It’s like thunder, lightning, the way you code me is frightening

Machine learning algorithms can predict when and where lightning will strike, according to new research published in npj Climate and Atmospheric Science.

That might sound impressive at first, but the study is still pretty crude considering its limits. The forecast can only guess that a thunderbolt will flash somewhere within a 30 kilometer radius (~18 miles) and at some point within 30 minutes of it touching down.

“Current systems are slow and very complex, and they require expensive external data acquired by radar or satellite,” said Amirhossein Mostajabi, co-author of the paper and a PhD student at the Swiss Federal Institute of Technology Lausanne.

“Our method uses data that can be obtained from any weather station. That means we can cover remote regions that are out of radar and satellite range and where communication networks are unavailable.”

Lightning strikes are, the paper claims, the second leading cause of death under “natural hazard processes” in Switzerland, after snow avalanches. A total of 1,023 people died from being struck by lightning from 1946 to 2015. The zap of electricity can also damage crops and set fire to buildings.

The team of researchers began the study by collecting data from 12 weather stations that logged lightning strikes in Switzerland from 2006 to 2017. They analyzed four weather variables: local air pressure, temperature, humidity, and wind speed, which affect how much moisture is available for a brewing thunderstorm.

Next, they split the data for training and testing. “Once the database was formed, pattern recognition and data mining algorithms were employed to identify regularities between predictors and responses using a portion of the data which, as mentioned above, is called the training set,” the paper said.

We trained an AI to predict how bad a forest fire will be. It’s just as good as a coin flip!

READ MORE

The machine learning algorithms learned to pick up on the combination of variables that were most likely to lead to lightning strikes. After training, they were fed the same four inputs to predict if a bolt will flash within a 30 kilometer radius around the weather station. Results revealed that the algorithms could predict a strike to within ten minutes of it occurring to over 71 per cent accuracy.

The model struggles with false positives, however, and tends to overpredict when lightning will strike. Deploying such a system in the real world would probably lead to a lot of false alarms.

“The main challenge in developing the appropriate predictive scheme was the high imbalance seen between lighting-inactive and lighting-active classes. The situation gets even worse when the lead time is increased,” the researchers admitted in the paper.

But since the model only requires monitoring the air pressure, temperature, humidity, and wind speed surrounding the weather station, it’s probably not too difficult for the researchers to test their algorithms in real time using sensors. Hopefully they make sure everything’s grounded first. ®

Sponsored: Disrupting with data – challenges for digital native organizations