Stochastic gradient descent algorithm for the predictive modelling of grate combustion and boiler dynamics.

Pressure and water level control are quite challenging in grate-fired boilers due to higher combustion lag. Fluctuation in the pressure and the water level is more evident in the coal-fired grate boilers due to the presence of lower volatile and higher char. This results in suboptimal operation and poor performance of the boiler. This paper presents a novel predictive and dynamic simulation model for the drum dynamics analysis of a grate-fired boiler by combining a data-driven model and a thermodynamic model. A data-driven methodology is employed for the estimation of combustion, heat transfers, and circulation performance of the boiler. A novel thermodynamic model is proposed for the boiler dynamics of a hybrid boiler. The proposed data-driven model has been integrated with the thermodynamic model to reduce the randomness and improve consistency. Pressure and water level errors are estimated by comparing the predicted value and experimental result and the multi-objective optimisation technique is employed for the minimisation of errors. The Stochastic Gradient Descent algorithm is proposed for its ability to quick learning and adaptation to variations in fuel and combustion characteristics. The model demonstrates good accuracy in predicting the combustion and boiler dynamics of a grate-fired boiler. The proposed model has good potential to be used for the reciprocating grate solid-fuel boiler control in fluctuating load conditions.