Variable Spatiotemporal Framework for Multivariate Time Series Prediction

Multivariate time-series forecasting plays a crucial role across diverse domains. The inherent variability of spatiotemporal dependencies among these variables introduces substantial challenges in forecasting. Traditional methods, with their homogeneous static frameworks, fail to capture the dynamic spatiotemporal nature of the data, thereby limiting the model's ability to accurately fit and represent the underlying patterns. In this paper, we propose a novel variable spatiotemporal framework for multivariate time series prediction. This framework constructs spatiotemporal-varying model parameters, achieving “dynamic-dynamic” alignment between the forecasting framework and real-world multivariate systems. On multiple multivariate time series datasets, our model demonstrates superior performance and maintains its advantage in long-term forecasting.