What an ‘intelligent hurricane’ can uncover

When detailed data can be captured about interconnected infrastructure systems, utility managers will be able to make better informed decisions in times of crisis. Kristen Mally Dean highlights research from the US Department of Energy’s (DOE) Argonne National Laboratory, where scientists are studying how optimisation theory and machine learning might better uncover meaningful patterns relevant to infrastructure sciences.

When Hurricane Maria hit the US territory of Puerto Rico in 2017, the archipelago endured a severe test of its energy systems’ resilience. With all of the electric power for the commonwealth supplied by a single major electrical system, the widespread damage to this system’s electrical transmission lines immediately cut power to thousands of people.

Like most places, Puerto Rico’s critical infrastructure systems are highly interconnected, meaning that the operation of one system can affect the operation of another. So significant portions of the lifeline critical infrastructure systems supporting society – clean drinking water, wastewater management, natural gas, communications, and transportation – were hobbled or, in some cases, completely shut down when the electrical system failed.

As an isolated, short-term interruption, one of these losses would likely be an inconvenience. Collectively and simultaneously, the cascade of failures across multiple, interconnected systems became life-threatening.

Have you read?
Forecasting the unpredictable… Whatever the weather
IoT + edge processing is making sci-fi a reality for energy management

Scientists at the US Department of Energy’s (DOE) Argonne National Laboratory are studying how optimisation theory and machine learning – an artificial intelligence technique that learns through repetition – might better uncover meaningful patterns relevant to infrastructure sciences.

Josh Bergerson, a principal infrastructure analyst in Argonne’s Decision and Infrastructure Sciences (DIS) division, was part of a large Argonne team that supported the US Federal Emergency Management Agency’s recovery efforts in Puerto Rico. He and the broader team were intent not only on helping local operators and decisionmakers restore and recover operations, but also on identifying and prioritising systems that would benefit most from improvements in security and resilience. The key to doing so required a closer examination of system interconnectedness and higher order dependencies.

“Dependencies refer to what you need to maintain operations. They are not necessarily within your control, and you might be relying on other businesses to maintain control,” said Bergerson. “Most operators know what they need in order to maintain their own operations, but while a water system operator might know who provides their electrical power, they might not understand what the electric company needs to maintain its operations. They generally aren’t privy to these higher order dependencies.”

You need hundreds of hurricanes to develop machine learning models

One reason for this gap in understanding is that no authoritative dataset of interdependencies generally exists.

“When you look at critical infrastructure, such as electrical power, there are all the rules and regulations [that make it a reliable and planned for system],” explained Bergerson. “When you look at water, it also has its own rules and regulations. But, when you look outside of a system, you don’t have anyone whose job it is to collect data on interdependencies. Someone in the field understands this power station is connected to that water system, but finding that person can be like finding the needle in the haystack.”

Bergerson and his fellow researchers from DIS developed a tool – the Puerto Rico Infrastructure Interdependency Assessment (PRIIA) – to help identify an area’s most critical assets and how widespread the impact could be should they fail.

In Puerto Rico, where there were tens of thousands of dependency connections, populating a useful dataset for this purpose through interviews or field data collection was a gargantuan, inefficient undertaking. The DIS team quickly recognised a need and an opportunity to enhance the PRIIA tool by collaborating with colleagues from Argonne’s Mathematics and Computer Science (MCS) division.

“Before we began working with MCS, [using PRIIA] was a manual process,” said Bergerson, who described the initial approach as theoretically turning off a single facility and examining “what if” outcomes. “With the tool’s enhancements, we are now able to do automated search and identification of the most critical systems using game theory. This empowers us to answer the question ‘what pieces, when turned off, cause the greatest failure?’”

An ‘intelligent hurricane’

Optimisation models were the first techniques used to enhance the PRIIA tool. Very effectively, they helped identify the infrastructure pieces whose disruptions had the most impact.

“You need hundreds of hurricanes to develop machine learning models,” said Sven Leyffer, a senior mathematician in MCS. “But we don’t want to subject the people of Puerto Rico to hurricanes to populate models. Optimisation allows us to simulate events with artificial data. We can be an ‘intelligent hurricane’ – a very mischievous hurricane – that hits where we want it to and that has the most impact.”

Such a “disaster on a chip” allows the scientists to propagate, study, plan for and potentially mitigate against real events.

Next, Bergerson and Leyffer collaborated to build a rudimentary “digital twin” of the systems in Puerto Rico to measure how, exactly, cascading failures spread. The digital twin is a simulation programme, or set of programmes, that is supposed to behave exactly as its real-life equivalent. It’s an important step toward addressing the complex problems of when, how, and at what level lifeline services are disrupted.

While the programme is still in its infancy, the researchers hope to further develop a more accurate and sophisticated version to capture higher-level details of interconnected systems. This would be valuable to utility managers and decision makers in circumstances outside of extreme weather or natural disaster. This might include cyber or physical attacks by malicious actors, or a major supply chain interruption. Such complex scenarios can threaten lifeline
infrastructure systems that could widely disrupt infrastructure, as occurred following Hurricane Maria. The machine learning tools developed to address the hurricane’s impact may help decision makers prepare, restore and repair systems vulnerable to many catastrophic outages.

“A digital twin is agnostic to where data comes from,” Leyffer explained. “It’s a bit like the camera on your phone. The camera can take pictures of lots of things, not just selfies or pictures of another face. The functionality [of a digital twin] can apply to other events or locations.”

Furthermore, an increasingly prevalent Internet of Things (IoT) – the great variety of networked devices that now collect and control information – may present additional opportunities to build better models of the systems and to help researchers deduce which parts of the systems are interconnected. The IoT will make it easier and faster to do or know things. For example, a sensor signal will communicate more quickly with another station to report a system disruption than a human operator, who must physically place the call. Yet researchers will need to balance that improvement in operating efficiency with security risks if that sensor can be deliberately misused.

The Internet of Things will make it easier and faster to do or know things

Ultimately, the IoT may generate such a grand volume of information that researchers will face a question with which Bergerson is already familiar: How does one collect, manage, store, protect, and ultimately use all that data?

For now, Bergerson is pleased with the improved modeling of the PRIIA tool.

“It is intended to enhance our ability to identify and evaluate dependencies,” said Bergerson. “We’ve taken what we started with in Puerto Rico and significantly improved that to move beyond a simple ‘what if’ view into a more probabilistic nature. Instead of estimating the most likely configuration of the network, we are running hundreds of configurations to enhance and make more robust our estimations of the probability of the configuration of the network.”

A hurricane may have sparked the collaboration between the infrastructure science group and the mathematics and computer science group, but the conversations happening after the storm are bringing exciting new possibilities to light.

“Our initial collaboration was on a project that was interesting but limited, and now we are finding a wealth of opportunities and research questions that we are working out together,” said Leyffer. “We are pushing through to new ideas.”

About the author

Kristen Mally Dean writes and edits for Argonne National Laboratory as well as other organisations at the forefront of science, business, art and technology. In addition to contributing to the publication of Argonne’s comprehensive AI for Science Report in 2020, she has written on various topics in decision and infrastructure science, energy and global security, and environmental science.