Combustion instability is one of the most plaguing challenges for general industrial thermal facilities, such as furnaces, gas turbine and hot air heaters. Considerable researches have been reported on developing effective active control means to mitigate combustion instability. The quick and accurate prediction of oscillating combustion characteristics is the premise of developing an active control model. In the present work, neural network models are constructed to predict temperature and key radical evolutions with historical data and system parameters as inputs. Active subspace (AS) method is adopted to identify the important direction of the system-parameter space in neural network, reduce the number of neurons in the input and middle layers, and simplify the network structure. Results show that the system-parameter space of eight dimensions can be reduced to a one-dimensional space using AS method. The neural network structures are greatly simplified, and the numbers of neurons are significantly reduced by 50%, which shortens the training time by 72%, and reduces single-step prediction time in hardware system by 60%. Compared with the neural network with full system parameters, the neural network with reduced parameters by AS method can capture the oscillating combustion characteristics and accurately predict the temperature evolution and mass fraction of radical H and CH.

对于诸如炉子，燃气轮机和热空气加热器的一般工业热力设施而言，燃烧不稳定性是最令人困扰的挑战之一。在开发有效的主动控制装置以减轻燃烧不稳定性方面已进行了大量研究。快速准确地预测振荡燃烧特性是开发主动控制模型的前提。在目前的工作中，构建了神经网络模型，以历史数据和系统参数作为输入来预测温度和关键自由基的演变。采用主动子空间（AS）方法识别神经网络中系统参数空间的重要方向，减少输入层和中间层的神经元数量，简化网络结构。结果表明，采用AS方法可以将八维系统参数空间缩减为一维空间。神经网络结构大大简化，神经元数量减少了50％，训练时间缩短了72％，硬件系统中的单步预测时间减少了60％。与具有完整系统参数的神经网络相比，采用AS方法减少参数的神经网络可以捕获振荡燃烧特性，并准确预测H和CH自由基的温度演化和质量分数。并将硬件系统中的单步预测时间减少了60％。与具有完整系统参数的神经网络相比，采用AS方法的具有减少参数的神经网络可以捕获振荡燃烧特性，并准确预测H和CH自由基的温度演化和质量分数。并将硬件系统中的单步预测时间减少了60％。与具有完整系统参数的神经网络相比，采用AS方法的具有减少参数的神经网络可以捕获振荡燃烧特性，并准确预测H和CH自由基的温度演化和质量分数。