基于某116 MW燃气热水锅炉的实际运行数据，采用多层感知器（MLP）神经网络和支持向量回归机（ε-SVR）数据辨识方法对其运行特性进行仿真建模，分析燃气消耗量及NO〖HT5”〗x排放量与锅炉运行工况之间的关系，并将两种方法的精确度和泛化能力进行比较。对比发现：MLP模型预测的燃料消耗量与实际数据的误差在-2%~3%之间，NO〖HT5”〗x的排放量误差在±5%以内；而ε-SVR模型预测的燃料消耗量误差在±2%以内，NO〖HT5”〗x的排放量误差在±3%以内，ε-SVR方法具有更高的准确性和泛化能力。

Simulation modeling on the performance characteristic of boiler was built through multilayer perceptron neutral network (MLP) and support vector regression (ε-SVR) data recognition algorithm based on the operating data of a 116 MW gasfired water boiler.The relations among gas consumption,NO〖HT5〗x discharge and operating conditions of boiler were analyzed and the accuracy and generalization ability were compared between MLP and ε-SVR algorithms.It indicates that the error of gas consumption predicted by MLP algorithm is between -2% and 3%,and about ±5% for the NO〖HT5〗x discharge,while the accuracy of gas consumption and NO〖HT5〗x discharge predicted by ε-SVR algorithm is about ±2% and ±3%,respectively.The results demonstrate that the ε-SVR algorithm has high accuracy and good generalization ability.

TK223

北京市重点研发计划（D171100001217001）