Evacuation modeling: finding the best time (and way) to get going

As a result, 200km² of surrounding land would be inundated with fast-moving water, threatening the lives of about 70,000 residents. An increasing population, combined with the silent threat of the Warragamba Dam spilling or failing, make this region one of the worst flood risks in Australia.

To evaluate the impact of major flooding on regional areas, we developed a tool to simulate the evolution of the flood, minute by minute. The algorithm forecasts a 24-hour flood in a few minutes and specifies which areas need to be evacuated and at what time, and when roads will be flooded.

The optimization algorithm then uses all this information to produce an evacuation plan indicating when each person needs to evacuate, where he or she should go, and route to follow.

Contrary to human decision makers, the optimization algorithm considers simultaneously all 70,000 evacuees, how they interact and compete for road capacities, and how the flood affects the transportation network. The whole plan is produced in a matter of seconds and rescheduling can be performed in real time as the event unfolds.

The video above illustrates the evacuation of 70,000 persons in the Hawkesbury Nepean area. The flood (in blue), rising from the Warragamba dam (at the back) inundates the flood plain. Vehicles (in green) are evacuated following precise evacuation routes and schedules to shelters (green boxes).

In this illustration, emergency services wait as long as possible before giving evacuation orders in order to avoid false alerts.

Lessons to learn
Our algorithms have led to fundamental insights about evacuations and human behavior. Letting every individual decide when, where and how to evacuate can have disastrous consequences.

In the Hawkesbury Nepean flood scenario, if everyone leaves at a reasonable time and goes to the closest evacuation shelter, more than 60 percent of the population will not reach safety and will end up trapped by major traffic jams.

It is not surprising that independent decisions by 70,000 individuals do not lead to an effective evacuation — but our algorithms evacuate every single person.

Even better, if as many as half of the population does not follow the plan exactly but leaves at the time it is instructed to, 97 percent of evacuees will reach safety.

Equally interesting is the fact that most existing evacuation algorithms are too optimistic: they delay the evacuation too much and, as a result, a substantial portion of the population cannot be evacuated.

After the 2005 Hurricane Katrina, the U.S. recognized the need to go beyond situational awareness and to adapt for disaster management the optimization algorithms used in airlines, logistic systems, and supply chains.

Technological advances may have significant impact, and save numerous lives, in Australia and the Asia-Pacific region. The next step, of course, is convincing more governments around the world to deploy them.

Further reading: A Conflict-Based Path-Generation Heuristic for Evacuation Planning

Victor Pillac is researcher at National ICT Australia (NICTA); Pascal Van Hentenryck is Optimization Research Group leader at (NICTA). This story is published courtesy of The Conversation (under Creative Commons-Attribution/No derivatives).