Enhancing parameter prediction in gas-fired boiler systems through node similarity-based feature aggregation

Accurate sensor network prediction is crucial for improving industrial boiler efficiency and safety. While existing predictive models show promise, they are constrained by several limitations: (i) insufficient integration of interpretable multi-level spatiotemporal information, (ii) over-reliance on static topologies and shallow features, and (iii) limited continuity and adaptability in complex environments. To address these challenges, we propose a novel framework to improve parameter prediction in gas-fired boiler systems through node similarity-based feature aggregation. First, we apply a node similarity-based multi-level aggregation strategy for interpretable multi-scale integration. Next, dynamic graph learning, utilizing a higher-order graph convolutional network, captures the evolving relationships between sensors and time steps. Additionally, continuous modeling is facilitated by a spatiotemporal ordinary differential equation solver, which overcomes the limitations of discretized time steps. Real-world evaluations show our approach improves accuracy and robustness, even with sensor failures. Furthermore, the continuous model supports predictions at any time step. This approach provides a foundation for data-driven parameter prediction and the modeling of interacting industrial components.