Graph Neural Network Based Multi-Channel Fusion Multi-variate Time Series Forecasting

Time series forecasting is an important and challenging research content,which has related applications in various fields.With the continuous development of deep learning and the rise of Graph Neural Networks(GNNs),multivariate time series prediction models based on spatial-temporal convolution have gradually replaced traditional statistical methods.Spatial-temporal convolution starts from two aspects of the multi-node relationship and timing features,extracts the spatial relationship and timing features.Most of the current spatial-temporal convolution models are static,which means that the entire relational network remains unchanged at all times.This static correlation network does not consider that the relationship between nodes in the time series will change dynamically and lacks dynamic adaptation to the system.In addition,the switching of the system between different states will also change the interaction model.The current algorithm lacks the learning of the interaction relationship between the nodes under various working conditions.In response to the above issues,this article has carried out the following two aspects of work.Focusing on the time-varying and global-stable features in the data,TVGNN and GSGNN models are proposed.To obtain instantaneous features,TVGNN employs convolution and matrix cross-products to expand the input sequence to the time-varying correlation network,and uses prediction loss for adaptive learning;GSGNN uses pooling operations to smooth feature information and adds stable loss to guide the algorithm.The algorithms are verified on the traffic and engine data sets.On the traffic data set,TVGNN and GSGNN can improve the prediction effect by 5% compared with recent models.To fuse the characteristics of different composition algorithms and control the proportion of each channel information through hyper parameters,a multi-channel fusion graph neural network MCFGNN is proposed.Compared with other models on the traffic and engine data sets,the sequence prediction effect has been improved by 10%.The experiment part further conducts heat map analysis,hyper parameter analysis,and ablation experiments to verify the effectiveness of each module.This disseration studies the time-varying and global-stable features in time series,designs composition algorithms and models,which combines meta-feature graphs,timevarying graphs and global-stable graphs,in the main part.Spatial-temporal convolution is used for time series forecasting.Experimental results show that our model effectively improves the accuracy of time series forecasting.