Using Graph Neural Networks (GNN) and Gated Recurrent Unit (GRU) based on Regression Modelling for Climatic Forecasting

Abstract

Most climate time series forecasts suffer from inaccuracy, which can negatively and directly impact the effects of forecasting on human life and living organisms, especially when the data has both temporal and spatial dimensions. The use of traditional forecasting methods when dealing with complex non-linear patterns, as well as the problem of data heterogeneity resulting from variations in the temporal (seasonal) and spatial dimensions, as is the case with multi-site climate data, can lead to reduced forecast accuracy. This study used multivariable climate time series data collected from meteorological stations in Nineveh Governorate, Iraq, for daily observations from the beginning of 1994 to the end of 2024. Graph neural networks (GNN), a deep learning technique, were used to process the spatio-temporal data by establishing node-to-node connections based on neighboring information to improve forecast quality by combining the spatio-temporal dimensions. To address the long-term and short-term memory challenges faced by climate data, a gated recurrent unit (GRU) was employed as an alternative to long short-term memory (LSTM) Networks. GRUs offer fewer variables and parameters, simpler computational processes, similar performance, high predictive accuracy, and faster computation speed. The GNN and GRU methods were hybridized using Multiple Linear Regression (MLR), with normal temperature as the response variable. This hybridization provided a statistical basis for determining the input variable structure of both methods, thereby improving the forecasting accuracy of the proposed hybrid methods, GNN-MLR and GRU-MLR, which are structured according to sound scientific principles. The results of this study demonstrate that GNN-MLR achieved more accurate forecasting results compared to GRU-MLR, and both models outperformed the traditional MLR model. The results also indicate that the nature of the data, in terms of low variability and stable values, suggests that the improvement in predictive results when using GRU-MLR hybridization was slight. From this, it can be concluded that the geographical proximity of cities, which ensures rapid air mass movement, reinforces the importance of the spatial dimension in climate relationships, in addition to the temporal dimension, unless there are problems with variability and stability. This is because all cities are affected by a uniform climate pattern due to their small size and proximity, thus allowing for an absolute advantage in using GNN-MLR hybridization.