项目名称: 基于部件神经网络模型的复杂制冷空调系统混合仿真方法
项目编号: No.51206123
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 工程热物理与能源利用学科
项目作者: 邵亮亮
作者单位: 同济大学
项目金额: 25万元
中文摘要: 制冷空调系统是能耗大户。为提高能源利用效率,制冷空调系统设计日趋复杂、对系统仿真技术的需求也不断增强。面对复杂制冷空调系统,系统仿真方法亟待突破以满足计算精度、速度、稳定性与通用性的综合要求。本课题提出并研究一种系统混合仿真方法:在部件层面发展神经网络建模方法,建立准确、快速、稳定的部件模型库;在系统层面保持热力学系统"框架"模型,保证系统模型的通用性。在部件神经网络研究上,通过对机理模型的无量纲化确定部件神经网络的无量纲输入输出参数,保证网络的最小化与通用性;通过自定义神经元激励函数改善网络性能;通过交叉验证保证训练样本集较小情况下神经网络的泛化性能。在系统热力学模型研究上,通过图论方法完成系统的通用数学描述;通过非线性方程组聚类算法实现快速稳健的系统仿真。以多联机系统为例验证混合仿真方法的精度、速度与稳定性。本课题的研究为自主研发高效通用的制冷空调系统仿真平台提供了新思路与理论基础。
中文关键词: 制冷;系统仿真;建模;人工神经网络;
英文摘要: Refrigeration and air-conditioning equipment is the major consumer of power. To improve the energy efficiency, the design of refrigeration and air-conditioning systems is getting more complex than before, which brings forward higher demands on the system simulation techniques. Facing up the complex refrigeration and air-conditioning systems, the system simulation method should be able to comprehensively address the needs of accuracy, speed, robustness and generality. In the project we proposes a hybrid system simulation methodology: at component level, we are going to develop accurate, fast, and robust neural network models, while at system level we will retain the generic thermodynamic modeling framework. To develop the component neural networks, physics-based models are used in dimensionless form to determine the dimensionless input/output parameters of a minimized generic neural network; the activation functions of neurons are customized to better fit the component characteristics; cross-validation are used to guarantee the generality of neural networks, particularly when the sample dataset is small. To develop the system thermodynamic model, graph theory is applied to describe the system architecture in a generic way. A new nonlinear equations solver using the clustering method to group different types of eq
英文关键词: Refrigeration;System simulation;Model;Neural network;