APSO-based Neural Network Hybrid Modeling For The CFB-FGD Processes

The technology of circulated fluidized bed for flue gas desulfurization (CFB-FGD) is based on the theory of circulated fluidizing bed. It augments the material concentration and lengthens the reacting time by recycling the reaction product and desulfurizer, thus increases the desulfurizer utilization ratio and the desulfurization efficiency. The CFB-FGD technology is suitable for the current condition of China because of its low investment cost, high desulfurization efficiency, reliable operating condition, convenient maintenance and other advantages.Over 80% of the power plants are coal-fired for power generation, which induce a mass of sulfur dioxide emission and bring serious environment pollution problems. So how to reduce the pollutant emission is one of important issues that should be solved urgently. For excogitating of control technology about CFB-FGD which has the trait of technique viable, high standard, low project alteration and working period, reliable investment and self-determination knowledge property right, the thesis sets up the CFB-FGD principle model. Because the CFB-FGD is a nonlinear, multi-variable and complicate system, its mechanism has not been completely mastered even now. The mathematics model doesn't satisfy the actual need, so the zoom out design and the operative process guidance is mainly based on the experience. This thesis emphasizes on the hybrid modeling based on BP neural network optimized with particle swarm optimization (PSO) algorithm to solve above problems.Firstly, the run course regulation of CFB-FGD is analyzed in great depth and the principle model is established. Efficiency predictions using this model have been compared with the test data. It can be seen that the calculated results from the model agreed well with the experimental results.Secondly, the reasons of premature convergence of PSO algorithm are analyzed systematically. An adaptive PSO (APSO) is proposed, in which the inertia weight of the particle is adjusted adaptively based on the premature convergence degree of the swarm and the fitness of the particle. The diversity of inertia weight makes a compromise between the global convergence and convergence speed. Simulation results show that APSO not only effectively alleviates the problem of premature convergence, but also has fast convergence speed.Thirdly, an approach that neural network optimized with APSO algorithm is proposed in the paper. Unlike conventional training method with gradient descent method only, this paper introduces a hybrid training algorithm by combining the APSO and BP algorithm. The APSO is used to optimize the architecture and the initial parameters of the BP neural network, including the weights and biases. It can effectively better the cases that network is easily trapped to a local optimum and has a slow velocity of convergence. The experiment results show the method one has greater improvement in both accuracy and velocity of convergence for BP neural network.Lastly, according to the principle model we have had, a hybrid model using neural network optimized with APSO to compensate the error from the principle model is proposed. Then we discuss about the structure, training and simulation of this principle-neural network hybrid model.The emulational result indicates that the hybrid model can simulate and predict the desulfurization efficiency perfectly.