【作者】 赵灵晓；

【导师】 谷波； 张春路；

【作者基本信息】 上海交通大学 ， 制冷与低温工程， 2010， 博士

【摘要】 制冷空调装置仿真技术已经有了相当的发展,为制冷空调装置的设计节能优化等研究做出了巨大的贡献。基于部件物理模型的系统仿真方法具有很好的通用性和趋势控制能力,是常见的制冷装置仿真方法。在不同的阶段和不同的情况下,仿真技术需要满足不同的需求。随着时代的进步,人们对仿真的要求不断提高,除了准确性之外,计算速度和求解稳定性成为了新的明确的需求。为了进一步提高制冷系统装置仿真技术的计算速度和求解稳定性,本文在“面向部件”的制冷系统仿真研究的基础上,引入神经网络模型提出了基于部件神经网络模型的制冷系统混合仿真方法。“面向部件”的系统仿真和神经网络模型各有其显著的优缺点,需要在两者之间寻找一个契合点以扬长避短。混合仿真方法是仿真需求不断提升的产物,是基于部件物理模型的制冷系统模型之外的一个延伸和扩展。本文的主要研究内容包括:1.本文明确提出了基于部件神经网络模型的混合仿真方法。详细阐述了混合仿真方法的必要性和实现过程。在系统层面,混合仿真方法保持了“面向部件”的系统仿真结构的的灵活性,部件模型可以重复利用,而且可以方便地增减或修改部分辅助部件;在部件层面,部件神经网络模型有助于降低部件模型的复杂程度,提高计算速度。基于部件神经网络模型的混合仿真方法既考虑了制冷系统的灵活性,又兼顾了计算速度和稳定性。为制冷系统仿真提供了新的可供选择的方向。2.深入分析神经网络建模的五个基本步骤,针对各个步骤可能存在的风险和问题展开了神经网络方法的改进研究,从而尽可能地降低神经网络模型的过拟合风险、提高模型在大范围工况内的精度、改善模型的趋势等。本文所提出的改进方法主要包括:a)对数据样本进行恰当合理的处理以改善模型的训练精度和趋势合理性,比如增加理论点、保持数据样本随输出参数的分布均匀性等。b)采用和物理模型相结合的方式对神经网络模型的输入输出参数进行分析选择,避免了参数的冗余和不足。c)自定义多项式传递函数,从理论上证明了多项式神经网络模型和多项式函数的等价性,有效避免了神经网络模型的过拟合的风险。d)根据研究对象的特性,优化神经网络的结构,既显著提高了神经网络模型的灵活性,又提高了模型的训练效率。作者应用这些改进方法建立了若干相关部件性能的神经网络模型,包括容积式压缩机、螺杆压缩机的多项式神经网络模型、毛细管和短管流量特性的通用神经网络模型、翅片管冷凝器性能的神经网络模型和翅片管蒸发器性能的神经网络模型。这些模型在部件层面进行了充分的精度和趋势验证,可以满足系统仿真的需求。3.混合仿真方法的实例验证。在部件模型的基础上实现了制冷系统的混合仿真,并进行了实验验证和分析:a)模拟了带经济器的水冷冷水机组的性能预测。由于采用了高精度而且连续的螺杆压缩机的神经网络模型,水冷冷水机组的模型实现了从满负荷到卸载负荷的连续仿真,显著改善了水冷冷水机组的部分负荷下的预测精度,96%的点误差在±5%以内。作者基于此高精度的冷水机组模型展开了经济器开关最优切换点的研究。b)模拟了轻型商用空调的性能预测。由于部件神经网络模型避免了迭代,随着部件物理模型不断被神经网络模型替代,在保证系统模型精度的前提下混合仿真方法不断提高了系统模型的计算速度。最终,与基于部件物理模型的系统模型相比,基于四部件神经网络模型的混合仿真方法的计算速度提高了12倍左右。最后,作者简要阐述了本文工作存在的不足和进一步的研究设想。更多还原

【Abstract】 Computational simulation on refrigeration and air-conditioning appliances has been fairly developed for several decades, and has made a huge contribution to the design of products, optimization and energy saving.“Component-oriented”system modeling, which bases on physics-based component models, has good generality to be applied to different systems and can give reasonable trends prediction. Computational simulation should meet different needs under different phase and conditions. Currently, the simulation speed and robustness are new and clear requirements besides accuracy.In order to meet these requirements, a new approach of refrigeration system simulation has been presented in the present work, namely, hybrid modeling of refrigeration system based on Neural Network (NN) component model. The hybrid modeling integrates the merits and avoids the shortcomings of“component-oriented”system modeling and NN modeling. The hybrid modeling is the product of new modeling requirement, and can be regarded as an extension of system model based on physics-based component model. In detail, the main contents of present work include:1. Propose the hybrid modeling of refrigeration system based on NN component model clearly. What is hybrid modeling, why propose and how to realize hybrid modeling are stated in detail. In system modeling level, the hybrid modeling keeps the generality of“component-oriented”system model; in component modeling level, NN model simplifies component model and remarkably increases simulation speed. The hybrid modeling shows many merits including the genrality, quick simulation and good robustness. Therefore, the hybrid modeling is a new alternative solution of refrigeration system modeling.2. Improvement of five steps in development of NN model. Deep analysis was carried out at first, and then corresponding improvement methods to overcome shortcomings and mitigate risks in each step were proposed respectively. NN model property is improved, including reduction of over-fitting risk, improvement of accuracy in a large range, prediction of tendency and so on. The main improvement methods include:a) Reasonable processing of data sample to improve model’s accuracy and trends, such as adding theoritical points, even distribution of training data along with outputs other than inputs, and so on.b) Choose input and output parameters based on analysis of physics-based model to avoid redundancy or omission parameters.c) Specify polynomial transfer function which is proved identical to polynomial correlation. Over-fitting risk is avoided by this way in theory.d) Optimize NN structure according to the property of the object.Through this way, both flexibility of neural network and training efficiency are improved.3. Application of the hybrid modeling. Develop refrigeration system modeling based on developed component models, and then do experimental validation and analysis:a) Performance simulation of two economized water-cooled screw chillers. Due to the accurate and continuous NN model of screw compressor, the chiller model can simulate from full load to unload continuously, and the prediction accuracy under part load conditions are improved obviously.b) Performance simulation of a light commercial air-conditioner. Since NN component model avoids iteration, its simulation speed is much quicker than physics-based model. In turns, the hybrid model of the light commercial air-conditioner based on four NN component models can save simulation time dramatically. Compared to physics-based system model, simulation speed is improved around 12 times, and system model runs more robustly under working conditions.Finally, the author summarizes the present work and proposes the further research ideas in the field.更多还原