Weather Forecasting Using METAR Inputs Through Machine Learning: Good Forecasts at Lower Computation and Data Acquisition Cost

[]
Modern weather forecasting is a computationally intensive activity, relying on expensive data sources like satellites and Doppler radar. This presentation explores an alternative, low-cost approach using machine learning to predict local weather conditions. By leveraging historical METAR data (routine reports published by airports), we can train a deep neural network to find predictive patterns that are difficult to discern with other methods.
This talk will detail the complete project lifecycle: from acquiring and cleansing over 7.5 million METAR records to the design, training, and tuning of the neural network using Pylorch. An emphasis is placed on explaining the steps generally used to develop machine learning applications. We will cover feature engineering, model architecture, and the results, which show the model can predict temperature 30 hours out with a Mean Absolute Error of about 2.6°C. While not matching the precision of state-of-the-art global systems, this method produces a reasonably accurate forecast using minimal resources, demonstrating the power of machine learning to create effective solutions without the need for supercomputers. At the close of the talk, we will demonstrate the tool in its current form.
This presentation will count for 1 Professional Development Hour (PDH) for the PE License in Wisconsin and Michigan.
[]
Speaker(s): Greg Hawley
Agenda:
5:00 Featured Speaker – Greg Hawley
D.J. Bordini Center at FVTC
[]
6:30 Adjourn to Good Company
110 N. Richmond St.
Appleton, WI
6:45 Social time, cash bar open
Order meals from the restaurant menu
7:00 Section announcements, door prize drawing
Room: BC112B, Bldg: D.J. Bordini Center at FVTC, 5 N. Systems Drive, Appleton, Wisconsin, United States, 54914