Using Predictive Analytics for Weather and Process Predictions

Author photo: Janice Abel
ByJanice Abel
Category:
Technology Trends

The US mainland and Puerto Rico recently experienced several massively destructive hurricanes. Looking at the predicted hurricane paths, I started researching the hurricane models (on my own time) as well as some of the predictive algorithms used to determine the path and position of hurricanes. Frequently, the predictions are dangerously inaccurate. But why? I know that deep learning algorithms require lots of data and in both cases, the weather, the data and conditions are always changing.

The US has several models but mostly relies on the Global Forecast System (GFS), run by the National Centers for Environmental Prediction, part of the National Oceanic and Atmospheric Administration. earliest GFS reports for Irma predicted it would hit Miami and track up the East Coast, bringing winds and heavy rains all the way to North Carolina. The European model, known as the ECMWF, predicted Irma would slam into the middle of Florida and then change to the west coast of the state, which turned out to be the case.

The European hurricane model also did a much better job of predicting Harvey’s track through Houston last month. European model factors in the immediate impact of winds and ocean water temperatures. This model has also more accurately predicted other storms, such as Sandy in 2012 and several northeasters. models have been improving over the past couple of years, but still are not accurate enough. Overall, predictions made four days in advance of landfall end up averaging about 175 miles off the mark.

predicting the weather predicting the process.jpg

ECMWF was the most accurate model for IRMA. Chart shows mean absolute errors for different hurricane models/ Source: Brian Tang, Atmospheric Scientist, University of Albany

nhc models have been improving.png

NHC models have been improving: Source: hurricanescience.org

Why is the European Model More Accurate?

Twenty-two nations support the ECMWF model, developed by the European Centre for Medium Range Weather Forecasts consortium. One reason the model is more accurate, is because the infrastructure used for the predictive algorithms includes the largest supercomputer complex in Europe and the world’s largest archive of numerical weather prediction data.

The US GFS was upgraded in 2017 – just before the start of the hurricane season, and perhaps new algorithms still need to be adjusted and need further learning form the data to be as accurate as the 12z model. According to the experts, the European model works on the most advanced computer hardware and has devised a system to assimilate real-time meteorological observations into its model to improve accuracy. Using real data, the model runs with very accurate initial conditions.

Like an industrial process, the behavior of the atmosphere is governed by physical laws that can be expressed as mathematical equations. For weather forecasts, the equations used include measured variables such as temperature, wind speed, direction, humidity, etc. For both industrial process prediction and weather prediction real-time measurement data is often integrated with algorithms generated from historical data. In both cases, the algorithms are about 90 to 95 percent accurate, and small changes in the initial conditions can have a major impact on the results.

So why the inaccuracy?

There are many reasons given for inaccuracies, including:

  • Lack of computing power
  • Accurate initial data (garbage in/garbage out) – small changes make a huge difference
  • Inaccurate computer models
  • Model or even human bias

Fortunately, the US is continuously updating and improving its weather models so, hopefully, someday we will have extremely accurate hurricane predictions.

For both weather and process predictions, we need more data and algorithms and additional ‘data learning’ to predict potentially hazardous conditions. Like complex weather models, process models are starting to get better and are more accurate and useful than even 15 years ago. Even though accurately predicting a hurricane may be more complex than modeling a process, the future looks promising and we should be able to accurately model both with a small margin of error. The future of operational analytics depends on data and more accurate predictive analytics.

 

Engage with ARC Advisory Group

Representative End User Clients
Representative Automation Clients
Representative Software Clients