Predicting Global Average Temperature Time Series Using an Entire Graph Node Training Approach

The data-driven approach has become significant in various scientific fields, such as climate modeling and weather forecasting, where a mechanistic description of physics and chemistry is either unavailable or insufficient for the desired purpose. However, the prominent nonstationarity poses a significant challenge to the accurate prediction of recent global average temperature (GAT) using conventional methods. To address this challenge, we draw inspiration from signal analysis's moving-window approach, wherein we split the GAT into shorter segments to alleviate nonstationarity. These segments are transformed into a time-symmetric graph (TSG) structure in the non-Euclidean domain. Consequently, we introduce the GlobalTempNet model, which incorporates a graph convolutional network (GCN) embedded with a residual neural network (NN) and a long short-term memory (LSTM) network. In addition, we propose the entire graph node training (EGNT) process, optimizing parameters by treating each sample as a graph node for feature aggregation and information updating. Validation using the HadCRUT5 dataset demonstrates that GlobalTempNet outperforms nine established models, showcasing higher prediction accuracy. Furthermore, long-term estimation and future prediction analyses reveal GlobalTempNet's capability to predict the climate change trend in the coming years. The model's applicability is confirmed across ten different temperature datasets. Consequently, the proposed GlobalTempNet, coupled with the EGNT process, emerges as a robust, reliable, and open-source method for global temperature prediction, offering promising potential as a tool for univariate time series analysis using graph NNs (GNNs).