Series clustering and dynamic periodic patching-based transformer for multivariate time series forecasting

Multivariate time series forecasting (MTSF) is widely employed in research-intensive domains, such as weather forecasting. Recently, Transformer-based models have outstanding ability to achieve SOTA performance, benefiting from its self-attention mechanism. However, existing models fall short in capturing multivariate inter-dependencies and local semantic representations. To tackle the above limitations, we propose a series clustering and dynamic periodic patching-based Transformer model named CMDPPformer, with two distinctive characteristics: (1) A channel-mixing module based on series clustering is proposed which can strengthen the association between variables with high sequence similarity, and weaken the effect of uncorrelated variables. Concretely, we use whole-time series clustering to group multivariate time series into clusters. After that, variables in the same cluster share the same Transformer backbone while variables in different clusters do not affect each other. (2) A dynamic periodic patching module is introduced which can better capture semantic information and improve Transformer’s local semantic representation. Concretely, multivariate time series after clustering are dynamically segmented into periodic patches as Transformer’s input token. Experimental results show that CMDPPformer can achieve an overall 13.76% and 10.16% relative improvements than SOTA Transformer-based models on seven benchmarks, covering four real-world applications: energy, weather, illness and economic.