Mobile weather forecast, why is it only 15 days?

Tech Fox

2024-08-02 11:12Published in Guangdong science and technology creators

The opening ceremony of the Paris Olympics, friends have watched it.

What impressed Lao Fox the most was not the bold picture of the CCTV commentary silent, nor the spicy-eyed Smurfs, but the rain that affected the effect of the live performance.

Many people may not know that in the Beijing Olympics 16 years ago, the original weather forecast was rain, and the wonderful performance of the Beijing Olympics was ensured through artificial rain elimination.

There is no doubt about the impact of weather forecasting on human production and life.

To put it more broadly, it affects the production and organization of agriculture and industry, and the farmers arrange agricultural affairs and enterprises arrange production, including such important activities as the opening ceremony of the Olympic Games, and weather forecasting has very important guiding significance.

To put it mildly, the weather forecast determines whether the old fox wears foot fitness or cave shoes to work.

Because the weather is so important, the weather forecast on TV has become a must-watch TV column for many people every day, and the BGM of "Fishing Boat Singing Night" is a common memory for many people.

I don't know if you have noticed, CCTV weather forecast, the predicted weather is only two days after tomorrow.

On mobile phones, weather forecasts are longer, providing 15 days of information both on plug-ins and third-party apps, with the exception of iOS, which is only 10 days.

If you want to improve the prediction time further, it is more difficult to be accurate due to the limitations of chaos theory, although Android phones have 90-day predictions, but they are also in units of ten days, and there are no more detailed weather changes.

Wouldn't meteorologists be able to accurately predict the weather for longer?

Last year, Huawei's Pangu meteorological model was published in the journal "Nature" in terms of its paper on the higher accuracy of weather prediction, and some time ago, Google's weather forecasting model Neural GCMs was also published in the journal "Nature".

The ancients had the need to accurately predict weather changes, and they did not have precision instruments to measure, but engaged in metaphysics.

On the unearthed Shang Dynasty oracle bone inscriptions, it is recorded that the ancients predicted the weather through divination, which is the earliest written record of weather forecasting.

As you can imagine, this is certainly not reliable. Later, the ancients observed various phenomena, summed up experience, and found out the law of weather changes.

The ancestors also left a lot of proverbs to help people predict the weather, such as hooked clouds in the sky and rain on the ground; Or the east wind is strong after the rain, and it will rain again the next day; Magpie branches bark, go out sunny day report, etc.

To a certain extent, it can also be called statistical prediction method, but this method lacks theoretical basis and is difficult to meet the needs of scientific development.

It's like building a house, if you don't have scientific data, you can build two or three floors based on your feelings, and you may be fine with two or three floors, but if you build a high-rise building, it may be a dilapidated house.

The key figure in the development of modern meteorology was the Norway physicist Wilhelm · Piekneis, who proposed in 1905 to describe the state of motion of the atmosphere with mathematical ideas, gave the equation system of atmospheric motion, and opened the door to modern meteorological research.

In 1910, William · Pieknis began to draw streamlines on weather maps, and through the collected ground and upper-air observations, he proposed a cyclone model for the development of mid-latitude cyclones and a polar front theory, which is the figure below. The result is a complete system that includes weather theory, weather map analysis and forecasting methods.

An important method of weather forecasting here is the weather map forecasting method, which uses various charts such as weather charts to predict the next changes in the weather based on the evolution history of the weather system, combined with physics, meteorology, and personal experience.

However, instead of using weather map forecasting, numerical weather prediction is used in today's weather forecasting.

As mentioned earlier, William · Pieknis gave the original equation system of atmospheric motion, which also became the basis for numerical weather prediction.

The system of atmospheric equations of motion includes equations of motion, continuous equations, thermodynamic equations, water vapor equations, and equations of state.

The so-called numerical weather prediction is to solve the system of equations according to the known initial conditions such as atmospheric temperature, humidity, and pressure, and find the future atmospheric temperature, humidity, pressure, and other data.

Don't panic, Lao Hu doesn't want to teach everyone to solve this equation, as long as you know that there is this equation.

How hard is it to solve this set of equations? In 1910, the United Kingdom scientist Richardson tried to use this equation to calculate the change in air pressure at a place in 6 hours, because the data was not processed correctly, resulting in wrong results.

If you want to predict the global weather changes, according to Richardson's vision, the earth is divided into many small regions in a huge spherical structure that mimics the earth, and each person is responsible for calculating the weather changes in one area, a total of 64,000 people.

It was later pointed out that Richardson had made another mistake and corrected the number of people who really needed to count - 204,800.

It wasn't until 1950 that the meteorologist Chani used ENIAC, the world's first computer, to simulate the first numerical prediction that matched reality, predicting the weather forecast 24 hours later, and the calculation process took exactly 24 hours.

It may sound like a wasted effort, but it means that numerical weather prediction can be commercialized, which is significant. It's like when you just graduate and just get a job, and your monthly salary is just enough to pay your rent and food, but for yourself, this is the beginning of real independence.

Taking the mainland as an example, a year ago, the mainland had a total of 7 atmospheric background stations, 27 climate observatoryes, nearly 70,000 ground automatic meteorological observation stations, 120 upper-air meteorological observation stations, 242 new-generation weather radars, and 7 Fengyun meteorological satellites in orbit.

*Fengyun Satellite Global Imagery

Therefore, today's weather forecasts are not only about temperature, pressure, and humidity, but also about visibility, lightning, typhoons, and even local extreme weather such as hail and tornadoes.

