Risk Characterization of the Electric Grid to Cascading Blackouts
Testing of the electrical grid for robustness to blackouts is currently performed in which one outage is simulated at a time to determine whether the failure of one component might lead to a cascading blackout in so-called “N-1 contingency testing”. This testing ignores the scenario that leads to cascading blackouts in which multiple contingencies occur at approximately the same time. These N-k contingency scenarios (where k is the number of combined outages that combine to create a cascade) are not tested for values of k greater than one because of the combinatorial explosion that results from choosing k elements from N for large N. Thus the combinatorial search space for combinations of contingencies is too large to search in a reasonable amount of time to ensure robustness to these N-k contingencies of even a medium-sized electrical grid. In our work, we use the Random Chemistry algorithm to quickly identify a subset of all combinations of contingencies that lead to cascading blackouts in a cascading failure simulator. This algorithm gives us an unbiased sampling of branch combinations that cause cascading failure faster than if we were to employ a random search on the combinatorial space. We then use this subset of N-k contingencies to estimate the number of expected total number of outage subsets that would lead to cascading failure of the grid. Using this estimate, we develop a risk statistic that can be used to characterize the reliability of a given electrical grid configuration to cascading blackouts.
Ananth Iyer
Faculty: Project Co-PI
Does this model also permit contingent actions (such as reacting to failures by shutting off power demand points) to be included in assessing risk ?
Mark Wagy
The model allows for that type of contingent action to be simulated, however we are not yet incorporating it into our risk statistic.
Paulette Clancy
Faculty: Project Co-PI
Why did you choose the Random Chemistry Model over other possible representations? What particular characteristics of this model were seen as especially relevant for modeling cascading blackouts?
Mark Wagy
Random Chemistry is an algorithm that allows us to more quickly find branch outages that might result in a cascading blackout. So Random Chemistry is not the model of the cascading blackouts, but rather an algorithm that we can use to quickly find branch combinations that might cause cascades. We use Random Chemistry because it is a method of finding these branch combinations more quickly than other methods.
Ranjit Koodali
Faculty: Project Co-PI
Can this model be used to predict brownouts?
Mark Wagy
My understanding is that brownouts are intentional, human-initiated outage events; as such, this model might not be an appropriate tool to use in their prediction.
Ranjit Koodali
Faculty: Project Co-PI
Thanks!
Ian Harrison
Faculty: Project PI
What algorithm is currently used by grid operators to prioritize repair of potentially malignant branch outages given the new NERC guidelines for n-k contingency testing? Could you explain the particular advantages of your algorithm? (I’m assuming the current NERC suggested method is not brute force computing…)
Mark Wagy
Just as a clarification: the Random Chemistry algorithm was developed by Maggie Eppstein (indicated in the poster reference), which was inspired by the concept of Random Chemistry by Stuart Kauffman – I am not claiming it to be my own.
In answer to your questions, I believe that the grid operators do perform brute-force testing for n-1 testing since it is feasible to do that many simulations (for n-1, they are not burdened by the combinatorial search space) and then focus on testing a small subset of the n-k outage scenarios for k greater than one (though this is with the intent not necessarily to repair outages but rather to adjust load to prevent possible cascading outages). To my knowledge, full n-k testing is not currently performed to characterize all malignancies for k>1. The Random Chemistry algorithm makes the discovery of n-k outages (particularly for k values of, say, 2 and 3) feasible in a shorter number of simulation evaluations than using either brute force or Monte Carlo-based methods for finding possible malignant subsets.
Ian Harrison
Faculty: Project PI
Thanks, this does seem to be computationally interesting problem.
Matthew Yates
Faculty: Project Co-PI
Can you explain some of the actions a grid operator may take based on your results?
Mark Wagy
A grid operator might, for example, adjust the dispatch of generators in response to a high risk characterization of the grid to a state of lower risk to cascading blackouts.