A Neural Network Time Series Model to Forecast Atmospheric Carbon Dioxide Concentrations in Central Siberia

Abstract

the rise in atmospheric carbon dioxide concentrations driven by natural and anthropogenic factors is one of the main causes of climate change, making accurate long-term forecasting of atmospheric carbon dioxide concentrations an important scientific challenge. In this study, we compared the performance of several machine learning methods for forecasting atmospheric carbon dioxide concentrations using time series data collected at the ZOTTO Station in Central Siberia in 2009–2022. We evaluated ensemble methods, including random forest and gradient boosting, together with a recurrent neural network based on the Long Short-Term Memory architecture. The predictor variables included fire intensity, a lag variable representing the number of days since the beginning of the observation period, and meteorological parameters, including air temperature, daily maximum and minimum temperatures, dew point temperature, relative humidity, wind speed, atmospheric pressure, and precipitation. The results showed that the Long Short-Term Memory model achieved higher forecasting accuracy on the validation dataset than the ensemble methods, with a 2.87 mean absolute error, .91 RMS error, and 0.83 coefficient of determination. We found that air temperature, relative humidity, dew point temperature, and wind speed were the most important predictors of atmospheric carbon dioxide concentrations.