Massive amounts of data can only be relied on by state-of-the-art supercomputers. Like the Tianhe-1 before the mainland, the current Shenwei computer has played an important role in meteorological research.

At present, the leading institution for numerical weather prediction is the European Centre for Medium-Range Weather Forecasts, which is the ECMWF frequently mentioned in the paper on Huawei's Pangu Meteorological Model.

*ECMWF Weather Forecast Map

Let's start by explaining the "medium-term" here: the weather forecast within 0-12 hours is called a short-term weather forecast, the weather forecast within 3 days is called a short-term weather forecast, and the 4-10 days is a medium-term weather forecast.

The graph below shows the evolution of the forecast level of ECMWF at half a standard atmospheric altitude (about 5500m above sea level), with the thick line being the Northern Hemisphere and the thin line being the Southern Hemisphere. Forecasts for 3, 5 and 7 days have all improved significantly, while weather forecasts for more than 10 days can only be said to have some reference value at the moment.

Therefore, the weather forecast app on the mobile phone only provides a maximum of 15 days of weather data, and the effectiveness of numerical weather prediction is limited to the medium-term time scale.

The core idea of numerical weather prediction can be said to be "mechanistic", which holds that as long as the relative positions of all the components of nature are known, the motion of all matter can be summarized through calculations.

However, later meteorologist Lorenz found that there was a chaotic phenomenon in atmospheric movements, which limited the accuracy of weather forecasts.

The atmospheric system is a very complex linear system, which is extremely sensitive to all kinds of errors, and the small errors in the observed data, model initialization, and calculation accuracy are constantly amplified in the calculation process.

Lorentz's better known explanation for chaos is the "butterfly effect" — a South American butterfly flapping its wings, which could trigger a tornado in Texas, United States.

And can AI weather forecasting, which is now in full swing, eliminate the distractions of chaotic phenomena? The AI didn't eliminate it, but found a way to get around the pit.

Unlike numerical weather prediction, which is based on physical models, AI weather forecasting is a statistical prediction method that summarizes patterns by classifying and summarizing a large number of comprehensive and sufficient meteorological data samples in the past. Logically speaking, it is actually very similar to the experience of the ancients in summing up the weather.

It's just that I want to summarize a huge amount of data, so I leave this work to AI, and the more data, the more accurate the prediction results.

For example, the Pangu model was trained on meteorological data measured by the European Centre for Medium-Range Weather Forecasts over a period of 39 years from 1979 to 2017, validated with 2019 meteorological data, and tested with meteorological data from 2018.

The test results showed that over a seven-day period, the root mean square error of each weather variable predicted by the Pangu Meteorological Model was 10% lower than that of the Integrated Forecasting System of the Medium-Term European Weather Forecast Center, and 30% lower than that of Qualcomm's AI model FourCastNet.

In terms of typhoon trajectory prediction, the paper compares Typhoon No. 25 and No. 26 in 2018, and the results show that the Pangu meteorological model has obvious advantages.

The European Centre for Medium-Range Weather Forecasts debated itself in an article on its website, taking another typhoon as an example, which means that our two forecasts are close to each other, but your central wind speed forecast is not as accurate as mine.

This article also acknowledges AI weather prediction, but here you will find that FourCastNet is back in the picture, and then, its performance is simply ignored.

FourCastNet, a predecessor of AI weather prediction, was slapped to death on the beach by the back waves.

Unlike Huawei's Pangu Meteorological Model, Google's latest Nueral GCMs are a hybrid model with two core parts: one is to simulate fluid dynamics and thermodynamics during atmospheric motion by dynamic equations like numerical weather prediction; The other part is machine learning, where neural networks solve processes that cannot be simulated mathematically.

However, in terms of the weather prediction effect of longer time ranges, such as 10-15 days, the Pangu model is not mentioned, and another domestic Fuxi meteorological model mentions the prediction performance of 15 days in the paper, which is close to the European medium-term weather prediction center.

Another advantage of AI weather forecasting, especially for large models, is energy savings.

After a large number of trainings, the Pangu meteorological model can complete the prediction of important meteorological elements in the world for 7 days in 10 seconds, and the calculation speed is more than 10,000 times faster than numerical weather prediction.

The supercomputer used for numerical weather prediction consumes more than 20,000 kilowatt-hours of electricity per hour, which is equivalent to the electricity consumption of a small town.

AI has shown great potential for weather forecasting, but it is not about obsoleteing traditional numerical weather forecasting, and it is more likely that the two models will coexist in the future, as in Google's Nueral GCMs, the two models will work together and complement each other.

If in the future, the accuracy and prediction level of weather forecasting can be further improved, and people have enough time to prepare for disasters, then the loss of life and property will be greatly reduced, which is also the goal that countless meteorological scientists strive for.

The old fox is looking forward to this day.

Resources:

Henan Daily: In the "weather forecast" more than 3,000 years ago, how did businessmen "divination"?

Zhejiang Weather Network: Talking about weather proverbs

Science and Technology China: The weather forecast is actually "calculated" in this way

Xu Xiaofeng: From Physical Models to Intelligent Analysis: A New Exploration to Reduce the Uncertainty of Weather Forecasting

Du Jun, Qian Weihong: Three leaps forward in weather forecasting

《nature》：Accurate medium-range global weather forecasting with 3D neural networks

《nature》：Neural general circulation models for weather and climate

ECMWF：The rise of machine learning in weather forecasting

China National Defense News: How much do you know about the performance of supercomputers?

Edit: Hungry Sheep

View original image 1.1M

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

•