Low-Rank Representation-Based Soft-Sensing Modeling for Secondary Air Velocity Field in Opposed Wall-Fired Boilers

Under the dual carbon targets, the secondary air velocity field of opposed wall-fired boilers faces challenges such as limited measurement points, high-dimensional coupling, and scarce on-site steady-state samples, which restrict refined air distribution regulation under deep peak operation. To address this issue, this study constructs a multi-condition computational fluid dynamics (CFD) numerical simulation dataset and proposes a low-rank representation soft-sensing method integrated with orifice outflow physical prior constraints to decompose the 40-dimensional velocity field into learnable spatial modal bases and low-dimensional coefficients, introducing a soft regularization loss to improve physical consistency. The results show that the proposed method achieves a mean absolute error (MAE) of 0.593 m/s and a root mean square error (RMSE) of 0.812 m/s on the test set, with R2 reaching 0.9717. Compared with the traditional point-by-point regression model, MAE and RMSE are reduced by 25.2% and 22.9%, respectively, and duct section pressure is identified as the core input feature. This method provides an effective solution for small-sample high-dimensional industrial flow-field soft-sensing, and offers a quantitative basis for low-cost optimization of on-site measuring point layout.