.png){kind=icon}
.png){kind=icon}
Return
Discovering explicit and implicit causality for bioprocess factor forecasting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ins.2025.122846.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate forecasting of key variables in industrial bioprocesses is critical for optimization and control but is hindered by the data's high dimensionality, nonlinearity, and dynamics. Existing data-driven models often fall short in capturing the complex causal relationships inherent in multivariate time series (MTS). To address these limitations, this paper introduces the Explicit and Implicit Causality Aware Graph Neural Network (EICA-GNN), a novel framework designed for bioprocess factor forecasting. The core of the EICA-GNN model is a causality-aware autoencoder that utilizes a temporal Transformer with a pair-wise causal convolution mechanism to learn robust causal representations from MTS data. A key innovation is the causality discoverer module, which constructs a comprehensive relational understanding by generating both an explicit and implicit causal graph from learnable attention masks and temporal convolution kernels. These generated causal graphs, combined with a dynamic-static graph, are integrated into a graph-based predictor for the final forecasting task. The model's effectiveness was validated on real-world datasets from an erythromycin production process, consistently outperforming nine baseline models across prediction horizons.
Keywords:
Multivariate time series
Causality-aware
Explicit and implicit causal graphs
Pair-wise causal convolution