Modelling and optimisation of post-combustion carbon capture process integrated with coal-fired power plant using computational intelligence techniques

Abstract

Coal-fired power plants are the major source of CO2 emission which contributes significantly to global climate change. An effective way to reduce CO2 emission in coal-fired power plants is post-combustion carbon dioxide (CO2) capture (PCC) with chemical absorption. The aim of this project is to carry out some research in model development, process analysis, controller design and process optimization for reliable, optimal design and control of coal-fired supercritical power plant integrated with post-combustion carbon capture plant. In this thesis, three different advanced neural network models are developed: bootstrap aggregated neural networks (BANNs) model, bootstrap aggregated extreme learning machine (BAELM) model and deep belief networks (DBN) model. The bootstrap aggregated model can offer more accurate predictions than a single neural network, as well as provide model prediction confidence bounds. However, both BANNs and BAELM have a shallow architecture, which is limited to represent complex, highly-varying relationship and easy to converge to local optima. To resolve the problem, the DBN model is proposed. The unsupervised training procedure is helpful to get the optimal solution of supervised training. The purpose of developing neural network models is to find a best model which can be used in the optimization of the CO2 capture process precisely. This thesis also presents a comparison of centralized and decentralized control structures for post-combustion CO2 capture plant with chemical absorption. As for centralized configuration, a dynamic multivariate model predictive control (MPC) technique is used to control the post-combustion CO2 capture plant attached to a coal-fired power plant. When consider the decentralized control structures based on multi-loop proportional-integral-derivative (PID) controllers, two different control schemes are designed using relative disturbance gain (RDG) analysis and dynamic relative gain array (DRGA) analysis, respectively. By comparing the two control structures, the MPC structure performs better in terms of closed-loop settling time, integral squared error, and disturbance injection.