Researchers at the University of San Diego and Penn State are harnessing machine learning technology to improve analog weather forecasting. Analog forecasting uses past weather conditions to create future predictions, while numerical weathering prediction relies on computer models. Machine learning technology can recognize human faces, and these scientists have found it can be applied to improve accuracy of surface wind speed and solar irradiance forecasts.
This breakthrough comes from the scientists comparing weather to historical forecasts and applying a deep learning algorithm to the process. They discovered that using machine learning for analog forecasting allowed them to search for more predictors than a traditional approach, which they claim results in more accurate forecasts. This process of clustering candidates for forecasts also helps to identify the most helpful features to look for in order to improve accuracy.
The results of the study, which was published in Boundary-Layer Meteorology, suggest machine learning has a significant potential to improve both the speed and accuracy of weather predictions. Guido Cervone, professor of geography and meteorology at Penn State and Hu’s advisor and co-author of the paper, commented “It is really during the last year or so that machine learning has been used as a central core of algorithms, often even replacing numerical model solutions.”
Weiming Hu, a machine learning scientist at the University of San Diego and a former doctoral student at Penn State, is the primary author behind this work. He received his doctorate in geography from Penn State and is passionate about understanding energy production risk factors of short-term forecasts. Hu believes that machine learning could offer a key solution to weather forecast uncertainties, particularly when it comes to creating ensembles with limited computational power.
As an alternative to traditional numerical weathering prediction (NWP), analog forecasting provides ensembles without repeating expensive model runs. While NWP has made strides in forecasting over the last couple of decades, Hu and his team strive to make even further advances.
Google’s FaceNet, which the Machine learning technology used in this study was inspired by, offers an additional benefit in its ability to compare human faces to a database of images. Similarly, Hu and Cervone hope their use of facial recognition on weather forecasts will also lead to improved accuracy and more efficient predictions.
