Risk Assessment Oil and Gas
How Do Data Quality and Quantity Affect Risk?
Download 361.05 Kb. View original pdf
|
ADNOC Toolbox Talk Awareness Material 2020, ADNOC Toolbox Talk Awareness Material 2020, TRA-Installation of Field Instruments, Road Maintenance Plan & Status-Map Format
| 3.5.3 How Do Data Quality and Quantity Affect Risk? A risk calculation is not just a model prediction, and so a real risk assessment will behave rather differently from a simple predictive model as its information input is reduced. Indeed, a simple predictive model may well predict a more favorable outcome when it is given less detailed or less precise input information. That is because the model prediction is a single scenario: the best estimate of the outcome. Furthermore, these kinds of models have “out of sight, out of 29 mind” behavior. With less detailed input information, the model will default to assumptions of intermediate input values and will therefore deliver an intermediate output prediction. An example would be the way a kriging model defaults to the sample average when there are no predictor data close enough to the location being predicted. Obviously, this would not be satisfactory for a risk assessment, because our ignorance of the input values does not really make the risk go away. The way risk assessment takes account of our uncertainty about input values is to represent every uncertain input as a probability distribution. The greater our uncertainty about the input, the broader that distribution. Now, the uncertainty about inputs gets propagated through the prediction component of the risk assessment. The output of the risk assessment is also a probability distribution, showing the distribution of outcomes. The greater the uncertainty about inputs, the greater the uncertainty about outcomes. The uncertainty about outcomes is reflected in the spread of the probability distribution that describes the result. If there is little uncertainty, the probability distribution is concentrated over a narrow range of outcomes, and then the acceptability of the risk depends merely on whether the outcomes in that narrow range are themselves in a range that we consider acceptable. But if there is great uncertainty, the probability distribution is spread over a broad range of outcomes. If the spread is great enough, a considerable portion of the distribution can “spill over” into a range of unacceptable outcomes, even if the peak is centered over outcomes that are acceptable. This would describe a situation where the “best estimate prediction” is an acceptable outcome, but the risk is still unacceptable because the uncertainty leaves too high a probability of outcomes that are far from the best estimate. In very conventional applications of statistics to risk calculations, the uncertainty about inputs is factored into the calculation through the use of confidence limits. For example, at a hazardous waste site, the decision about whether a given unit needs to be remediated may depend on whether the upper 95% confidence limit of the average concentration of containment exceeds some tolerance level. Use of the upper 95% confidence limit factors in the uncertainty. The greater the uncertainty (owing to small sample size or variability among the sample values), the higher the upper 95% confidence level will be above the sample average. In this way, the uncertainty exerts a kind of penalty in the decision process, forcing a greater margin of safety as the uncertainty increases. (This is not necessarily the best way to take uncertainty into account in a risk decision process, but it is one that many people are familiar with, and it is sufficient to illustrate the point.) In a similar way, we might consider a map of our oil development site and ask what portions of the site should be off limits to road construction because the risk is “too high.” If our 30 input information were of low resolution, a good deal of the off-limits area would be so classified only because our uncertainty about those locations was so high that the equivalent of the “upper 95% confidence level” would spill over into the range of unacceptable values. Basically, the map would contain a lot of gray areas that would be classified as high risk. With better input information (more spatial resolution, more precision) many of the gray areas on the map would be resolved into black and white. In the black and white areas the predictions are more certain, and some of the gray area that was previously classified as high risk because of the uncertainty will be reclassified as low risk, and more of the remaining areas that are still classified as high risk will be so classified because of a reasonably secure prediction that the outcome there really will be unacceptable. The high-risk gray area of the map shrinks as the input information improves. |