Monitoring fuel properties during boiler operation is important, whereas intermittent laboratory analyses provide limited temporal resolution. This study presents a physics-constrained artificial neural network (ANN) soft sensor for estimating fuel throughput, lower heating value, moisture, ash, elemental composition, and excess air ratio from 12 measurements in a 100 MWe biomass-fired circulating fluidized-bed boiler. To address limited fuel-analysis data, one million synthetic samples were generated from mass and energy balances using fuel-property and operating ranges obtained from plant data. Before ANN development, identifiability analysis showed that the available measurements were insufficient for unique estimation of the target fuel properties, as different fuel-property combinations could produce similar measurements. Accordingly, two constrained formulations were introduced by fixing either the H/C ratio or ash content at their respective average values to improve identifiability. Separate multilayer perceptron soft sensors were then trained for each formulation using the corresponding synthetic datasets, while dependent fuel-state variables were reconstructed using physical relations. The training objective combined prediction accuracy with physical consistency. For field application, plant measurements were preprocessed into 113 daily states. Synthetic-to-field discrepancies were reconciled by optimizing input correction factors within ±3% using energy and elemental balance residuals. Measured fuel throughput and laboratory analyses were excluded from the optimization and reserved for independent evaluation. The reconciliation reduced the test-set balance loss by 77% and 85% for the H/C-fixed and Ash-fixed formulations, respectively. The corresponding fuel-throughput mean absolute percentage errors were 3.36% and 4.34%, with the H/C-fixed formulation showing better agreement with the measured values. The Ash-fixed formulation showed greater overlap with laboratory-derived reference distributions for several fuel properties. Overall, the proposed framework provides physically consistent fuel-property estimates from limited plant measurements while accounting for their dependence on the imposed physical constraint.
 

biomass,fuel throughput,fuel properties,circulating fluidized bed,Artificial neural network