Artificial intelligence improves shipping forecasts in port areas

A technology created by the University of São Paulo’s Center for Artificial Intelligence increases the accuracy of shipping and other weather forecasts by 20%, and is being tested at the Port of Santos.

Agência FAPESP* – Predicting the behavior of the ocean is a fundamental part of port operation optimization. Knowing which ships may dock and undock in accordance with tide, visibility, wind and other conditions is indispensable to efficiency for any harbormaster.

In light of this importance, researchers at the Center for Artificial Intelligence (C4AI) have developed algorithms to predict weather and ocean conditions with 20% higher accuracy than conventional methods.

C4AI is an Engineering Research Center (ERC) established by FAPESP and IBM at the University of São Paulo’s Innovation Center (INOVA-USP).

“Our forecasts are produced in the following manner. We teach the AI what happened in past weather phenomena, so that the machine learns the consequences of certain occurrences, such as heavy rain, high temperatures etc. It can then predict the unfolding of each condition as decision support, making the forecasting process faster and more accurate,” said Marlon Sproesser Mathias, a researcher at the Institute for Advanced Studies (IEA-USP) and a member of C4AI.

The algorithms are fed with raw data from weather stations and sensors installed in various parts of the ocean, from shipping lanes to the high seas.

“Shipping forecasts and meteorological predictions generally, regarding winds, currents, waves, visibility, tides, and so on, are getting more and more accurate, but there’s still room for an increase in levels of agility and reliability. More accurate forecasting means safer ports, lower shipping tariffs, shorter waiting times and more efficient logistics,” said Eduardo Aoun Tannuri, a professor at the School of Engineering (POLI-USP) and also a member of C4AI. 

Furthermore, reliable weather forecasting is key to more sustainable port activities, such as dredging to maintain or increase navigable depths. Optimization of these activities mitigates their environmental impact. Efficient use of big data and better knowledge of what can be done with the system’s analytical skills enable port authorities to handle larger volumes with less harm to the environment. “Our algorithms enable terminals to render large-draft ships more manageable with less dredging, thus less environmental impact,” Tannuri said.

A merger of physics and AI

With so much data from sensors and other sources, producing forecasts and estimates by conventional methods using only physical models would be difficult and time-consuming, so the researchers decided to merge two knowledge areas – physics and AI – in a combination of machine learning with coastal engineering and oceanography. 

“If it rains heavily in a shipping lane, for example, we know this may cause an increase in the current. It’s a matter of physical relationships. But AI takes time to learn this scenario, which has to happen many times for the correlation to become clear. If an event hasn’t happened yet, or in the case of extreme events, the machine won’t have learned about it. That’s particularly significant in a world of increasingly intense climate change,” Mathias said. “As we add physical knowledge to the AI’s model, its ability to adapt to different scenarios improves because it has the information needed to know what’s physically possible. It goes beyond modeling with data and physical simulation. We try to merge the best of both approaches.”

Automatic environmental monitoring systems exist internationally, but merging analysis of physical phenomena with machine learning for shipping and weather forecasting in port areas is entirely novel and differs from the type of analysis previously performed with the model in other locations, such as reservoirs or lakes.

C4AI’s method is being tested at the Port of Santos in São Paulo state and the Port of Paranaguá in Paraná state. Forecasting there has improved considerably thanks to integration with historical data. It has become faster and more accurate, suggesting that trials on a larger scale can be performed there and elsewhere in the near future, as the model can be applied in any port.

* With information from Henrique Fontes, a member of the Center for Artificial Intelligence (C4AI).