Risk of Large Oil Spills: A Statistical Analysis in the ... - ACS Publications

Nov 5, 2012 - of spills, corresponding to the frequency, the pie chart on the .... parameter is the IMO rule to mandate double hull tankers, that led ...
0 downloads 0 Views 1MB Size
Download PDF
Article pubs.acs.org/est

Risk of Large Oil Spills: A Statistical Analysis in the Aftermath of Deepwater Horizon Petrissa Eckle,* Peter Burgherr, and Edouard Michaux Paul Scherrer Institut, Villigen PSI, Switzerland S Supporting Information *

ABSTRACT: The oil spill in the Gulf of Mexico that followed the explosion of the exploration platform Deepwater Horizon on 20 April 2010 was the largest accidental oil spill so far. In this paper we evaluate the risk of such very severe oil spills based on global historical data from our Energy-Related Severe Accident Database (ENSAD) and investigate if an accident of this size could have been “expected”. We also compare the risk of oil spills from such accidents in exploration and production to accidental spills from other activities in the oil chain (tanker ship transport, pipelines, storage/refinery) and analyze the two components of risk, frequency and severity (quantity of oil spilled) separately. This detailed analysis reveals the differences in the structure of the risk between different spill sources, differences in trends over time and it allows in particular assessing the risk of very severe events such as the Deepwater Horizon spill. Such top down risk assessment can serve as an important input to decision making by complementing bottom up engineering risk assessment and can be combined with impact assessment in environmental risk analysis.

■

INTRODUCTION Following the explosion of the exploration drilling rig Deepwater Horizon (DWH)1 in the Gulf of Mexico in April 2010 that killed 11 workers, ∼680 000 tons of oil ((4.6−6.2) million barrels) were spilled.2,3 BP paid out 14 billion dollars in immediate cleanup costs,4,5 and so far around 30 billion in total, but both the environmental and full economic consequences of this accident will only become clear over the coming years. Two questions will be addressed in this paper: The first is if the DWH spill can be considered an outlier or an accident that could have been “expected” based on historical experience. Second, we will compare the risk from such spills resulting from accidents on platforms, rigs and oil wells in exploration and production to the risk of spills from other activities in the oil chain, such as transport, storage and refining. Our analysis is based on a comprehensive global data set of oil spills (>200 t) that took place between 1974 and 2010, comprising around 1200 oil spills. Following the classical definition of risk being the product of probability of occurrence and consequences, the frequency and severity (amount of oil spilled) of accidental oil spills are analyzed. The results of this type of analysis can provide important inputs to objectively describe the current situation, but also to evaluate improvements over time, identify potential future trends, and provide essential inputs when establishing acceptable risk levels.6,7 Furthermore, such information can help to compare industry performance against regulatory risk targets and support decision-making and policy formulation processes. Finally, it can serve as an input to the next step of oil spill consequence assessment, which is the © 2012 American Chemical Society

estimation of environmental impacts. These effects are of course highly dependent on the actual location of a spill (e.g., distance to coast, ecological sensitivity of affected area), the prevailing weather situation affecting for example current and wind conditions, type of oil involved (crude vs products), and last but not least the response and mitigation measures (e.g., chemical agents, barriers, clean up).8−10 However, this second part of the consequence assessment was beyond the scope of the present study. With this top down approach to risk assessment we want to complement the extensive research on specific risks and impacts of oil spills, for example, in different environments, based on the chemical composition,11 or with specific geographical coverage, for example, for the U.S.12 Our analysis could also be used as an input to spill trajectory models such as, for example, ref 13. An in-depth analysis of tanker spills based on ENSAD can, for example, be found in refs 8 and 14. In our analysis we exclusively look at accidental oil spills, excluding other sources of spills,15 such as, for example, operational spills, (see, e.g., ref 16) and spills resulting from vandalism and terrorism17,18 because these cannot be assumed to follow the same statistical behavior as accidental spills, as the implementation of measures to avoid operational spills does not necessarily proportionally reduce the risk of very large spills. Received: Revised: Accepted: Published: 13002

