现代重型燃气轮机燃烧室的研发过程中,产生了大量的数值模拟和试验数据。其中,燃烧室火焰筒采用了先进发散冷却技术。该技术中,壁面冷效受发散孔型、主流/冷气压损等参数影响,难以建立基于理论模型和经验关联式的方法进行回归预测。针对工业研发产生的壁面冷效数据集,该文提出了基于人工特征选择、小批量化和归一化的数据预处理技术,建立反向传播(back propagation,BP)神经网络和径向基核函数(radial basis function,RBF)神经网络2种优化的神经网络模型,并通过交叉验证与网格搜索对算法模型进行超参数寻优,最终实现燃烧室壁面冷效的高精度场重建和高精度预测,场重建精度达90%,预测精度达85%,满足了工业燃烧室设计迭代中的实际工程需求。

[Objective] The effusion cooling technique is applied to hot sections in the modern heavy-duty gas turbine combustor, for example, the liner provides efficient protection from high-temperature gas using a small amount of cooling air. However, establishing a theoretical model or empirical correlations for regression and prediction of wall cooling efficiency is difficult. Because it is affected by many factors, including cooling hole type, patterns, and other flow parameters. The prediction methods based on computational fluid dynamics (CFD) techniques also have the disadvantages of long-term consumption and resource occupation. This paper proposes a prediction method based on artificial neural networks, namely back propagation (BP) and radial basis function (RBF) networks, that reach high accuracy for field reconstruction and cooling efficiency prediction. [Methods] This study collects simulation data from the design process of an industrial combustor and analyzed all dimensions of the wall adiabatic film cooling efficiency data set from the working conditions. Artificial feature selection, mini-batch, and normalization are used to preprocess the data. Important feature regions A and B are manually extracted by calculating the standard deviations of the cooling efficiency matrix considering different cooling air pressure loss conditions combined with CFD analysis. Then, the training set, the validation set, and the test set are divided based on cross-validation. For the field reconstruction, the BP network and the RBF network are selected. For its fast convergence and high precision, the BP network adopts the L-M algorithm. For the prediction, the BP network and generalized regression neural network (GRNN) are selected. The BP network adopts a Bayesian regularization algorithm to avoid overfitting. The GRNN network outperforms the RBF network for prediction and also takes less time for grid searching. For regression, mean square error (mse) and network prediction accuracy (accu) are defined to measure variance and deviation, respectively. [Results] The results showed that: 1) Through artificial feature selection and mini-batch, the samples in region A were compressed by ten times while perfectly extracting cooling efficiency characteristics. 2) For the field reconstruction, under the non-reactive condition 3, the field reconstruction accuracy of the optimized BP and RBF networks reached 92.2% and 95.5%, respectively. Under the reactive condition 10, the same networks reached 90.5% and 92.0% respectively. Compared with the BP network, the RBF network had better global approximation ability than the BP network. 3) For the prediction, under the reactive condition 10, the prediction accuracy of the BP and GRNN networks reached 88.3% and 86.8%, respectively, both lower than the field reconstruction accuracies. [Conclusions] The rapid field reconstruction and high prediction for wall adiabatic film cooling efficiency are realized by proposing a universal data preprocessing technique for the combustor wall cooling efficiency data set and an optimization technique for the BP and RBF networks. In summary, the field reconstruction accuracy reaches above 90%, and the prediction accuracy reaches above 85%, meeting the actual engineering requirements in the modern industrial gas turbine combustor design.