cond-mat0401084/aan.tex
1: \documentclass[aps,prl,preprint,groupedaddress,showpacs]{revtex4}
2: \usepackage{graphicx}
3: 
4: \begin{document}
5: \title{Structural Vulnerability of the North American Power Grid}
6: 
7: \author{R\'eka Albert$^{1,2}$ Istv\'an Albert$^2$ and Gary L. Nakarado$^3$}
8: \affiliation{1. Department of Physics, Pennsylvania State University, University Park, PA 16802}
9: \affiliation{2. Huck Institute for Life Sciences, Pennsylvania State University, University Park, 
10: PA 16802}
11: \affiliation{3. National Renewable Energy Laboratory, Golden, CO 80401}
12: 
13: \begin{abstract}
14: 
15: The magnitude of the August 2003 blackout affecting the United States has put 
16: the challenges of energy transmission and distribution into limelight. 
17: Despite all the interest and concerted effort, the complexity and interconnectivity 
18: of the electric infrastructure have so far precluded us from understanding why certain events 
19: happened. In this paper we study the power grid from a network perspective and determine 
20: its ability to transfer power between generators and consumers when certain nodes are disrupted. 
21: We find that the power grid is robust to most perturbations, yet disturbances affecting key 
22: transmision substations greatly reduce
23: its ability to function. We emphasize that the global properties of the 
24: underlying network must be understood as they greatly affect local 
25: behavior. 
26: \end{abstract}
27: 
28: \pacs{89.75.Fb, 02.10.Ox, 84.70.+p, 89.75.Hc, 89.75.Da} 
29: 
30: \maketitle
31: 
32: During the past decades the North American power infrastructure has evolved into what many experts 
33: consider the largest and most complex system of the technological age. Geographically, the power 
34: grid forms a network of over 1 million kilometers of high voltage lines that are continuously 
35: regulated by sophisticated flow control equipment\cite{roadmap}. As a result of the recent
36: deregulation of power generation and transmission, about one-half of all domestic 
37: generation is now sold over ever-increasing distances on the wholesale market before it is 
38: delivered to customers\cite{roadmap}. Consequently the power grid is witnessing power flows in 
39: unprecedented magnitudes and directions\cite{rely}. 
40: 
41: As the power grid increases in size and complexity, it is becoming more important to understand 
42: the emergent behaviors that can take place in the system. Performing an analytic description of the 
43: electromagnetic processes integrated over the whole grid is a daunting, if not impossible, task. 
44: Instead the power industry must resort to constructing models that can be used to simulate the network's 
45: response to various external parameters. Generally these models attempt to simulate actual 
46: electrical flow characteristics in smaller systems like a single distribution grid\cite{flow}. 
47: In the present analysis we propose an alternative approach based on recent advances in understanding 
48: the structure of large complex networks\cite{ab02}. We choose to investigate the network 
49: representation of the power grid from a topological perspective with the hope of finding 
50: properties and behaviors that transcend the abstraction. 
51: 
52: 
53: We have built the network model based on data stored in the POWERmap mapping 
54: system developed by Platts\cite{platts}, the energy information and market services unit of 
55: the McGraw-Hill Companies. This mapping system contains information about every power plant, 
56: major substation and  $115-765$ kV power line of the North American power grid. Our model 
57: represents the power grid as a network of 14,099 nodes (substations) and 19,657 edges 
58: (transmission lines). We distinguish three types of substations: generators are the sources
59: for power, transmission substations transfer the power among high voltage transmission lines, and 
60: distribution substations are at the “outer edge” of the transmission grid, and the centers of local 
61: distribution grids. Only the identity of generating substations was directly available from our data 
62: sources. We identify distribution substations by the criterion of having a single high-voltage 
63: transmission line connected to them, with the expectation that the flow out of them is continued on
64: smaller voltage feeder lines leading to consumers\cite{miss}. A total of 1633 nodes are power plants,
65: we classify 2179 nodes as distributing substations, with the rest being labeled as transmission 
66: substations.