July 20, 2012 October 23, 2012 November 5, 2012 November 5, 2012 dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008

Environmental Science & Technology

Article

The values of the thresholds are chosen based on two competing goals: on the one hand, the analysis should be based on as many data points as possible and include also smaller accidents. On the other hand, the bias introduced by underreporting needs to be considered: The data set becomes rapidly less complete for smaller accidents, as underreporting becomes more likely, in addition geographical bias is larger for smaller accidents, that is, very severe events are reported globally, whereas the databases for smaller events in non OECD countries are less reliable. The year 1974 was chosen since it can be considered to be the time from which serious reporting on oil spills started.30

Furthermore, measures to reduce accidental oil spills primarily focus on technical requirements on the vessels and port state control, whereas for operational and/or intentional spills it is mostly increased surveillance and governance mechanisms.17 In general, catastrophic events often initiate changes in regulations to improve safety and reduce effects on human health and the environment (e.g., refs 19−21). In oil spill risk from accidents, few very severe events are responsible for a high share of the aggregate total consequences. Out of the total 1213 accidental oil spills considered, the two largest spills, Deepwater Horizon and Ixtoc I (1979, 480 000 tons) account for 11% of the total oil spilled and the 50 largest spills account for 61% of the total amount. This high contribution of the largest events to the total spill amount means that average and aggregate spill volumes do not show the complete picture of the risk. Instead careful analysis of in particular the risk of such rare but very severe events is needed. Both Deepwater Horizon and Ixtoc I were accidents in drilling, suggesting a separate analysis of this risk might be important in view of the rapid increase in deep (>305 m = 1000 ft) and ultradeep (>1524 m = 5000 ft) offshore drilling, where both a geographical expansion as well as a trend toward drilling at ever greater depths can be seen.9,22,23

■

DATA: SPILL SOURCE CHARACTERIZATION Spill sources were categorized into four categories corresponding to functional steps along the oil chain: the data set “exploration/production” comprises spills during exploration and extraction, that is, spills from oil wells as well as drilling and exploration platforms and rigs, “ships” refers to tanker spills during transport of both crude oil and refined products, “pipelines” contains spills from both on- and offshore, crude and product pipelines, and “storage/refineries” comprise spills at fixed installations. This splitting into four categories reflects an important trade off in the statistical analysis of risk: On the one hand, it would be desirable to characterize the risk with high resolution, that is, for each type of infrastructure, country etc. separately, on the other hand, a large data basis is needed to determine the risk of in particular severe events with any meaningful uncertainty interval. The data was therefore split into subcategories and the non parametric Kolomogorov−Smirnov (KS-) test31 was used to check if subsets were significantly different in the severity distribution. This comparison between the four subsets of data showed mutually significant differences in severity. Storage and refinery were combined as no significant differences could be found. (This is not surprising as spills in refineries are often from storage tanks on site.)

■

DATA: SEVERE ACCIDENT DATABASE ENSAD For this study, energy related, accidental global oil spills larger than 200 t in the period between 1974 and 2010 were extracted from our uniquely comprehensive global Energy-Related Severe Accident Database (ENSAD), resulting in a total of 1213 accidents. ENSAD was established in 1998,24 recognizing the need for a systematic and comprehensive treatment of accident risk in energy generation. ENSAD collects accidents throughout entire energy chains, starting, for example, for the oil chain with exploration and extraction, intermediate processing in refineries to the final use in power plants and heating, considering also intermediate transport stages, both of crude oil and the refined products. The database is regularly updated25−28 from a variety of commercial and open sources. Primary information sources of ENSAD with a broad coverage of energy accidents include among others MHIDAS (Major Hazard Incident Data Service; not any more available), HINT (Hazards Intelligence), OSH (Occupational Health and Safety) Update, whereas other sources such as CTX (Center for Tankship Excellence), WOAD (Worldwide Offshore Accident Databank) or Incident News by NOAA (National Oceanic and Atmospheric Administration) specifically collect oil chain accidents. A detailed overview of relevant information sources considered by ENSAD is given in the following publications and references therein.24,29 ENSAD focuses on severe accidents surpassing thresholds in at least one of several criteria like human health (fatalities and injured), environment (amount of oil/products spilled), societal and economic impacts.24 The general threshold for a spill to be considered severe is 10 000 tons of oil spilled, for the purpose of this study however, ENSAD was updated with all spill incidents down to 200 tons in size. Two arguments can be given for setting a threshold in the analysis: The structure of the risk in general means that smaller accidents contribute less to the total, limiting the error introduced by this lower cutoff. The exact point of cutoff is debatable, we made a big data collection effort for this paper to reduce our usual limit of 10 000 tons for a severe oil spills for this paper.

