COMPARATIVE ANALYSIS OF NEURAL NETWORK MODELS FOR FORECASTING THE DYNAMICS OF GREENHOUSE GAS CONCENTRATIONS IN THE SURFACE AIR LAYER ON BELY ISLAND

Bobakov, V. C., Buevich, A. G., Butorova, A. S., Sergeev, A. P.

Institute of Industrial Ecology of Ural Branch of the Russian Academy of Sciences (Yekaterinburg, Russian Federation)

The article was received on July 23, 2025

Bibliographic description: Bobakov V. C., Buevich A. G., Butorova A. S., Sergeev A. P. Comparative analysis of neural network models for forecasting the dynamics of greenhouse gas concentrations in the surface air layer on Bely Island. Arctic: Ecology and Economy, 2026, vol. 16, no. 1.

Abstract:  In recent years, the application of machine learning technologies has become increasingly relevant for moni-toring greenhouse gas emissions in the Arctic, where intense permafrost degradation and increasing methane emissions into the atmosphere are observed. In the study, we have developed and compared recurrent and graph neural network-based models to predict the dynamics of concentrations of major greenhouse gases in the sur-face air on Bely Island, the Yamalo-Nenets Autonomous Area, Russia. The original dataset consisted of surface concentration measurements of four gases for the summer months of 2016—2017: carbon dioxide (CO2), meth-ane (CH4), carbon monoxide (CO), and water vapor (H2O). For graph forecasting, we used the Multivariate Time Series Graph Neural Network (MTGNN) model; for recurrent forecasting, we used the Long Short-Term Memory (LSTM) and Long Short-Term Memory Network (LSTNet) models. The forecast accuracy of the models was assessed using the following metrics: mean absolute error (MAE), root mean square error (RMSE), normalized root mean square error (NRMSE), and correlation coefficient (Corr). In general, the MTGNN model demonstrated lower error values compared to recurrent models (LSTM, LSTNet). Compared to LSTM and LSTNet, MTGNN had 25—55% lower errors. The more accurate graph model required significantly more time to train and had a sig-nificantly larger number of parameters for optimization.

Keywords: time series, greenhouse gases, Arctic, forecast, neural networks, graph neural networks, LSTM

Funding:  The study was supported by Ministry of Science and Higher Education of the Russian Federation, project FUMN-2024-0003.

DOWNLOAD (in Russian)