67:  
68: We consider the power from a generator to be accessible to a consumer if there is a path of 
69: transmission lines between the two. In practice, the existence of 
70: a connection between two substations does not always imply that power can be transferred across it 
71: as there may be capacity or other constrains present. By ignoring these our model provides an 
72: idealized view, a “best case scenario” regarding the characteristics of the grid. We find that the 
73: network representation of the power grid contains a single connected component, meaning that 
74: there is a path of transmission lines between any power plant and any distribution substation. 
75: This observation implies that in the best case scenario each distribution substation can possibly 
76: receive power from any generator. 
77: 
78: Recent advances in mapping the topology of complex networks have uncovered that a large fraction 
79: of them are highly heterogeneous with respect to the number of edges incident on a node (also 
80: called the node degree). In these networks the majority of the nodes have low degrees, but there
81: is a continuous hierarchy of high-degree nodes (hubs) that play an important role in the system.
82: The degree distribution of these networks follows a power-law $P(k)\sim k^{-\gamma}$ with the exponent
83: $\gamma$ mostly between $2$ and $3$. It was demonstrated both numerically and analytically that these
84: so-called scale-free networks are resilient to the random loss of nodes, but are vulnerable to attacks 
85: targeting the high-degree hubs\cite{ajb00,cnsw00,cebh01}. Therefore it is important both from a 
86: theoretical and practical standpoint to determine whether the connectivity of the power grid is reliant
87: on a small set of hubs and whether their loss will cause a large-scale breakdown of the power grid's 
88: transmission capability. 
89: 
90: As the node degree is a good indicator of its topological importance, we first determine the degree 
91: distribution of the power grid. We find that the cumulative degree distribution defined as 
92: $P(k>K)=\sum_{k>K} P(k)$ follows an exponential  
93: \begin{equation}
94: P(k>K)\sim exp(-0.5 K) 
95: \label{deg_eq}
96: \end{equation}
97: (see Fig. \ref{degree_fig}).
98: This functional form agrees with previous results on the degree distribution of the Western power 
99: grid\cite{asbs00} and its classification as a single-scale network. The cumulative degree
100: distribution shows that the probability of high-degree nodes is less than in a scale-free network, 
101: but higher than in a random network with the same number of nodes and edges. Power engineering
102: principles suggest that the hubs of the power grid should belong to central station generators, and 
103: transmission substations should not have more than a few edges. Indeed, the inset to Fig. \ref{degree_fig}
104: shows that the fraction of generating substations among substations of
105: a given degree increases with this degree. Surprisingly, however, there are several high-degree 
106: transmission substations (e.g. $50$ have degree higher than $10$), including the node with highest 
107: degree. 
108: 
109: \begin{figure}
110: \includegraphics[width=8cm,angle=-90]{aan_fig1.ps}% 
111: \caption{\label{degree_fig} The probability that a substation has 
112: more than $K$ transmission lines. The straight line represents the exponential 
113: function (\ref{deg_eq}). Inset: the fraction $F_g(k)$ of generating substations among substations with 
114: degree $k$.}
115: \end{figure}
116: 
117: As the role of the power grid is to transport power from generators to consumers, a possible measure 
118: for the importance of a node corresponding to a substation is its betweenness (or load)\cite{gkk01,n01}.
