PhyG-Net: A Physics-Informed Adaptive Regime-Aware Gating Network for long-term temperature forecasting in biomass boilers

Accurate long-term forecasting of outlet water temperature is a prerequisite for the intelligent operation and fault diagnosis of biomass boilers. However, industrial thermal systems exhibit a “dual characteristic” dilemma: significant thermal inertia during steady states and high-frequency fluctuations during combustion disturbances. Existing deep learning models predominantly employ static feature fusion strategies, which often fail to adaptively reallocate attention, leading to significant phase lags and “physically blind” predictions that violate thermodynamic principles. To bridge these gaps, this paper proposes PhyG-Net, a Physics-Informed Adaptive Regime-Aware Gating Network. We design a dual-stream encoder that decouples system dynamics into global inertia (modeled by an inverted Transformer) and local transients (modeled by physically grounded multi-scale convolutions). A novel Adaptive Regime-Aware Gating Mechanism acts as a differentiable switch, enabling seamless transitions between inertia-dominant and transient-dominant regimes based on real-time contexts. Furthermore, a Physics-Informed Curriculum Learning framework progressively enforces velocity and acceleration constraints to ensure thermodynamic consistency. Extensive experiments on a 2.5-ton biomass boiler demonstrate that PhyG-Net achieves state-of-the-art performance, reducing the Mean Squared Error (MSE) by approximately 18.6% compared to the iTransformer backbone at a 192-step horizon. Sensitivity analysis identifies L = 256 as the optimal lookback window based on an empirical “Inverted-U” pattern. Crucially, PhyG-Net effectively mitigates the characteristic 15–20 min phase lag observed in pure Transformers, providing more temporally aligned predictive information for control-oriented decision support and interpretable indicators for unsupervised fault detection.