Generalized predictive control using improved recurrent fuzzy neural network for a boiler-turbine unit

Abstract

The ultra supercritical (USC) for the boiler-turbine unit has become an advanced power generation technology due to its high combustion efficiency and low-carbon emission. Considering the complex nonlinearity and uncertainty in USC boiler-turbine units, a novel recurrent fuzzy neural network (RFNN) is produced to model dynamic responses of nonlinear systems and improve the control performance of the outputs in the boiler-turbine unit. However, the number of fuzzy rules in previous networks is mostly predefined with experts’ experience, which contributes to the decline in the generalization and efficiency of system modeling. This challenge is tackled in this paper by a subtractive clustering (SC) algorithm, which can determine the optimal number of fuzzy sets in the proposed RFNN. Additionally, aiming at minimizing the tracking errors of the outputs in the boiler-turbine unit, a global generalized predictive control (GPC) strategy is further designed to control the fuel flow, the feedwater flow, and the steam governor valve in the boiler-turbine unit. With the real-time data generated in a 1000MW USC boiler-turbine unit, the proposed SC-RFNN-based GPC can achieve a testing root mean-square-error (RMSE) of 0.0411 as well as the lower integral absolute error (IAE) values of system outputs.