Skip to main navigation Skip to search Skip to main content

Physics-constrained Bayesian transformer for temperature soft sensing with uncertainty quantification in nuclear waste incinerators

  • Xi'an Jiaotong University
  • China Nuclear Power Engineering Co., Ltd
  • China National Nuclear Corporation

Research output: Contribution to journal › Article › peer-review

Abstract

Accurate temperature monitoring in nuclear waste incinerators is challenged by thermal inertia, measurement uncertainties, and the need for reliable confidence estimates in safety-critical operations. Conventional thermocouple-based methods suffer from large dynamic errors under rapidly changing conditions, and existing soft sensor approaches often lack physics-informed constraints and robust uncertainty handling. This paper proposes Physics-constrained Adaptive Thermal Intelligence (PCATI), a physics-constrained Bayesian sequence-modeling framework for uncertainty-aware temperature soft sensing. The framework features three key innovations: (1) a Dynamic Temperature Compensation (DTC) model based on detailed heat transfer analysis, employing Kalman filtering and recursive least squares to mitigate thermal inertia in real time; (2) a Bayesian Gated Convolutional Transformer (BGCT) architecture that combines adaptive feature extraction, temperature-aware attention mechanisms, and principled uncertainty quantification via variational inference; and (3) a physics-constrained training strategy that directly embeds energy conservation, steady-state equilibrium, and temperature evolution constraints into the optimization process. The PCATI framework is evaluated on five real combustion cycles collected from a single nuclear waste incinerator facility, using wall measurements to estimate internal temperatures. Across the evaluated comparisons, PCATI achieves RMSE values consistently below 6.2°C, an average prediction interval coverage probability of 95.06%, and error reductions of 69.1%–75.4% under dynamic conditions, placing it among the best-performing methods in the evaluated five-cycle benchmark. By combining high predictive accuracy, calibrated uncertainty quantification, and thermodynamic consistency, this work advances physics-informed soft sensing for safety-critical thermal monitoring and motivates further validation of physics-constrained deep learning across broader high-temperature industrial settings.

Original languageEnglish
Article number123274
JournalMeasurement: Journal of the International Measurement Confederation
Volume292
DOIs
StatePublished - 15 Dec 2026

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

  1. SDG 7 - Affordable and Clean Energy
    SDG 7 Affordable and Clean Energy

Keywords

  • Dynamic temperature compensation
  • Nuclear waste incineration
  • Soft sensor modeling
  • Thermal monitoring
  • Uncertainty quantification

Fingerprint

Dive into the research topics of 'Physics-constrained Bayesian transformer for temperature soft sensing with uncertainty quantification in nuclear waste incinerators'. Together they form a unique fingerprint.

Cite this