■

METHODS AND DATA ANALYSIS To capture the full information about the risk, the two components of risk need to be determined separately: The frequency of spills and the relative distribution of the size of the spills, that is, the severity distribution. Smaller spills have a higher probability than larger spills, the most important parameter in the severity distribution is thus how fast the probability drops with spill size. Figure 1 gives an overview over the data in the four data sets. The pie chart on the left shows the distribution of the number

Figure 1. On the left the number of incidents in the different infrastructure categories is shown, corresponding to the historical frequency. The chart on the right shows the amounts per infrastructure category, corresponding to the severity. 13003

dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008

Environmental Science & Technology

Article

threshold choice for extreme value anlysis can be found in ref 38 One possible way to find an appropriate threshold is thus to test the fit for different thresholds and comparing the resulting fit with the data using the KS test. The details of this test are provided in the Supporting Information. For the exploration/production data set, the results show that the GPD provides indeed a good fit for the data, and the fit is clearly improved by using a slightly higher threshold of 600 tons compared to the original 200 tons. For the ship and pipeline data set, the fit is rejected, storage/ refinery could be fitted using the GPD, but very low p-values close to the acceptance threshold indicate that the fit is not very good and becoming worse with higher thresholds instead of better as expected. The reason is that these data sets show features that can not be appropriately reproduced by a simple distribution: Distinct features are visible in the case, for example, for ship spills, where the distribution of spill severity is determined by the composition of the global oil tanker fleet as well as by the average share of oil spilled in the case of an accident. The shipping fleet is not a continuous distribution but can be divided in classes built to fit important transit passages such as the Panama Canal or the Suez Canal. For analysis, these data sets are thus given as empirical distributions. The data fit to the exploration/production data was performed as a Bayesian model, with broad, uninformative priors for both parameters of the GPD, implemented using the library functions in OpenBUGS.39 Bayesian modeling is extensively treated in ref 40, and Bayesian modeling for extreme value distributions is discussed in ref 41. An essential advantage of this approach is that the fit parameters are treated as distributions, so that the width of these distributions indicates the uncertainty of the fit. The results for the parameters of the GPD as well as the corresponing 5−95% range of the parameters are given in Table 1 for both the data set with and without the new data point of Deepwater Horizon. The corresponding risk distributions for both cases are shown in Figure 4.

of spills, corresponding to the frequency, the pie chart on the right shows the totals of oil spilled from the different sources, i.e. the total severity. Already from this graph it can be concluded that few exploration/production spills contribute disproportionately to the total amount of oil spilled, suggesting a higher severity per spill.

