Physics-Informed Spatiotemporal Graph Forecasting Framework with Drift-Triggered Continual Learning for NOx Prediction in Municipal Solid Waste Incineration Power Generation
Municipal solid waste incineration (MSWI) represents an important waste-to-energy technology. However, its nitrogen oxide emissions create environmental risks, and accurate NOx prediction remains a challenging task. This challenge arises from variable waste properties, strong nonlinearity in the combustion process, and time-varying operating conditions. Most existing data-driven models are trained under a stationary data distribution. They may produce physically inconsistent predictions, and their performance tends to degrade when process operating conditions shift. This study proposes a physics-informed spatiotemporal graph forecasting framework equipped with drift-triggered continual learning for NOx prediction in MSWI plants. The framework contains three core components. The first is a plant-specific spatiotemporal graph network, which captures spatial coupling across process compartments and temporal dependencies within sensor measurements. The second is a differentiable physical module. It imposes process consistency constraints on the incineration system and steers the network to generate physically plausible predictions. Third, a drift-triggered continual learning mechanism adapts the model online when operating conditions change, while replay and protection strategies reduce catastrophic forgetting and preserve physical and historical consistency. Comparative experiments against state-of-the-art time-series forecasting models verify the effectiveness of the proposed method.
Authors
- Haixu Ding (ORCID: https://orcid.org/0000-0001-7822-7705)
- Junfei Qiao
- Tao Yu
- Yuntao Tang
- Guangwei Chen
Publication Details
- Journal
- Preprints.org
- Published
- 2026-11-17
- DOI
- https://doi.org/10.20944/preprints202611.1681.v1
- Primary Topic
- Air Quality Monitoring and Forecasting
- Type
- preprint