Our team is reproducing the paper “Residential Links Under the Weather,” authored by Padmanabhan, Schulman, Levin, and Spring. The original paper ran their proprietary software ThunderPing for eight years, following the forecasts of weather in the U.S. and pinging up to 100 hosts from each last-mile provider in the area for six hours before, during, and six hours after a forecasted weather event. This paper was the first attempt to quantify the effect of weather on residential outages.
The problem is important because severe weather is the number one cause of power outages in the U.S. and costs the economy billions of dollars a year in lost output and wages, spoiled inventory, delayed production, inconvenience, and damage to the grid infrastructure. With the likelihood of severe weather events likely to increase in the future due to climate change, this question becomes even more pressing to answer. For exampple, heavy downpours have increased nationally, especially over the last three to five decades; since 1991, the amount of rain falling on the heaviest rain days has been significantly above average. There have been increases in flooding events in the Midwest and Northeast. Winter storms have also increased in frequency and intensity since the 1950s .
The original team showed that a variety of weather conditions inflated the likelihood of Internet dropouts, and that this inflation depends upon the type of weather, link type, and geographic location. Additionally, the team found that the time to recover from a dropout increases during weather events. The datasets and analyses produced by the team will enable a wide range of future studies into areas including the combined effects of wind speed and rain, the effects of cloud cover, the relationship between duration of a weather condition and recovery times, and much more.
To reproduce this original paper, our team first replicated several early figures in the paper using a year of data provided by the original team. After verifying that this data looks accurate, we searched for independent data sets, settling on the ANT Internet outage data and several datasets on snow and rain events in the U.S. We combined these datasets to try and produce similar trends to those found in the original paper.
Our graphs using the original data matched the original figures well.