Machine learning technology powers analog weather forecasts

Machine-learning techniques that can recognize human faces could also help improve weather forecasts, according to a team of scientists.

Machine learning scientist at the University of San Diego and former PhD student at Penn State University.

Scientists have applied deep learning algorithms to analog weather forecasts that use past weather conditions to make future predictions. In this case study, using machine learning to analyze predictions of surface wind speed and solar irradiance in Pennsylvania from 2017 to 2019, he found that the accuracy of analog predictions improved. .

“You’re going to want to know what your energy will be for the day,” said Hu, who has a doctorate in geography from Pennsylvania State University. “The important thing to understand about risks is that whether you overestimate or underestimate, penalties such as power shortages and overproduction will occur. It shows that we can improve the accuracy of our predictions.”

Analog forecasting is an alternative to numerical weather forecasting (NWP) that uses computer models to simulate how initial weather conditions will change in the days or weeks ahead. Although the NWP has made great advances in forecasting over the last few decades, uncertainties remain.

These uncertainties are partially addressed by running a number of simulations called ensembles. The simulations show a range of possible future atmospheric conditions, but they are also computationally intensive and expensive to create, the scientists said.

“But with analog prediction, you can generate ensembles without running expensive model iterations,” says Hu. “It works by finding the past predictions that are most similar to the target prediction, and the past observations associated with the most similar past predictions make up the members of the ensemble.”

Analog ensembles combine deterministic forecasts (one very detailed run of the NWP model) with past weather observations (temperature, pressure, humidity, etc.) from past forecasts that are similar to the current forecast. is generated.

The best similar forecast is selected based on a similarity metric that weighs individual weather predictor variables, but the process uses a constrained exhaustive search that limits the number of predictor variables that can be used to reduce the Relationships between variables are not considered.

“This was the limit of analog ensemble prediction,” says Hu. “In this paper, we try to address this problem by introducing a machine learning approach that learns the complexity between predictor variables.”

Scientists say the machine-learning technique takes every weather variable, such as temperature, pressure, and humidity, and turns them into a latent space — clustered patterns that help select ideal forecasts and similarities. It is said to convert to

“This approach attempts to identify the most useful features to look for to improve analog predictions,” Hu said. “Simply put, clustering the candidates yields the most accurate predictions and eliminates low-similar data points from low-similar predictions.”

The machine learning technique overcomes the computational limitations posed by predictor weight optimization in traditional analog ensemble prediction, the scientists said.

“Machine learning has been operational for many years to make predictions faster and more accurate, but its role has been largely limited to post-processing and data preparation,” says Geography and Meteorology at Pennsylvania State University. Atmospheric Sciences Professor Guido Cerbone said. Advisor and co-author of papers. “It’s been a year or so here he has machine learning as the core of algorithms, often even replacing numerical model solutions.”

Research results published in journals boundary layer meteorologysuggesting that machine learning allows more predictors to be used, producing predictions with higher accuracy.

“Our work shows that machine learning models can also be used in the geosciences to probe complex features,” Hu said. “Earth science deals with hundreds of variables. In our search, we had over 300 variables, and most of the time they are highly correlated. We have shown that we can indeed detect all these relationships from large datasets.”