Recurrent neural networks integrate multiple graph operators for spatial time series prediction

For multivariate time series forecasting problems, entirely using the dependencies between series is a crucial way to achieve accurate forecasting. Real-life multivariate time series often have complex time dependence, spatial dependence and high nonlinearity simultaneously, so Euclidean space is no longer sufficient to describe them. graph neural network presents a vital idea to solve this problem by modelling multivariate time series as graphs. Using the nature of graphs makes it possible to capture the dependencies between multivariate time series. However, no graph structure can perfectly characterize the relationships among multivariate time series; the facts underlying multivariate time series are much more complex. Therefore, we propose an integrated model (iGoRNN), which improves the model’s understanding of the deep relationships of multivariate time series by fusing the information captured by multiple graph operators through an integrator with a specific structure. In addition, we conducted experiments on the Metr-LA and PeMS-BAY datasets. The experimental results show that the proposed model outperforms the baseline model in three evaluation metrics, MAE, MAPE and RMSE, and can forecast complex multivariate time series.