■

METHODS AND DATA ANALYSIS: SEVERITY DISTRIBUTION The severity distribution can either be given as empirical distribution constructed from data points directly, resulting in an empirical frequency-consequence curve or as a fitted probability distribution (see, e.g., Figure 4). Empirical distributions do not require any assumptions about the shape of the distribution, and thus offer a reliable characterization of the risk in particular if a large set of observations is available. A second important criterion for the use of empirical distributions is that natural upper boundaries to severity should exist so that historical worst cases can be considered a good proxy for the risk of very severe events. This is the case, for example, for ship spills, where a large historical record is available and the spill amount is limited by the capacity of the tanker ship (currently the biggest crude carriers have a capacity of 550 000 DWT (dead weight tons) 32). Similar arguments apply to storage tanks. Fitted probability distributions on the other hand allow quantifying the uncertainty of the severity distribution, a very important aspect in particular for the extreme risks at the tail of the distribution where very few observations are available. Quantification with an analytical distribution is particularly important if historical worst cases in severity can be expected to be exceeded in the future, as no obvious upper boundary exists. This is the case for exploration/production spills, as the source of the spill is an essentially unlimited reservoir. Pipeline spills present an intermediate case, as the distance of the spill source to the next valve as well as time to shutdown play a role in the case of an accident. To fit the data, the Generalized Pareto (GPD) distribution was chosen as probability distribution. In the framework of extreme value theory it was shown that the probability of extreme events over a sufficiently high threshold can be modeled with this distribution.33,34 The GPD covers a large range of possible behavior of the severity distribution, from exponentially falling probability with severity, corresponding to a very low probability of very severe events to power law behavior, describing, for example, the behavior of severity of earthquakes. The GPD distribution is used to model extreme events in widespread areas such as financial markets, insurance claims or severity of natural catastrophes.35,36 In a similar context to oil spills, the GPD was also used to characterize the risk of fatalities for severe accidents in different activities along the oil chain, and found to fit the data well.37 The generalized Pareto distribution is defined by two parameters, a shape parameter ξ and a scale parameter σ . The cumulative distribution function is given by

Table 1. Results of the GPD Fit to the Exploration/ Production Data Set. For Comparison Also the Parameters for the Data Set Without the DWH Data Point Is Given data set exploration/production exploration/production w/o DWH

shape (5−95% values) 1.963 (0.838, 3.621) 1.765 (0.682, 3.431)

scale (5−95% values) 16669 (4267, 37535) 15962 (4052, 35230)

based on years 1974−2010 1974−2010

Within the model, also derived measures such as the return frequency of a given size can be directly calculated, again resulting in a distribution of values. Figure 2 shows as an example the distribution of the return frequency of an event of the size of the DWH spill, calculated directly within the model. With this method, the 5−95% percentiles are given for the parameters of the distribution as well as for the return frequency (see results in Table 2).

⎛ ⎞−1/ ξ ξ F (y ) = 1 − ⎜1 + (y − θ )⎟ ,y>θ ⎝ ⎠ σ

■

METHODS AND DATA ANALYSIS: FREQUENCY Severe accidents are rare, independent events, so that the number of accidents per year can be assumed to follow a Poisson distribution. The main parameter of interest is thus the

The function is defined for values above a threshold parameter θ. The choice of threshold has to strike a balance between a good fit to the data and retaining as many data points in the data set as possible. A detailed discussion of 13004

dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008

Environmental Science & Technology

Article

Figure 2. Distribution of the return frequency for the data set with and without the DWH data point. Without the DWH data point, the resulting distribution is broader, reflecting higher uncertainty. (Data cut off at 100 years for display).

Figure 3. Number of accidents over time for all for data sets in blue, also shown is the fitted frequency over time (red solid line) and the 95% confidence interval around the fit (red dashed line).

Table 3. Comparison of the Expected Return Frequency of an Event Exceeding the Spill Size of DWH with and without Including the DWH Horizon Data Point

mean frequency, as well as trends over time. The data sets were first tested for trends using the non parametric Mann-Kendall test.42 Significant trends were found for all data sets except exploration/production, so that for this data set no trend parameter was included in the model. The frequency was modeled for each data set i separately as a Poisson distribution with mean λi = exp(λi0) with an exponential trend with parameter λ1i for the three data sets where a trend had been found. 0

Poisson(λi , t ) = exp(λi +

return frequency

1/freq. at 668 000

5−95%

with DWH without DWH

17 23

8−91 10−177

led to a phase out of old tankers, reducing the average age of the fleet. As the results for the time trend parameter in Table 2 show, a strongly falling trend was found for the frequency of ship spills. For a detailed analysis of tanker spills see also, refs 8 and 14. Rising trends were found for storage/refinery and pipeline spills. No trend could be found for exploration/production. One reason is the generally lower number of events, so that a trend would need a longer observation period to become statistically significant. On the other hand many factors influence the risk: Drilling technology, regulation and practices in the industry change, offshore drilling continues to reach ever greater depths, new geographical locations with different conditions are explored etc. The combined effect of these developments is not obvious, thus demanding continuous close attention to the evolution of accident frequencies. For the same reason, it is not easy to predict, how an increased activity in offshore drilling would affect the expected accident frequency. Unfortunately the limited number of available data points does not allow an analysis on the drivers of the risk, e.g. the number of drilling wells, the amount of oil produced etc. Frequencies are therefore calculated on the raw data

