At MTUCI, researchers forecasted temperature changes in Antarctica using a neural network.

In recent decades, climate science has rapidly advanced, with the amount of data collected steadily increasing. Approaches to studying the climate have evolved from simple descriptions to complex forecasts, and methods for data processing and analysis have improved, including the use of big data technologies and artificial intelligence.

Antarctica, located at the South Pole with its extreme temperature conditions and isolation from civilization, presents a unique opportunity for climate study. Preliminary research on temperature trends in Antarctica, based on meteorological station data, has indicated that these trends align with overall climate patterns, albeit with less fluctuation; notably, a slight warming trend characteristic of our era has been recorded.

The challenge of analyzing temperature dynamics is crucial in the context of assessing climate changes on the planet. Climate change poses one of the most serious threats to humanity's future, as it can lead to significant ecological, economic, and social consequences. Both global warming and the potential onset of a new ice age create risks for the survival of human civilization.

Researchers from the Department of Ecology at MTUCI have analyzed climate changes at the pole and developed a predictive model using a fully connected neural network to forecast temperature changes in Antarctica. The structure of the neural network model includes the number of layers, neurons, activation parameters, loss functions, and optimizers.

“In developing the neural network, which consists of two fully connected layers, we utilized the ReLU activation function. This function provides non-linearity and aids the model in better identifying complex patterns in the data. To assess the testing results, we applied the Median Absolute Error (MAE) metric, which clearly demonstrated the accuracy of the forecast and showed a significant impact of extreme climate conditions on the forecasting accuracy across several stations. In particular, the results at 'cold' stations were slightly worse than at 'warm' ones,” explained the head of the Department of Ecology, Associate Professor, PhD in Biological Sciences Victoria Erofeeva.

At the same time, the scientists emphasize that the rationale for using a fully connected neural network lies in its adaptability to complex non-linear relationships in climate data. The high degree of accuracy of the model compared to real observations highlights its effectiveness in capturing and predicting temperature trends in this region.

“Testing was conducted from 2020 to 2024. The neural network was tested at several Antarctic stations and showed promising results. Our task was to compare the capabilities of the neural network in forecasting climate for regions of the continent with various temperature regimes. A detailed analysis of temperature fluctuations and the identification of key periods of change allowed us to create an objective representation of climate dynamics at the pole.

To train the model, we used available data from polar meteorological stations from 1958 to 2019, which included temperature changes, normalization and preprocessing of data, as well as a validation strategy that enabled us to achieve high forecasting accuracy. In the course of the study, considering the climate characteristics of the continent, we divided the stations into 'warm' (Rothera, Bellingshausen, Vernadsky) and 'cold' (Vostok, Amundsen-Scott). Preliminary results indicated that trends at 'warm' stations, where average temperatures exceed -30 °C, differ from trends at 'cold' stations with significantly lower temperatures,” clarified Senior Lecturer at the Department of Ecology, Zhanna Zhukova.

Notably, in addition to testing models on historical data, the scientists conducted temperature forecasting for the next five years, which will allow for an assessment of the models' ability to adapt to future climate changes.

The research demonstrates a significant contribution of fully connected neural networks to forecasting temperature changes, highlighting their effectiveness in capturing complex climate patterns. The main trends identified during the work indicate the potential for adaptation to changing climate conditions through advanced machine learning technologies.

By paying special attention to analyzing the differences between 'cold' and 'warm' stations, the scientists identified some peculiarities in forecasting under more extreme climate conditions, which opens new prospects for further improvements to the models and their adaptation to diverse climate conditions.

In the future, there are plans to optimize approaches to data analysis and enhance the accuracy of long-term climate predictions in the region, particularly considering the expansion of geographical coverage and the refinement of forecasting models.

The article is based on material published on the scientific electronic library ELIBRARY.RU titled “Using Neural Networks for Modeling Climate Changes.”