Prediction of ash-induced agglomeration in biomass-fired fluidized beds by an advanced regression-based approach

• Modeling of maximum agglomeration-free operating temperatures for 83 biomass fuels.
• Application of advanced regression methods based on a large set of experimental results (350 individual data-points).
• Translation of prediction error into generally applicable agglomeration probabilities.
• Validation of predictions through long-term tests.

Energy crops and biogeneous residues offer the highest potential for future growth in biomass utilization. Traditional forest grown wood types, along with their consistent combustion characteristics, will thus be replaced by fuels with highly heterogeneous composition. Reliable prediction of their combustion characteristics and in particular of their ash behavior is essential for plant designers and operators trying to harvest this potential for energy conversion. In fluidized bed combustion, the fuel–ash induced agglomeration of the bed materials is one such behavior that needs to be described. In this paper a black-box model for agglomeration prediction was created through multivariate regression modeling using R-statistics v3.0.2. It based on the input variables bed ash concentration, particle size, fluidization velocity and fuel ash composition and predicts the maximum operable agglomeration-free temperature. Three linear and nine non-linear modeling algorithms have been applied to the data, optimized and validated in independent subsets. This validation was performed on results of controlled agglomeration tests, partly performed on our own test reactors and partly derived from literature. The final data set comprises 350 test results, covering 83 different fuels tested in seven different reactors. The validation revealed good predictive performance of the regression models, in particular of non-linear ensamble algorithms such as random forests, or cubist. These exhibit average deviations of around 60 K between model predictions and experimental results, which is very promising given the complexity of the system. After transformation of these prediction errors into agglomeration probabilities, a set of operational parameters unlikely to cause agglomeration can reliably be identified. A final evaluation of selected cases in controlled long-term tests could confirm the validity of these predictions.