λi1·t )

Analogous to the fit of the severity distribution, the model was also implemented as a Bayesian model with uninformative priors, yielding distributions for both the mean frequency and the trend parameters. Figure 3 shows the number of accidents over time for the four data sets as well mean and 5−95% intervals of the modeled number of accidents per year. Table 2 also gives the resulting frequencies for the end of the evaluation period, the year 2010 for the data sets with trends. Please note the frequencies are for spills >200 tons for the ship, pipelines, and storage/refinery data sets and for spills >600 tons for the exploration/ production data set, as for this data set the severity distribution was fitted to this reduced data set (see below). As the results for the time trend parameter in Table 3 show, a strongly falling trend was found for the frequency of tanker spills despite the fact, that the number ton miles have increased over time.43 An important reason can be found in the application of electronic navigation systems. Another important parameter is the IMO rule to mandate double hull tankers, that

Table 2. Frequencies for the Four Data Sets. For Comparison Also the Results for the Data Set Without the DWH Data Point Is Given frequency year 2010 [Spills/year] mean (5−95 %)

time trend parameter mean (5−95 %)

ship pipeline storage/refinery

6.597 (5.650, 7.626) 9.957 (8.106, 12.010) 6.133 (4.637, 7.849)

−0.061 (−0.067, −0.055) 0.036 (0.025, 0.047) 0.052 (0.035, 0.069)

exploration/production exploration/production without DWH

0.540 (0.353, 0.762) 0.514 (0.337, 0.724)

data set

lambda (5−95%) 4.135 (4.043, 4.226) 0.972 (0.686, 1.249) −0.121 (−0.574, 0.310)

threshold 200 200 200 600 600

13005

dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008

Environmental Science & Technology

■

■

RESULTS AND DISCUSSION: EXPECTED RETURN PERIOD OF A DEEPWATER HORIZON SIZED SPILL From the fit to the data, the expected return period R of an event of the size of DWH can be directly calculated from the graph in Figure 4 as RDWH = 1/frequency (680 000 tons). The

Article

RESULTS AND DISCUSSION: SHIP SPILLS With about 75% of spills, ships dominate both frequency and total spill volume historically. The biggest ship spill occurred off the Coast of Trinidad and Tobago in 1979 and resulted from the collision of two tankers, the Atlantic Empress and the Aegean Captain releasing 287 000 tons of crude oil in total. Such an event involving two tankers can probably be considered the worst case, that is, an upper limit to spills from tanker ships. Figure 5 shows the risk displayed as

Figure 4. Comparison of fit and data. The fit is displayed as the complementary cumulative distribution function (CCDF), that is, 1CDF, multiplied with the frequency. The data is correspondingly calculated as 1-ECDF, where ECDF is the empirical cumulated density function.

Figure 5. Empirical frequency consequence curves (1-ECDF) × frequency constructed from global historical ship spills exceeding 200 tons. The data set was split into four time intervals as the frequency of spills is found to decrease over time.

resulting return period is calculated to be 1/0.0583 = 17 years with an uncertainty interval of between 8 and 91 years (5− 95%) (see Table 3). To answer the question if such an event can be considered an outlier that significantly changed the expected risk; we also performed the same analysis on a data set where the data point of DWH had been removed. Both fits to the data are shown in Figure 4. The resulting frequency of this analysis is 23 years with an uncertainty interval of 10−177 years. Included in this difference is a ∼5% change in frequency per year, due to the removal of one event. This means, that the expected risk has not really changed and that the DWH accident can not be considered an outlier based on this global data set. The additional data point has however affected the uncertainty interval that is much narrower in the new data set, that is, 8−91 years compared to 10−177 years if the new data point is not considered (also see Figure 2). The high uncertainty is a direct result of the structure of the risk with few but very severe events. Similarly to other types of risk that follow this pattern such as earthquakes, it means that great care needs to be taken ̈ approaches to risk when historical data is analyzed. Naive analysis, such as extrapolation of trends in yearly spill volumes cannot be used for such risk, frequency and severity have to be treated separately. In addition, as we need to consider events that have a return period that is longer than the historically available data set, risk is easily underestimated if the inherent uncertainty is not taken into account. This has implications both for regulation and risk management, where risk is generally compared to acceptance criteria. To be conservative, the lower limit of the uncertainty interval around the return frequency interval would need to be considered as a basis for decision making as this worst case cannot be excluded by this statistical analysis.

