https://doi.org/10.1140/epjs/s11734-024-01368-z
Regular Article
Application of graph-structured data for forecasting the dynamics of time series of natural origin
1
Ural Federal University, Mira Str., 19, 620002, Ekaterinburg, Russia
2
Institute of Industrial Ecology UB RAS, S. Kovalevskoy Str., 20, 620990, Ekaterinburg, Russia
Received:
11
September
2024
Accepted:
10
October
2024
Published online:
23
October
2024
The study aimed to compare graph and non-graph methods in the task of forecasting the dynamics of time series of natural origin. Non-graph methods were represented by long short-term memory (LSTM), temporal convolutional network (TCN), and long short-term memory network (LSTNet), while graph methods involved multivariate time series graph neural network (MTGNN) and graph convolutional recurrent network (GCRN) which was fed with graph-structured data obtained from initial time series by two different approaches. The initial data were synchronized measurements of main greenhouse gas concentrations and meteorological parameters: a total of nine variables. The dataset consisted of three sub-datasets: time points for summer months of three different years, which were consequentially combined into one time series used for forecasting. To assess the performance of models, mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), normalized root mean squared error (NRMSE), correlation coefficient Corr, and determination coefficient R2 metrics were calculated and standard deviation, root mean square, and correlation coefficient were visualized in the Taylor diagram. In addition, the number of learning parameters and training time were assessed for each model. In general, GNN models outperformed non-graph models in each of the assessed metrics. Standard deviation, root mean square, and correlation coefficient visualization in the Taylor diagram also proved the superiority of GNNs. Both graph and non-graph models have managed to capture the trend in the data, however, GNN models were more accurate and handled peaks and sharp transitions in data better than the non-graph models while demanding a larger number of learning parameters.
Copyright comment Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
© The Author(s), under exclusive licence to EDP Sciences, Springer-Verlag GmbH Germany, part of Springer Nature 2024. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
