Spatial information bottleneck graph structure learning based multivariate time series prediction for industrial processes

Abstract

Prediction-based graph structure learning enhances both prediction accuracy and interpretability by identifying the underlying causes of prediction fluctuations, making it particularly valuable for industrial process monitoring. However, industrial data often exhibit strong spatio-temporal heterogeneity due to the presence of diverse physical measurements and redundant sensor placements, posing significant challenges for effective graph structure learning. Furthermore, when increasing the look-back window length to improve prediction accuracy, the heterogeneity of time series introduces more noise, making it difficult for graph structure learning methods to establish effective edge connections. Meanwhile, homogeneous time series provide redundant spatial features, causing prediction-based graph structure learning methods to fail. This paper is the first to study how to control the learned graph structure density in a multivariate time series prediction model to achieve a reasonable balance between prediction accuracy and structural accuracy. This paper proposes a Spatial Information Bottleneck (SIB) method to simultaneously address the aforementioned two challenges. The SIB method introduces the spatial feature prioritization principle, whereby the prediction model preferentially utilizes neighborhood node features for forecasting in homogeneous time series pairs, thereby enabling graph structure learning to establish connections between homogeneous time series pairs. Second, SIB performs independent information compression on each time series feature, which suppresses prediction-irrelevant noise in heterogeneous time series to varying degrees, thereby mitigating the impact of noise on prediction accuracy under long-sequence inputs. Experiments on industrial process data with accessible ground truth graph structures show that the model based on this method not only enhances prediction accuracy but also generates graph structures that align with physical processes for result interpretation.