119: The betweenness of a node in a network is defined as the number of shortest paths that traverse 
120: it\cite{gkk01,n01}. Assuming that power is routed through the most direct path, the betweenness of a substation 
121: is a proxy for how much power it is transmitting, and for this reason we will use the alternative term 
122: load to denote it. Since it is the transmission substations' role to route power from generators 
123: to distribution substations, we focus our attention to them. We determine the shortest paths starting 
124: from all generation substations and ending on an all other reachable substations. For each transmission node we accumulate the number of paths that 
125: pass through it; being at the start or at the end of a path does not count. The highest possible load 
126: is $1633\times 12466\simeq 20$ million. We find that substations can have a load anywhere between 1 
127: and 4 million, and determine the cumulative load distribution, i.e. the probability that a node's load 
128: $l$ is larger than a given value $L$ (see Fig. \ref{load_fig}). The functional form of the 
129: cumulative load distribution is
130: \begin{equation}
131: P(l>L)\sim (2500+L)^{-0.7}
132: \label{load_eq}
133: \end{equation}
134: Fig. \ref{load_fig} illustrates that $40\%$ of the 
135: substations participate in tens or hundreds of paths only, but $1\%$ of them are part of a million 
136: or more paths. These high-load substations, although possibly not hubs regarding their degree, 
137: play an important role in power transmission. 
138: 
139: \begin{figure}
140: \includegraphics[width=8cm,angle=-90]{aan_fig2.ps}% 
141: \caption{\label{load_fig} The probability that a substation has more than L transmission paths 
142: passing through it. The 
143: continuous curve has the generalized power law form (\ref{load_eq}). Inset: histogram of the length
144: of the shortest alternative path $r$ between the endpoints of an edge. In order to be able to include edges
145: with no alternative path, the abscissa is inverted. }
146: \end{figure}
147: 
148: A fundamental requirement of the power grid is robustness, the ability to withstand and tolerate
149: errors (random failure) and targeted attacks\cite{ajb00,cebh01,cnsw00}. To ensure the reliability of 
150: power distribution, the transmission grid was conceived in such a way that there is more than one 
151: electrical path between any two points in the system \cite{power_book}. We wanted to verify whether the 
152: actual topology of the current power grid has this feature of global redundancy, or it has lost it during its 
153: growth and evolution. A possible measure of network redundancy is the so-called edge range, defined as 
154: the distance between the two endpoints of an edge if the edge connecting them were removed \cite{mnl02}.
155:  The inset of Fig. \ref{load_fig} shows the frequency of 
156: different edge ranges $r$ plotted as a function of $r^{-1}$. We find that parallel edges and short 
157: alternative paths are fairly frequent. However, around $15\%$ of the edges in the power grid have 
158: an infinite range. In addition to the $2179$ edges ending in distribution 
159: substations, close to $900$ edges connecting generators and/or transmission substations
160: are radial. These radial edges represent a clear vulnerability, as their loss disconnects 
161: their endpoints and creates isolated clusters in the power grid.
162: 
163: While the connectedness of the power grid allows for the transmission of power over large distances, 
164: it also implies that local disturbances propagate over the whole grid. The failure of a power line 
165: due to lightning strike or short-circuit leads to the overloading of parallel and nearby lines. 
166: Power lines are guarded by automatic devices that take them out of service when the voltage on them 
167: is too high. Generating substations are designed to switch off if their power cannot be transmitted; 
168: this protective measure has the unwanted effect of diminishing power for all consumers. Another 
169: possible consequence of power line failure is the incapacitation of transmission substations, 
170: possibly causing that the power from generators cannot reach distribution substations and ultimately 
171: consumers. 
172: 
173: In the unperturbed state each distribution substation can receive power from any of the
174: $N_g=1633$ generators. As substations lose function, the number of generators connected
175: to (and able to feed) a certain distribution substation $i$, $N_g^i$, decreases. We introduce
176: the concept of connectivity loss to quantify
177: the average decrease in the number of generators connected to a distributing substation,
178: 
179: \begin{equation}
180: CL=1-\left\langle\frac{N_g^i}{N_g}\right\rangle_i,
181: \end{equation}  
182: where the averaging is done over every distributing substation. In summary, the connectivity loss 
183: measures the decrease of the ability of distribution substations to receive power from the generators, and
184: in the following we will express it as a percentage.