empirical frequency consequence curves. As a strong downward trend in the frequency was found (see Figure 3), the data is divided into four time periods as shown in the graph to test, if also the severity distribution changed over time. To test this, the severity distributions of the four periods were tested for mutual similarity using the KS test. No significant difference in severity could be found for spills in the last three decades, so that the severity distribution for current risk can be constructed based on all data from 1981 to 2010. The data set from 1974 to 1980 showed significant differences to the later decades and is thus excluded from the severity distribution shown in Figure 6. The severity distribution could be further broken down to the distribution of tanker sizes in the global fleet and the

Figure 6. Frequency vs spill quantity for all four infrastructure subsets. DWH marks the volume of the Deepwater Horizon spill. 13006

dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008

Environmental Science & Technology

Article

(2) Federal Register FR Doc No: 2010−29944, Volume 75, Number 228 (Monday, November 29, 2010). http://www.gpo.gov/fdsys/pkg/ FR-2010-11-29/html/2010-29944.htm (accessed June 12, 2012). (3) Griffiths, S. K. Oil release from Macondo well MC252 following the Deepwater Horizon accident. Environ. Sci. Technol. 2012, 46, 5616−22. (4) Deepwater Horizon Accident. http://www.bp.com/ sectiongenericarticle800.do?categoryId=9036575&contentId=7067541 (accessed July 10, 2012). (5) BP BP and PSC Reach Definitive Settlement Agreements and Seek Preliminary Court Approval, Press Release, 18.04.2012 2012. (6) Burgherr, P.; Eckle, P.; Michaux, E. Oil tanker transportation risk: driving factors and consequence assessment. In 11th International Probabilistic Safety Assessment and Management Conference & European Safety and Reliability Conference 2012 (PSAM11 & ESREL2012); Helsinki,Finland, 2012; p 10. (7) Vanem, E.; Endresen, O.; Skjong, R. Cost-effectiveness criteria for marine oil spill preventive measures. Reliab. Eng. Syst. Saf. 2008, 93, 1354−1368. (8) Burgherr, P. In-depth analysis of accidental oil spills from tankers in the context of global spill trends from all sources. J. Hazard. Mater. 2007, 140, 245−256. (9) Jernelöv, A. The threats from oil spills: Now, then, and in the future. Ambio 2010, 39, 353−366. (10) Webler, T.; Lord, F. Planning for the human dimensions of oil spills and spill response. Environ. Manage. 2010, 45, 723−738. (11) Edwards, B. R.; Reddy, C. M.; Camilli, R.; Carmichael, C. A.; Longnecker, K.; Van Mooy, B. A. S. Rapid microbial respiration of oil from the Deepwater Horizon spill in offshore surface waters of the Gulf of Mexico. Environ. Res. Lett. 2011, 6, 035301. (12) Anderson, C. M.; Labelle, R. P. Update of Comparative occurrence rates for offshore oil spills. Spill Science 2000, 6, 303−321. (13) Price, J. M.; Johnson, W. R.; Marshall, C. F.; Ji, Z.-gang; Raineyà, G. B. Overview of the oil spill risk analysis (OSRA) model for environmental impact assessment. Technology 2003, 8, 529−533. (14) Burgherr, P.; Eckle, P.; Michaux, E. Oil tanker transportation risk: driving factors and consequence assessment. In PSAM11 & ESREL 2012; Helsinki, Finland, 2012; pp 1−10. (15) Schmidt-Etkin, D. Spill Occurrences: A World Overview. Oil Spill Sci. Technol. 2011, 7−48. (16) United Nations Environment Programme Environmental Assessment of Ogoniland; T. Davis, T. Jones, , Ed.s; United Nations Environment Programme: Narirobi, Kenya, 2011; pp 1−262. (17) Hassler, B. Accidental versus operational oil spills from shipping in the Baltic Sea: Risk governance and management strategies. Ambio 2011, 40, 170−178. (18) Giroux, J. A portrait of complexity: New actors and contemporary challenges in the global energy system and the role of energy infrastructure security. Risk, Hazards Crisis Public Policy 2010, 1, 31−54. (19) Knapp, S.; Franses, P. H. Does ratification matter and do major conventions improve safety and decrease pollution in shipping? Mar. Policy 2009, 33, 826−846. (20) Homan, A.; Steiner, T. OPA 90’s impact at reducing oil spills. Mar.Policy 2008, 32, 711−718. (21) Vinnem, J. E. Evaluation of offshore emergency preparedness in view of rare accidents. Saf. Sci. 2011, 49, 178−191. (22) Mcmahon, N.; D, P. Deepwater E & P in a global context. Analyst 2010. (23) Atlas, R. M.; Hazen, T. C. Oil biodegradation and bioremediation: a tale of the two worst spills in U.S. history. Environ. Sci. Technol. 2011, 45, 6709−15. (24) Hirschberg, S.; Spiekerman, G.; Dones, R. Severe Accidents in the Energy Sector, 1st ed.; . PSI Report No. 98-16; Paul Scherrer Institut: Villigen PSI, Switzerland, 1998. (25) Burgherr, P.; Eckle, P.; Hirschberg, S. SECURE Deliverable No 5.7.2 Final Report on Severe Accident Risks including Key Indicators; Brussels, Belgium, 2011.

