【作者】 孙雷；

【导师】 任志波；

【作者基本信息】 河北大学 ， 管理科学与工程， 2013， 硕士

【摘要】 目前我国电力行业依然以火力发电为主，但是其在生产过程中的排放物却严重污染了环境，尤其是二氧化硫对大气的污染，所以控制火电行业二氧化硫的排放量是我国目前治理环境污染的一项重要工作。而在生产过程中，影响二氧化硫排放的因素很多，如何对这些因素进行正确合理的分析以便从中找到有价值的信息是目前研究的重点和难点。通过对脱硫系统的分析和研究，可发现，采用智能算法来实现脱硫系统的仿真和优化比较方便，并且准确性高，还可以排除人为因素的影响。神经网络算法是模拟人脑的高复杂的神经元，并将神经元相互连接组建的网络系统。其自学习能力和大规模并行处理能力在工程仿真中的应用相当普遍。遗传算法的指导性搜索使得其在工程寻优方面比较迅速并且准确性高，不易出现局部最优的结果。神经网络主要采用梯度下降法来实现仿真过程中对权值和阈值的更改，梯度下降法有很多种，其中，动量梯度下降法是在梯度下降法的基础上增加一个动量来控制下降的速度和方向，共轭梯度下降法是增加一个共轭方向，使得在下降过程中始终沿着梯度的共轭方向下降迭代。本文基于以上两种算法提出了带动量的共轭梯度下降模型，该方法增加了下降速度，提高了仿真的精度。并利用遗传算法来对神经网络仿真函数进行寻优，但是由于神经网络的缺陷，使用遗传算法在训练过程中会出现寻优结果越界的问题，针对此问题本文提出了基于惩罚值的遗传算法优化模型，实现了对遗传算法的有效控制。更多还原

【Abstract】 At present, thermal power plants is the main part in our country’s power industry. But theenvironment is polluted by emissions in the production process, especially the sculpturedioxide which cause atmospheric pollution seriously. So we should Control the sulfur dioxideemissions of thermal power industry in order to control of environmental pollution of ourcountry. But in the process of production, there are many influence factors which influencethe so2’s emissions. It is difficultly to find reasonable analysis in order to find valuableinformation. Genetic algorithm is more quickly and higher accuracy in engineeringoptimization search and will not appear local optimal results.According to the analysis and study of desulfurization system, we can find that it is moreconvenient to use intelligent algorithms to realize the simulation and the optimization of thedesulfurization system with the higher accuracy, and it can eliminate effect of artificial factor.Neural network algorithm is to simulate the human brain neurons, and interconnects neuronswith the network system. It is been application quite common in engineering simulation withits self learning ability and massively parallel processing capabilities.Gradient descent is the main method to realize the changes of the neural network weightsand threshold. There are many types of gradient descent method. Momentum gradient descentmethod adds a momentum in the gradient descent method in order to control the speed anddirection of descent. Conjugate gradient descent method adds a conjugate direction, whichmakes the descent direction always along the conjugate gradient direction. Based on theabove two kinds of algorithm, this paper puts forward a conjugate gradient descent withmomentum, which increase the speed of descent and improve the accuracy of the simulation.Then we use genetic algorithms to optimization neural network simulation function. Becauseof the characteristic of the neural network, the optimization results will cross the border; thispaper presents a genetic algorithm based on the penalty value optimization model, whichrealize the effective control of genetic algorithm.更多还原