185:  
186: \begin{figure}
187: \includegraphics[width=8cm,angle=-90]{aan_fig3.ps}% 
188: \caption{\label{source_fig} Connectivity loss in the power grid resulting from the failure of a 
189: fraction $f_g$ of generators. The straight line represents the minimum loss due to the 
190: node removal itself. Circles: random removal of generators; triangles: removal starting from the 
191: highest-degree generators. The curves are averages of ten runs, where either 
192: the list of generators or the list of generators with the same degree was randomly permuted.}
193: \end{figure}
194: 
195:  First we investigate the effect that the failure 
196:  of a power-generating substation has on consumers. Since initially the network contains a single
197:  connected component every consumer can reach all generators, and their connectivity is $100\%$. As 
198:  the number of generators decreases this value will decrease due to both loss of the generators themselves and 
199:  due to loss of routing capabilities at the generating substation level. 
200:  We remove nodes corresponding
201:  to generators either randomly, or in the decreasing order of their degrees, and monitor the connectivity loss
202:  as a function of the fraction of generators missing. The minimum possible loss is equal to the fraction 
203:  $f_g$ of inactive generators and is due to the loss in generation only (straight line on 
204:  Fig. \ref{source_fig}). We find that the 
205:  connectivity loss caused by removing power substations remains very close to this minimum value 
206:  (Fig. \ref{source_fig}), even though generating substations tend to be the 
207:  largest hubs in the system. The removal of generating substations does not alter the overall 
208:  connectivity of the grid thanks to a high level of redundancy at the power generating substation 
209:  level.
210:  
211:  \begin{figure}
212: \includegraphics[width=8cm,angle=-90]{aan_fig4_rev.ps}% 
213: \caption{\label{trans_fig} Connectivity loss in the power grid due to the removal of nodes corresponding
214: to transmission substations. We remove a fraction $f_t$ of transmission nodes with four different
215: algorithms: randomly (circles), in the decreasing order of their degrees (triangles) or loads (diamonds),
216: and by recalculating the load every ten steps and removing the ten nodes with highest load (squares). The curves corresponding to 
217: random and degree-based node removal were averaged over ten runs. The load-based and cascading removal 
218: curves represent a single run.}
219: \end{figure}
220:  
221: The situation can be dramatically different when the nodes that we remove are transmission nodes. If 
222: the power grid were highly redundant the loss of a small number of transmission substations
223: should not cause power loss as power is rerouted through alternative paths. We find that even the removal
224: of a single transmission node causes a slight connectivity loss. We remove transmission nodes one by 
225: one,
226: first randomly, then in the decreasing order of their degree or load. 
227: For a random failure the connectivity loss is fairly low and stays proportional with the number of nodes 
228: lost. The connectivity loss is 
229: significantly higher, however, when targeting high degree or high load transmission hubs 
230: (Fig. \ref{trans_fig}). The grid can withstand only a few failures of this nature before considerable parts of the network
231: become disconnected leading to substantial connectivity loss at consumer level. For example,
232: failure of only $4\%$ of the nodes with high load may cause up to $60\%$ loss of connectivity. We also study an algorithm where we periodically recalculate the
233: load of all transmission nodes during node removal, and select the nodes with highest load to
234: be deleted next. This is a possible illustration of a propagating (cascading) power failure, 
235: where it is more likely that substations that have the highest load in the perturbed configuration
236: will fail next. Fig. \ref{trans_fig} illustrates that this cascading failure has the most damaging effect, 
237: as the loss of only $2\%$ of the high-load transmission substations leads to a connectivity loss of
238: almost $60\%$, and all distribution substations become virtually powerless at $f_t \simeq 8\%$. In conclusion,
239: the transmission hubs ensuring the connectivity of the power grid are also its largest liability in 
240: case of power breakdowns.
241:   
242:  
243:  
244: This vulnerability of the electric power grid is inherent to its organization and therefore cannot 
245: be easily addressed without significant investment. Possible solutions include increasing the 
246: redundancy and capacity of the existent structure or decreasing the reliance on transmission by 
247: incorporating more generation at the distribution substation level. Such distributed generation 
248: by small local plants can supplement power from the grid under normal operation conditions and 
249: can greatly mitigate the effects of blackouts on the population. Targeted use of generation 
250: located near the point of use might prove to be the only viable economical alternative.
