目前，我国的发电方式仍然以燃煤火力发电为主。随着能源危机的日益凸显和环保意识的深入人心，电站燃煤锅炉的节能环保问题成为亟待解决的要务之一。要想实现电厂锅炉节能环保的燃烧，一要依赖建立的精确锅炉燃烧特性模型，二要取决于高效的优化技术。然而，锅炉的燃烧过程涉及到复杂的湍流、传热传质等多个化学反应，且多个运行参量高度耦合，导致传统机理式建模方法和经典优化算法难以解决锅炉燃烧优化问题。因此，本文对计算智能技术中的超限学习机(Extreme Learning Machine，ELM)和涡流搜索(Vortex Search，VS)算法进行深入研究，将其改进应用到锅炉燃烧特性建模和可调运行参量优化中，以此来实现锅炉高效和低氮氧化物排放的燃烧。该研究工作不仅具有很强的理论意义，而且还有很广阔的应用前景。主要研究内容如下： 首先，针对涡流搜索算法后期收敛速度慢，易陷入局部最优的不足，提出两种改进的VS算法。第一种是利用逻辑自映射混沌增加其局部开发性能和利用Lévy飞行策略增加其全局探索能力；第二种是利用Bloch球面坐标改善初始候选解质量和利用广义反向学习加快收敛速度。然后，采用13个著名经典基准函数进行测试，并利用Wilcoxon符号秩检验，在统计意义下与VS算法、其它智能优化算法进行比较，验证上述改进算法的可行性和有效性。 其次，针对ELM泛化能力弱的不足，根据样本的输入变量与输出变量的非线性程度，合理地在ELM中增加松弛因子，提出自适应超限学习机(Adaptive ELM， AELM)，并给出收敛性分析的理论证明。然后，利用UCI数据库中的11个标准回归数据集进行测试实验。最后，考虑到在线学习算法能够降低空间和时间复杂度，并能够提高算法的实时性，利用矩阵MP广义逆理论，结合最小二乘的思想，给出该网络的在线学习算法。 再次，针对并行感知超限学习机(ELM with Parallel Layer Perception，PELM)随机生成非线性部分的输入权值和阈值，影响泛化能力和稳定性能的问题，利用输入样本和方程组的最小范数最小二乘解生成下层网络的输入权值和阈值，提出双线性并行感知超限学习机(Two-linear PELM，TPELM)。然后，通过UCI中11个标准回归数据集验证TPELM的有效性。最后，给出基于输入样本增量的在线TPELM学习算法。 然后，根据循环流化床锅炉燃烧历史数据，利用TPELM和AELM分别建立锅炉氮氧化物排放浓度和热效率的批量离线模型。通过与PELM、ELM和其它模型详细的比较，验证所提出的模型具有更加优秀的泛化性能及稳定性能。为了实现实时跟踪锅炉的燃烧动态，而不再重复学习已经训练过的数据，利用OAELM和OTPELM分别建立氮氧化物排放浓度和热效率的在线模型。 最后，基于建立好的循环流化床锅炉燃烧特性模型，应用改进的涡流搜索算法离线和在线优化锅炉热效率或氮氧化物排放浓度，实现热效率提高或氮氧化物排放浓度降低的目标。进一步地，利用Pareto占优的思想，在第一种改进的涡流搜索算法中引入外部存储库，同时优化热效率和氮氧化物排放浓度，给出多组合理的可调参量配比方案，使热效率提高的同时氮氧化物排放浓度降低。 关键词：计算智能技术；涡流搜索算法；自适应超限学习机；双线性并行感知超限学习机；锅炉燃烧优化；热效率；氮氧化物排放浓度

 At present, coal-fired thermal power generation is still the main form of power generation in China. With the increasing energy crisis and the awareness of environmental protection, the realization of energy saving and environmental protection of coal-fired boilers in power plant is one of the urgent problems to be solved. In order to realize the optimal operation state of boiler combustion, one depends on accurate combustion characteristics model of a boiler, and the other depends on the efficient optimization technology. However, combustion process of a boiler involves many chemical reactions such as turbulence, heat and mass transfer. Furthermore, multiple operating parameters are highly coupled, so it is difficult to solve the problem of combustion optimization by the traditional mechanical modelling method and the classical optimization algorithm. Therefore, Extreme Learning Machine (ELM) and Vortex Search (VS) algorithm in computational intelligence technology are deeply studied. Their improvements are applied to establish combustion characteristic models and optimize adjustable operation parameters of a boiler. In this way, boiler combustion with high efficiency and low Nitrogen oxide emission can be realized. The research not only has a strong theoretical significance, but also has a very broad application prospect. The main research contents are as follows: Firstly, in order to solve the problems that the VS algorithm converges slowly and is easy to fall into local optimization, two improved VS algorithms are proposed. The first one is to use Logic self-mapping chaos to increase its local exploitation performance and Levy flight strategy to increase its global exploration ability; the second one is to use Bloch spherical coordinates to improve the quality of initial candidate solutions and generalized opposition-based learning to accelerate the convergence speed. The test is carried out by using 13 well-known classical benchmark functions to verify the feasibility and effectiveness of the improved algorithms. Wilcoxon signed rank test is used to compare the VS algorithm and other intelligent optimization algorithms in statistical sense Secondly, in view of weak generalization ability of the ELM, according to the nonlinear degree of the input variables and output variables of the sample, a relaxation factor is reasonably added to the ELM, and an adaptive ELM (AELM) with the relaxation factor is proposed. The theoretical proof of convergence analysis is given. Then, 11 standard regression data sets in UCI database are used to test and compare the generalization ability of AELM with other networks. Finally, considering that the online learning algorithm can reduce the complexity of space and time, and improve the real-time performance of the algorithm, the online learning algorithm of AELM is given by combining the idea of least squares and using MP generalized inverse of the matrix. Thirdly, since the input weights and thresholds of the nonlinear part of the ELM with parallel layer perception (PELM) are randomly generated, the generalization ability and stability of PELM are affected. A two-linear PELM (TPELM) is proposed by using the input samples and the minimum norm least square solution of the equations to generate the input weights and threshold of the lower network. Then, the validity of TPELM was verified by 11 standard regression data sets in UCI. Finally, an online TPELM learning algorithm based on the increment of input samples is presented. Fourthly, according to the historical combustion data of circulating fluidized bed boiler (CFBB), the batch offline models of Nitrogen oxide emission concentration and thermal efficiency of the boiler are established by using TPELM and AELM respectively. Through detailed comparison with PELM, ELM and other models, it is verified that the proposed models have better generalization performance and stability performance. In order to track the combustion dynamics of the boiler in realtime, and no longer repeat to learn the data that has been trained, the online models of NOx emission concentration and thermal efficiency are established using OAELM and OTPELM respectively. Finally, based on the established combustion characteristic model of CFBB, the improved VS algorithms are applied to offline and online optimizations of boiler thermal efficiency or Nitrogen oxide emission concentration to achieve the goals of increasing thermal efficiency or decreasing Nitrogen oxide emission concentration. Furthermore, by using the idea of Pareto dominance, the external repository is introduced into the first improved VS algorithm. Then, the algorithm is applied to simultaneous optimization of thermal efficiency and Nitrogen oxide emission concentration. The multi-group ratio schemes of the adjustable parameters are given, which can increase the thermal efficiency while decrease the nitrogen oxides concentration. Keywords: computational intelligence technology; Vortex Search algorithm; adaptive Extreme Learning Machine; two-linear ELM with parallel layer perception; boiler combustion optimization; thermal efficiency; Nitrogen oxide emission concentration