Application of machine learning algorithms in boiler plant root cause analysis: a case study on an industrial scale biomass unit co-firing sugar cane bagasse and furfural residue at excessive final steam temperatures

THE CURRENT WORK sets out to showcase the power of statistical learning algorithms to mine boiler operational data in an attempt to create a predictive model capable of capturing the plant specific behaviour. The machine learning predictive model can be used to perform investigations such as: boiler diagnostics, sensitivity analysis on operational parameters and root cause analysis to determine cause of upset/detrimental conditions. A data mining analysis was performed on an industrial scale biomass boiler co-firing sugar cane bagasse and furfural residue, which operated at excessive final steam temperatures (420?440 ?C) when compared with the design steam temperature (400 ?C). The goal of the analysis was to find the cause of the excessive final steam temperatures and propose remedial action. The analysis comprised of using artificial neural network, support vector regression (E-SVR) and Random Forest machine learning algorithms to mine the operational data acquired from the boiler?s distributed control system and generate a statistical predictive model. A sensitivity analysis was performed on the boiler input parameters (fuel moisture, fuel density, fuel feeder speeds, induced draught fan speed, forced draught fan damper position, etc.) using the machine learning model, to find the inputs that caused excessive temperature excursions. The model was able to accurately capture the boiler trends, and was used to identify that it was the fuel moisture, density and upward flow velocity in the furnace that caused the flame to be positioned much higher in the furnace than intended. The higher flame position caused an increase in thermal radiation heat transfer to the radiant superheater above the design values, which resulted in the higher final steam temperature.