基于神经网络的变截面再生冷却结构优化设计

针对目前再生冷却结构优化研究存在参数对比范围窄、依赖于经验关系式等问题,根据航空发动机燃烧室特点提出了一种变截面宽度再生冷却通道,并采用神经网络模型结合数值模拟结果,以通道出口燃油温度相对标准差、燃气侧最高壁温及壁温相对标准差为目标,预测了全参数范围内不同槽宽和、槽宽比及肋高下目标函数的变化规律,预测结果表明：当槽宽和较小时,增大肋高可以强化换热,但当槽宽和较大时,需减小肋高才能强化换热,这也揭示了为何有些文献中关于肋高对换热性能影响的结论会相反；此外,存在一个最佳槽宽比范围可使得三个目标函数均最低；增大槽宽和可以明显降低燃气侧壁温及其不均匀度,减小肋高可以缩小不同管道出口燃油温度的差异。从预测空间内可选取多组综合流动换热性能较优的结构,优化后三组目标函数的加权值降低了9.09%。

Structural Optimal Design of Regenerative Cooling with Variable Section Based on Neural Network

Nowadays, there are some problems such as narrow range in parametric contrast and dependence on empirical relations existing in the research of structural optimization about regenerative cooling. According to the characteristics of aero-engine combustors, a new structure of channels adopted regenerative cooling with variable sectional width was proposed. Neural network combined with numerical stimulation results was selected with the aim of RSD (Relative Standard Deviation) of fuel temperature in channel exit, maximum temperature of hot wall and RSD of hot wall temperature. Therefore, the changing rules of targeted function in different slot width sum, slot width ratio and fin height within the full parameter range were predicted. The results demonstrated that when the slot width sum is relatively small, heat transfer will be strengthened by increasing the fin height. However, when the slot width sum became relatively large, heat transfer will be enhanced by reducing the fin height. This explains the reasons why controversial conclusion about the relationship between fin height and heat transfer was presented in several essays. Moreover, there is an optimal range of slot width ratio that can make all these three objective functions lowest. The temperature of hot wall and its non-uniformity decline by increasing slot width sum, and the fall of the fin height leads to the difference of fuel temperature at the outlet of channels deceasing. Multiple structures with better performance in comprehensive flow-heat transfer can be obtained from the range of prediction. After this optimization, the weighted value of three objective functions dropped by 9.09%.