251: 
252: \begin{acknowledgements}
253:  The authors wish to thank Donna Heimiller and Steven Englebretson for 
254: their help in obtaining the POWERmap network data. This research was partially supported 
255: by the Midwest Research Institute (contract number AAX-3--33641-01).
256: \end{acknowledgements}
257: 
258: 
259: \begin{thebibliography}{10}
260: \bibitem{roadmap}
261: Electricity Technology Roadmap, 1999 Summary and Synthesis, by the Electric Power Research 
262: Institute, \url{http://www.epri.com/corporate/discover_epri/roadmap/}
263: \bibitem{rely}
264: North American Electricity Reliability Council reliability assessment report, 1998,
265: \url{http://www.nerc.com/~filez/rasreports.html}
266: \bibitem{flow}
267: Dromey Design electrical distribution analysis software, 
268: \url{http://www.dromeydesign.com/dess/lfa.htm}
269: \bibitem{ab02}
270: R. Albert and A.-L.  Barab\'asi , {\it Reviews of Modern Physics} {\bf 74}, 44-94 (2002); A.-L. Barab\'asi
271: {\it Linked: The New Science of Networks} (Perseus Publishing, Cambridge, 2002); D. J. Watts 
272: {\it Six Degrees: The Science of a Connected Age} (W. W. Norton $\&$ Co., New York, 2003); S. N. Dorogovtsev 
273: and J. F. F. Mendes, {\it Evolution of Networks: From Biological Nets to the Internet and WWW} (Oxford
274: University Press, Oxford, 2003); M. E. J. Newman, {\it SIAM Review} {\bf 45}, 167 (2003).
275: \bibitem{platts}
276: Platts Global Energy, \url{http://www.platts.com/electricpower/index.shtml}.
277: \bibitem{miss}
278: This method may miss distribution substations with more than one incoming transmission line. 
279: Unfortunately no information regarding the directionality of the transmission lines was available
280: from our data sources.
281: \bibitem{ajb00}
282: R. Albert, H. Jeong and A.-L. Barab\'asi, {\it Nature} {\bf 406}, 378 (2000).
283: \bibitem{cebh01}
284:  R. Cohen,  K. Erez, D. ben-Avraham  and S. Havlin, {\it Phys. Rev. Lett.} {\bf 85}, 4626
285: (2000); R. Cohen,  K. Erez, D. ben-Avraham  and S. Havlin, {\it Phys. Rev. Lett.} {\bf 86}, 3682 (2001).
286: \bibitem{cnsw00}
287: D. S. Callaway, M. E. J. Newman, S. H.  Strogatz and D. J. Watts,  {\it Phys. Rev. Lett.}
288:  {\bf 85}, 5468 (2000).
289: \bibitem{asbs00}
290: L. A. N. Amaral, A.  Scala, M. Barth\'el\'emy  and H. E. Stanley, {\it  Proc. Natl. Acad. Sci. 
291: USA} {\bf 97}, 11149 (2000).
292: \bibitem{gkk01}
293: K.-I. Goh, B. Kahng and  D.  Kim, {\it Phys. Rev. Lett.} {\bf 87}, 278701 (2001).
294: \bibitem{n01}
295: M. E. J. Newman,  {\it Phys. Rev. E} {\bf 64}, 016132 (2001).
296: \bibitem{power_book}
297: H. Saadat, {\it Power System Analysis} (McGraw-Hill, Boston, 1999).
298: \bibitem{mnl02}
299: A. E. Motter, T. Nishikawa and Y.-C. Lai, {\it Phys. Rev. E} {\bf 66}, 0651103(R), (2002).
300: \end{thebibliography}
301: 
302: \end{document}
303: 
304: