need the inforation when the tabular representation techniques are saturated. Because of the above-entioned challenges in processing large volues of data fro the contingency analysis process, it is certain that we will need a second layer of analytical tools to analyze the data and extract useful and necessary inforation for power grid operators. This layer of tools not only provides inforation about the current power grid status, but they also analyze historical data and generate syste trending inforation to enable predictive capabilities. With this kind of real-tie decision support, the operators will then have no need to review the assive aount of data but be presented with actionable inforation of the current status and syste trends. The next two sections present the application of visual analytics and graph trending techniques to construct such decision support tools for power grid operators. This decision support syste will be able to fully utilize the contingency analysis results to predict potential probles of the power grid and adapt the power grid to adverse situations. Though this paper presents power grid exaples, the decision support syste can be extended for coplex network operations in other industries. Exaples include gas pipeline systes, telecounication systes, and aviation networks. Risk Assessent of Contingency Violations Contingency violations are defined as operational paraeters (i.e. power on a line or voltage at a substation) exceeding their liits. For exaple, the power that a transission line can transfer has a liit due to theral or stability constraints (NERC 997). Exceeding the liits will result in equipent failure and/or syste instability. Thus, the risk of a transission line can be defined as the relative loading R% with respect to the liit P ax : Pik R ik % 00% () P i ax where ik denotes the ith line of the kth contingency. The risk of a substation can be defined siilarly with the only difference being that the substation voltage has both lower and upper liits (V in and V ax ). V ik Vi in V i ax Vi in V V R % 00% () ik i ax i in where ik denotes the ith substation of the kth contingency. For each contingency k, the risk of lines and substations can be categorized as: 0, RT %, R 00%,, safe R ik % T %,00%, alert () violation where R T % is the pre-specified alert risk level. Copared with the tabular for shown in Figure, the iproveent is that ()-() convert the contingency data into quantitative risk levels, which indicate the severity if an operational paraeter exceeds a liit. This conversion also goes beyond the violation data. Risk levels are defined as how close the operational paraeters are to the liit, even if there are no violations, as shown in (). Defining the risk levels of individual eleents (lines and substations) is the first step in converting contingency data into actionable inforation. Each contingency will generate a set of risk levels as defined in ()-(). If a total of K contingencies are analyzed, there will be K sets of risk levels. Across all of the contingencies, the risk level of the ith eleent can be defined statistically as the axiu, suation, or ean of the individual risk levels. Using the axiu as an exaple is shown below: R axr %, k =,,,K () i % ik Two questions reain: How to present the risk levels in an easy-to-understand anner? And how to define risk levels for the whole network and for regions of interest? The next section presents an approach based on visual analytics techniques. Visual Analytical Application Failure of one eleent in a power grid could propagate into other areas of the grid. Given the different geographic locations of lines and substations and the heterogeneous structure of a power grid, the propagation would likely be different, and a sae risk level of different lines or substations would have various levels of ipact to the power grid. We assue that higher risk levels and risks in dense areas would have a larger ipact on the reliability of the syste. We further assue that the sae risk level would propagate into the sae radius of a geographic area, which is deterined using visual analytics techniques. The result of this application is a contoured ap with the color indicating the risk levels. Then it is very easy for operators to see the vulnerable areas of the grids without the need to sift through individual nubers. Visual Analytics-Based Contoured Maps The visualization starts with assigning the lines and substations the risk level as defined in () on the geographical ap of the power grid. Then the propagation is visualized as fading colors fro the center as shown in Figure. The ipact area of a substation has a circular shape, while a line has an elliptical shape. Individual risk 6
areas are then superposed to for the collective risk areas. The sae superposition is done aong ultiple contingencies as well. Superpose Figure. Collective risk area using fading colors and superposition of individual risk areas The ipleentation uses a hash table to store all the pixels of the lines and substations. Each pixel has a value deterined by the risk level of the line or substation. When lines are crossing, the larger value reains in the table so the highest risk is represented (Figure ). order for it to be easy to interpret, a green/gray/red color ap is selected. Considering (), the color ap can be understood as green, gray and red correspond to three risk categories safe, alert and violation. The final visual representation uses HaveGreen (Wong et al. 006) as the application fraework, which provides the interface for navigating and zooing over the power grid. The graphics is developed in C# using Managed DirectX. An exaple of the color contoured ap is shown in Figure 6. This exaple uses actual odel and data of the western North Aerican power grid. 00 contingencies are analyzed, and 00 sets of risk levels are overlaid on the single ap to visualize the collective risk of the contingencies on the syste security. The red color (shown as darkest areas in the figure) indicates vulnerable portions of the power grid and brings attention to network operators. Copared with Figure, this color contoured ap has the obvious advantage of bringing inforation rather than raw data to operators. Figure. Hash table for storing pixels The next step is to create the color filter for displaying the resultant collective risk. The filter is circular shaped with values conforing to that of a Gaussian curve (Figure 5). The Gaussian curve is noralized so that the peak height is equal to one. The radius of the filter is a paraeter that is set by the user. We define the Gaussian curve to have three standard deviations within one radius. Next we iterate through all the pixel points associated with the lines and substations stored in the hash table. Green Gray Red Gray Green Figure 5. Gaussian color filter and green/gray/red color ap. At each one of these points, the value in the table is ultiplied by the Gaussian curve. These values are then added to an output graphic atrix representing the final contour. The outcoe of the Gaussian filtering is the output atrix defining each point in the ap with a floating point nuber. Then these floating point nubers are assigned to a color ap to obtain the final contour. In Figure 6. Western North Aerican power grid risk ap with 00 overlaid contingency analysis results Syste and Regional Risk Levels The color contoured ap visually shows the risk across the network. This section eploys statistical analysis ethods to quantitatively calculate the risk level R% of the network and individual regions. The risk level is defined as a cobination of arithetic average and geoetric average. R% a a (5) where a and a are weighting constants. and are the arithetic average and geoetric average, respectively. The statistics is perfored over all the pixel points on the ap. Each pixel has a color value corresponding to the risk level at that pixel. If we categorized the pixels into M categories and there are N pixels in each category with the sae color value (R%), the arithetic and geoetric averages are calculated as follows: 6
R%, =,,,M (6) N N N N R%, =,,,M (7) For regional risk levels, the sae process can be applied but only the pixels in the region are considered. Figure 7 shows the risk levels of the western North Aerican power grid over a orning load pick-up period. When the syste total power consuption is at a low level (the beginning of the period), increasing load does not increase risk levels as uch as when the total load is at a higher level toward the end of the period. This is consistent with operational experience. Syste Risk Level Syste Stress Level 6 Figure 7. Syste risk level and stress level over tie It is worth pointing out that the sae statistical analysis can be perfored on the risk levels calculated fro (). The advantage of perforing the analysis in the visual space is that the propagation and collective risk areas are considered, which is ore reasonable and realistic for actual power grids. Visual Trending Analysis Converting the data into risk levels and visualizing the as contoured aps enables visual network anageent, which akes it easier to gain situational awareness and recognize probles. Operators would now have ore tie to focus on urgent issues rather than spending tie analyzing uniportant data. Operators could also observe the evolving patterns of the visual aps to deterine network reliability and security trends. For exaple, an increase in color intensity and size of the risk contour would indicate a deteriorating network situation that would raise awareness for the operator. In a siple network, evolving patterns are siple to understand and visual exaination of the aps would be adequate to deterine trends. However, in a coplex network, evolving patterns can be coplicated and the nuber of the patterns can be 8 0 Stress Level nuerous at any given tie. Figure 8 shows a sipler case with a few obvious red areas and several fuzzy gray areas. All areas evolve in tie. An operator ay be able to recognize the pattern of areas and erging into one single area. But it would be very difficult to deterine how the other areas are evolving and how to quantify the iplications. And ore iportantly is to use the results of the contingency analysis to deterine the trend and predict the network status in the future. Evolve Figure 8. Power grid risk evolving patterns To enable this predictive capability, we developed a ethod for visual trending analysis. The ethod is based on the syste and regional risk levels as defined in (5). The trend is obtained by fitting a curve to historical risk levels of the network or regions, and extrapolating to predict the future syste situation, as shown in Figure 9. Figure 9. Illustration of visual trending analysis Coplex evolving patterns ay exist in a network. Soe of coplex patterns are shown in Figure 0. Two areas can erge into one, or one can split into two, or one area steals a portion of another area. Merging Merging/ Splitting R% Past Now Future tie Stealing Stealing/ Splitting Figure 0. Coplex evolving patterns of network risk ipact areas To autoatically identify all the coplex patterns, the actual ipleentation of visual trending analysis cobines structural analysis and statistical analysis, as shown in 5 6
Figure. Statistical analysis is used to calculate the risk indices of individual areas, while the structural analysis uses a relation atrix to capture the relationship between areas, i.e. how two areas overlap or differ at the pixel level. The nuber in the relation atrix is the su of the risk level for each pixel in the overlapped area, except the last row and last colun are for the adjacent area. The nubers in Figure correspond to the case shown in Figure 8. This trending analysis approach has been able to identify all the coplex evolving patterns shown in Figure 0. 0 A A 0 0 0 0 A Figure. Cobination of structural analysis and statistical analysis using a relation atrix for visual trending analysis. The green dashed line in Figure 7 is the predicted syste risk level, each point based on the three prior risk levels. It can be seen that the prediction is reasonably close to the actual syste risk level (blue line). Figure further shows the trends for the five ost critical regions in the power grid for this exaple, corresponding to the sae syste conditions in Figure 7. The regional risk trends are ore extree than the syste trend. The syste trend is relatively flat as changes in different regions ay cancel each other s ipact. Therefore it is iportant to observe regional trends to recognize potential regional issues. situational awareness of the network without sifting through large aounts of raw data. Predictive capability is established by analyzing the trend of the network risk level with an approach that cobines structural analysis with statistical analysis. Exaples using actual odels and data of the western North Aerican power grid deonstrate the validity of the predictive analytics. Given that these results were obtained in a research environent and based on siulation, an iportant step to bring these proposed ethods into practice will be deonstrating this approach in an actual power syste environent and evaluating the decision support tool with experienced operators. Further work should also focus on developing ethods to iprove the prediction by including probability to enable better forecasting for network operations. This forecasting capability refers to ultiple future paths, each of which has a certain probability. Another enhanceent is helping operators decide the outcoe of various reedial actions through an interactive analysis function. This interactive analysis would identify reedial actions that can turn the red to green on the contoured ap. Soe of this work is ongoing and results are expected to be published in the near future. Acknowledgeent This work is supported by the Inforation and Infrastructure Integrity Initiative of the Pacific Northwest National Laboratory. The Pacific Northwest National Laboratory is operated by Battelle for the U.S. Depart of Energy under Contract DE-AC06-76RL080. The authors would like to extend special thanks to Ning Zhou, Jeff Dagle, Ji Thoas and Mark Hadley, all with Pacific Northwest National Laboratory, for their insightful coents and support to this work. References Congress U.S. 005. Energy Policy Act of 005. URL: http://www.doi.gov/iepa/energypolicyactof005.pdf. DOE. 00. U.S.-Canada Power Syste Outage Task Force: Final Report on the August, 00 Blackout in the United States and Canada: Causes and Recoendations. Figure. Exaple of regional risk trends of the western North Aerican power grid Conclusion and Future Work This paper described visual analytics techniques that were successfully applied to coplex network operations by converting large aounts of operational data into actionable inforation. Operational data are translated into risk levels and then visualized as a color contoured ap. This application will greatly iprove real-tie decision support for network operations. Operators can quickly gain Huang, Z., Guttroson, R., Nieplocha, J., and Pratt, R. 007. Transforing Power Grid Operations via High-Perforance Coputing. Scientific Coputing April. Kosterev, D.N., Taylor, C.W. and Mittelstadt, W.A. 999. Model Validation for the August 0, 996 WSCC Syste Outage. IEEE Trans. Power Syst. (): 967-979. The National Visualization and Analytics Center (NVAC). 008. Accessed /008. URL: http://nvac.pnl.gov/. North Aerican Electric Reliability Corporation (NERC). 997. NERC Planning Standards. URL: www.nerc.co/pub/sys/all_updl/pc/pss/ps9709.pdf. 6
North Aerican Electric Reliability Corporation (NERC). 005. Transission Syste Standards Noral and Eergency Conditions. Accessed /008. URL: www.nerc.co. Wong, P.C., Chin, Jr., G., Foote, H.P., Mackey, P.S., and Thoas, J.J. 006. Have Green A Visual Analytics Fraework for Large Seantic Graphs. In IEEE Syposiu on Visual Analytics Science and Technology. Baltiore, MD. 65