Graph neural network model for multivariate time series forecasting
Most of the existing graph neural network models for forecasting multivariate time series capture the time series characteristics in a static way based on predefined graphs,and may be lack of capturing the dynamic adaptation of the system and some potential dynamic relationships between time series samples.A graph neural network model for multivariate time series prediction (MTSGNN) was proposed.In a graph learning module,the static and dynamic evolution graphs of time series data were learned in a data-driven way to capture the complex relationships between time series samples.The information interaction between the static and dynamic graphs was realized by the graph interaction module,and the convolution operation was used to extract the dependency in the interaction information.A multi-layer perceptron was used to forecast the multivariate time series.Experimental results on six real multivariate time series datasets showed that the forecasting effect of the proposed model was significantly better than those of the current state-of-the-art methods,and it had the advantages of small parameter quantity and fast operation speed.
multivariate time seriesgraph neural networkstatic graphdynamic graphgraph interaction
万方数据