proportion of oil generally released to the environment in the event of a spill. Such an analysis is however beyond the scope of this paper. Accidental tanker spills have significantly decreased over the past four decades, both in terms of numbers and volumes of spills, which is in contrast to increases in total oil movement and also to popular perceptions after recent catastrophic events.6 Key factors for these improvements include on the one hand improvements in navigation technology (e.g., GPS, Electronic Chart Display System (ECDIS) and the Automatic Identification System (AIS)), and on the other hand the conversion of the world tanker fleet from vulnerable single hull to safer double hull constructions, and accordingly a decrease in the average age of the world tanker fleet8,20,44. While the effect of the former is basically undisputed, the latter aspect has also been critically reviewed.45 Finally, the influence of flag state on tanker spills is a controversially discussed topic.8,46

■

RESULTS AND DISCUSSION: RISK COMPARISION OF SPILL SOURCES Based on the separate analysis of frequency and severity described above, the risk distributions can be constructed. Figure 6 shows the results for all four subsets of the data as a comparison. For exploration/production, the fit to the data set including the DWH accident is shown, for the other three data sets, the empirical frequency consequence curve is displayed, that is, 1-ECDF multiplied by the frequencies given in Table 2. This graph shows that while the total spill frequency is very similar for ships, pipeline and storage/refineries, the overall risk of pipeline spills is lower as it shows lower potential for very severe accidents. The total frequency of exploration/production spills is an order of magnitude lower than for the other categories, however for severe events, this category dominates the risk, as exploration/production spills differ considerably in the structure of the risk to other sources. Ship, pipeline and storage/refinery spills are more frequent than exploration/ production spills, but the latter can be much more severe, as the reservoir from which the oil is spilling is very large compared to man-made reservoirs such as tankers or storage facilities. The results of this top down approach to risk analysis based on global data sets covering long time periods can now be used as a complement and benchmark to bottom up engineering risk analyses of individual installations, as well as an input to consequence analysis for environmental risk assessment

■

ASSOCIATED CONTENT

S Supporting Information *

Table 1. This material is available free of charge via the Internet at http://pubs.acs.org.

