Conversations About AI: Part 7 – Your Next Weather Forecaster May Be An AI

With the world experiencing warmer daytime highs and nighttime lows, and headlines linking the increased heat to climate change, the weather, always a subject of conversation, is becoming a headline grabber. This is putting human meteorologists under a microscope as they attempt to divine future forecasts using satellites, radar, weather station data, and older rules of thumb such as reading barometers and looking up at the sky.

Getting the weather forecast right this summer seems to have gone by the wayside where I live. Accuracy appears to be taking a holiday. So it was interesting to come across articles about using artificial intelligence (AI) to forecast the weather, doing it with speed and accuracy.

Why are speed and accuracy so important? Munich RE, the reinsurer that underwrites millions of policies, has predicted that extreme weather events by 2050 will result in 1 million lives lost and USD 1.7 trillion annually. That’s why weather forecasting operations use supercomputers today to estimate probabilities.

So meteorologists now have competition beginning with these three companies: Alphabet, Huawei and Nvidia. Each is using AI to model the weather based on historical data. And each claims to be doing it better and faster than any human meteorologist. So how are they doing this? Here’s a deeper look.

Alphabet’s Google DeepMind 

The company has developed an AI in partnership with the United Kingdom’s Met Office, one of the most advanced meteorological services on the planet. The AI is called Deep Generative Models of Rainfall or DGMR and it is doing nowcasting which constitutes high-resolution, short-term predictions of precipitation. It seems fitting considering the UK’s reputation as a pretty wet place that the AI has first focused on rain.

DGMR has trained by studying historic radar images of rainfall allowing it to predict future precipitation patterns by location and amount. Early implementation has DGMR being used to help with air traffic control and energy management.

The model is a conditional generative adversarial network that takes four observed radar frames to predict the next 18 output frames. The historical radar observations used to teach the AI cover a period from 2016 to 2019. And DGMR takes just over a second currently to generate a forecast. It is proving to be 93% accurate.

Huawei’s Pangu Weather AI

Pangu-Weather AI is a forecaster that predicts not only the weather but also the path of tropical cyclones. Huawei, the Chinese telecommunications company, developed the model using its CLOUD team. The deep learning AI studied 43 years of weather data. Recently, it accurately predicted the trajectory of Typhoon Mawar five days before the storm changed course near Taiwan.

In an interview, Tian Qi, Chief Scientist at Huawei CLOUD described the complexity that is involved in developing accurate AI weather predictors.

He stated, “Weather forecasting is one of the most important scenarios in the field of scientific computing because meteorological prediction is a very complex system, yet it is difficult to cover all aspects of mathematical and physical knowledge. AI models mine statistical laws of atmospheric evolution from massive data.”

The current Pangu

At present, Pangu-Weather mainly completes the work of the forecast system, and its main ability is to predict the evolution of atmospheric states. Our ultimate goal is to build a next-generation weather forecasting framework using AI to strengthen existing forecasting systems.”

Pangu Weather-AI uses a 3D Earth-Specific Transformer (3DEST) architecture to process complex non-uniform 3D meteorological data. It has studied historical data from 1979 to 2021 and can do 1-hour, 3-hour, 6-hour and 24-hour predictions. It can produce a complete 24-hour global weather forecast in 1.4 seconds.

NVIDIA’s FourCastNet

FourCastNet is an open-source machine-learning weather modeller. The link I have provided takes you to an essay with embedded source code.

FourCastNet uses a neural network to produce its accurate short-to-medium term (up to 10 days) weather forecasts with a high degree of accuracy. It has trained using historical data from 1979 to 2015 provided by Europe’s ECMWF Integrated Forecasting, a state-of-the-art meteorological system. The model uses 20 different variables to make accurate and speedy forecasts.

The Future of Meteorology

Are the days of the weather person numbered?

Peter Dueben who heads up the aforementioned ECMWF, doesn’t believe conventional human weather predictors will go the way of the dinosaur.

All the current AIs he notes have been trained on historical data. But as we are learning from the past few years, and in the last few months, climate change is making weather patterns more unpredictable. So a combination of human expertise and AI may end up tracking, recording and forecasting the weather well into the future.