基于支持向量机与高斯分布估计的低NOx排放

燃烧优化的核心在于建立有效而快速的建模工具及寻优算法，以便于在线应用。为了研究新方法的适用性以及克服常用算法的缺点，本文利用支持向量回归建立了大型四角切圆燃烧电站锅炉NO_x排放特性模型。利用大量的热态实炉试验NO_x排放数据对模型进行了训练和验证。结果表明，支持向量回归模型能获得较神经网络模型更加准确的预测结果，相对于神经网络，支持向量回归能更好处理大样本量数据的非线性问题。随后，采用一种基于高斯概率密度 (GPDD)的分布估计优化算法对NO_x排放模型进行了寻优。研究发现，与遗传算法相比，GPDD具有更好的寻优能力与更快的收敛速度。结合支持向量回归与高斯概率密度分布(GPDD)算法能有效降低燃煤锅炉NO_x排放量，不到1 min的优化时间便于在线应用。研究结论可为该算法在实际电厂中推广应用提供参考依据。

Low NOx emissions based on support vector machine and Gaussian estimation of distribution

Quick and effective modeling tools and searching algorithms are the critical issues in realizing on-line combustion optimization in coal-fired utility boilers.In order to explore the applicability of novel modeling tools and optimization algorithms and to overcome the drawbacks of the existing methods, in the present study support vector regression (SYR) model was proposed to capture the functional relation between the NO_x emissions and operational parameters of a utility boiler.A large number of thermal field test samples, which were recorded by DCS in the actual power plants, were employed to establish the models.It was found that the predicted NO_x emissions by SVR showed better agreement with the measured than those by neural networks.SVR model was more suitable to non-linear problem with a large number of samples.Subsequently, a Gaussian probability density distribution (GPDD) based optimization algorithm was described and then applied to searching the optimal inputs of SVR model for NO_x reduction.The results showed that GPDD outperformed the existing GA.Less than one minute of optimization time period required for GPDD was suitable for on-line application.The current work will lay a foundation for the further extension of GPDD’s application to actual power plants.