Model combines physical parameters and machine learning to predict storm surges

Predicting extreme weather events is essential for preparing and protecting vulnerable areas, especially in times of climate change. The city of Santos, on the coast of São Paulo state in Brazil, is the largest port in Latin America and is the focus of an important case study, especially due to storm surges that threaten infrastructure and the local ecosystem.

Highlighting a key area of ​​focus for Santos, an article reports on the results of a study using advanced machine learning tools to optimize existing extreme weather forecasting systems. Proceedings of the AAAI Conference on Artificial Intelligence.

The study involved many researchers and was coordinated by Anna Helena Reali Costa, professor at the University of São Paulo's School of Engineering (POLI-USP). The first author is Marcel Barros, a researcher at POLI-USP's Department of Computer Engineering and Digital Systems.

The models used to predict sea level, high tides, wave heights, etc. are based on differential equations that include temporal and spatial information such as astronomical tides (determined by the relative positions of the Sun, Moon and Earth), wind conditions, current speeds and salinity.

Although these models have been successful in some areas, they are complex and rely on many simplifications and assumptions. Moreover, new measurements or other data sources cannot always be integrated into the models to improve the reliability of their predictions.

Modelers are increasingly using machine learning techniques that can identify patterns in data and extrapolate to new situations, but training algorithms to perform complex tasks such as weather forecasting or storm surge prediction requires a huge number of examples.

“In our research, we combine the two worlds and develop machine learning-based models that start with a physical model and improve it by adding measurement data. This field of research is known as Physics-Informed Machine Learning, or PIML,” Barros explained.

Harmonizing these two sources of information is essential for more accurate and precise predictions. However, using sensor data poses significant technical challenges, especially due to its irregularity, missing data, time lag, and variable sampling frequency. A failed sensor can take days to come back online, but storm surge prediction mechanisms must be able to operate continuously even without missing data.

“To deal with situations where we are dealing with highly irregular data, we developed an innovative way to represent the passage of time in a neural network. This representation allows the model to know the location and size of missing data windows and therefore take them into account when predicting tides and wave heights,” Barros said.

This innovation allows for better modeling of complex natural phenomena and can also be used to model other phenomena with irregular time series, such as health data, sensor networks in manufacturing, and financial metrics.

“Moreover, our model combines different types of neural networks to integrate multimodal data such as satellite imagery, tables and forecasts from numerical models. In future, it will be able to integrate other types of data such as text and audio. This approach is an important step towards more robust and adaptable forecasting systems that can deal with the complexity and variability of data related to extreme weather events,” said Leali Costa.

The model has three key strengths, she added: it combines physical and numerical models, uses neural networks to represent time in a new way, and has a multi-modal architecture to handle data in different formats.

“This study presents a methodology to improve the accuracy of forecasting extreme events, such as the Santos storm surge. At the same time, it highlights the challenges and potential solutions in integrating physical models and sensor data in complex situations,” she said.