■

AUTHOR INFORMATION

Corresponding Author

*Phone: +41 (0) 56 310 2649; e-mail: [email protected]. Notes

The authors declare no competing financial interest.

■

REFERENCES

(1) Anderson, J.; Burkeen, A. D.; Clark, D.; Curtis, S.; Jones, G.; Kemp, R. W.; Kleppinger, K. D.; Manuel, B.; Revette, D.; Roshto, S.; Weise, A. DEEP WATER: The Gulf Oil Disaster and the Future of Offshore Drilling; National Commission on the BP Deepwater Horizon Oil Spill and Offshore Drilling, 2011. 13007

dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008

Environmental Science & Technology

Article

(26) Burgherr, P.; Hirschberg, S.; Hunt, A.; Ortiz, R. A. Severe accidents in the energy sector. Final Report to the European Commission of the EU 5th Framework Programme “New Elements for the Assessment of External Costs from Energy Technologies” (NewExt), 2004. (27) Burgherr, P.; Hirschberg, S.; Cazzoli, E. Final Report on Quantification of Risk Indicators for Sustainability Assessment of Future Electricity Supply Options, NEEDS Deliverable no. D7.1 - Research Stream 2b. NEEDS project 2008. (28) Burgherr, P.; Eckle, P.; Hirschberg, S. Final Report on Severe Accident Risks including Key Indicators, 2010 (29) Burgherr, P.; Hirschberg, S. A comparative analysis of accident risks in fossil, hydro and nuclear energy chains. Hum. Ecol. Risk Assess. 2008, 14, 947−973. (30) McMahon Anderson, C.; Labelle, R. P. Comparative occurrence rates for offshore oil spills. Spill Sci. Technol. Bull. 1994, 1, 131−141. (31) Smirnov, N. . Tables for estimating the goodness of fit of empirical distributions. Ann. Math. Stat. 1948, 19, 279−281. (32) United Nations Conference on Trade and Development Review of maritime transport 2011 2011. (33) Pickands, J., III Statistical inference using extreme order statistics. Ann. Stat. 1975, 3, 119−131. (34) Balkema, A. A.; de Haan, L. Residual Life Time at Great Age. Ann. Probab. 1974, 2, 792−904. (35) Coles, S.; Casson, E. Extreme value modeling of hurricane wind speeds. Struct. Safe. 1998, 20, 283−296. (36) Embrechts, P.; KlüppelbergC.; MikoschT. Modelling Extremal Events for Insurance and Finance; Springer: New York, 1997 (37) Eckle, P.; Burgherr, P. Bayesian data analysis of severe fatal accident risk in the oil chain. Risk Anal. 2012. (38) Tancredi, A.; Anderson, C.; O’Hagan, A. Accounting for threshold uncertainty in extreme value estimation. Extremes 2006, 87− 106. (39) Lunn, D.; Spiegelhalter, D.; Thomas, A.; Best, N. The BUGS project: Evolution, critique and future directions (with discussion). Stat. Med. 2009, 28, 3049−3082. (40) Gelman, A.; Carlin, J. B.; Stern, H. S.; Rubin, D. B. Bayesian Data Analysis; 2nd ed.; Chapman & Hall/CRC: Boca Raton, FL, 2003. (41) Coles, S. G.; Powell, E. A. Bayesian methods in extreme value modeling: A review and new developments. Int. Stat. Rev. 1996, 64, 119−136. (42) Mann, H. B. Nonparametric Tests Against Trend. Econometrica1 1945, 13, 245−259. (43) ITOPF Oil Tanker Spill Statistics 2011; London, UK, 2011. (44) Yip, T. L.; Talley, W. K.; J., D. The effectiveness of double hulls in reducing vessel-accident oil spillage. Mar. Pollut. Bull. 2011, 62, 2427−2432. (45) Devanney, J. The Tankship Tromedy: The Impending Disasters in Tankers; The CTX Press , 2010 (46) Alderton, T.; Winchester, N. Globalisation and deregulation in the maritime industry. Mar. Policy 2002, 26, 35−43.

13008

dx.doi.org/10.1021/es3029523 | Environ. Sci. Technol. 2012, 46, 13002−13008