How to get better weather forecasts with artificial intelligence

Using machine learning, i4cast® wind and wave forecasts are up to 40% more accurate. An excellent result!

But what is machine learning and how does it contribute to sea and weather forecasts? That's what we're going to see now. Follow me!

How are sea and weather forecasts made?

Sea and weather forecasts are generated by sophisticated numerical models, a type of computer program that simulates environmental conditions.

These numerical models use data on air temperature, rainfall, winds, waves, among others, to solve complex mathematical equations that describe the behavior of oceanic and atmospheric phenomena, and thus predict what will happen in the coming days. These complex calculations require large computing infrastructure or supercomputers and can take hours to complete.

You can find out in more detail how weather forecasts and ocean forecasts are made by clicking here.

As good as the technology is, there is a major challenge associated with sea and weather forecasts: the chaotic nature of the atmosphere and ocean. Because of this, small deviations in the first hours of the forecast can result in large errors days ahead, increasing the uncertainties of the results.

That's why today's forecast is much more accurate than tomorrow's forecast, and the more days ahead, the more difficult. This is where machine learning can help.

What is machine learning?

Machine learning is a branch of Artificial Intelligence.

In this process, computers can learn from a large amount of complex data and develop the ability to recognize patterns or predict future trends, outcomes and behaviors, without being specifically programmed to do so.

That is, with machine learning it is possible to train the computer to learn a certain activity that, after this training phase, will be performed without human intervention.

A simple example of using machine learning is your email spam filter. Initially the computer, or predictive model, is trained to recognize spam emails in your inbox. After this training phase, algorithms are created – finite sequences of instructions to solve a problem – that start to operate autonomously, freeing your inbox from receiving that many unwanted emails.

Why use machine learning?

There are many advantages of using machine learning techniques, but the main one is efficiency.

These predictive models built by data-driven algorithms are much more adaptable and flexible than following pre-programmed static instructions. All this with a high processing power, which allows analyzing an absurd amount of data in short periods.

With machine learning, it is possible to analyze increasingly complex and numerous data, automatically and quickly. And of course, virtually eliminating human error from the process.

Right. But how do we apply this to sea and weather forecasts?

Why use machine learning in weather forecasts?

As I said before, there are several challenges associated with sea and weather forecasting, but the accuracy of forecasts can be significantly increased using machine learning techniques.

The use of artificial intelligence techniques, such as machine learning, ensures that corrections in predictions are made with data measured by sensors in real time, reducing uncertainties and making predictions more accurate, especially in the next 24 hours.

Even better, with the use of Machine Learning algorithms, predictions can be updated much more frequently, with a new prediction being generated every hour, for example.

That's because it only takes a lot of computational power to train the Machine Learning algorithm. Once trained, its application is simple and light, and can be repeated several times.

This is a great benefit, considering that numerical models take hours to generate a new prediction and, therefore, the application of machine learning techniques can generate corrections on existing predictions in a matter of seconds!

How to get better sea and weather forecasts using machine learning?

At i4sea, we use machine learning to develop a predictive model that uses data measured by local sensors to adjust the numerical model results and its predictions. And we do this process several times a day.

For this, several training rounds are carried out with a large volume of possible environmental scenarios until reaching a certain performance. Once trained, machine learn corrects the model prediction using real-time sensor data.

Depending on the variable we want to correct, we can apply different machine learning methods, ranging from simpler methodologies, such as linear regressions, to complex deep learning methodologies, involving neural networks.

The results obtained are encouraging!

In wave forecasts, we observed gains of 40% in accuracy for 6h forecasts and 25% gains in accuracy for 20h forecasts.

In wind forecasts, which are even more chaotic and therefore more difficult to predict, we observed an accuracy gain of 42% for 3-hour forecasts and 15% for forecasts between 3 and 6 hours.

This performance is continuously monitored and the training is redone whenever more data is available, because with a greater volume of data, the greater the diversity of patterns that the machine learning algorithm is exposed to during its training.

A real example of the benefit of having more accurate sea and weather forecasts is the improvement of port planning at Porto do Açu, where i4cast® decision-making support is twice as assertive as global forecasts in real moments of environmental restrictions.

What do twice as assertive decisions mean to you?