Dynamic Personalized Graph Neural Ordinary Differential Equation Network for Multivariate Time Series Prediction of Chemical Processes

The prediction of multivariate time series (MTS) has increasingly captured substantial interest in practical domains such as chemical process flow and data analysis, which representing a profoundly challenging task. Numerous traditional MTS forecasting models demand substantial computational resources and may fall short of fulfilling real-time requirements. In this paper, given the intricate interdependencies within and between variables, a novel Multivariate Time Graph Neural Network (MTGNN) integrated with an ordinary differential equation (ODE) (MTGNN-ODE) approach is proposed to address the challenges of MTS prediction in chemical engineering. First, MTGNN represents the dependencies among variables in MTS data using the adjacency matrix to learn of spatio-temporal features. Then, The ODE is utilized to achieve the continuous propagation of the graph structure to a deeper level, and solves the problem of over-smoothing. In comparison with the baselines on a petrochemical process dataset, proposed method demonstrates a 15.6% enhancement in RSE and a 5.1% improvement in CORR. The prediction of key points in the chemical process is of great significance to achieve closed-loop control and real-time optimization of product quality.