【作者】 杨广全；

【导师】 朱昌明；

【作者基本信息】 上海交通大学 ， 机械设计及理论， 2007， 博士

【摘要】 在经济不断发展、科学技术日新月异的今天,电梯作为建筑物内的主要运输工具,已成为我们日常生活中一个不可缺少的重要组成部分。随着建筑物规模的扩大及电梯部数的增加,电梯客流呈现新的变化,随之对电梯群控系统的性能提出了更高的要求。因此,一方面需要对不断发展变化的电梯交通流进行调查研究,分析其规律特征,另一方面要求研究处理电梯交通流的群控策略。本文针对电梯群控系统中的一些关键问题进行了研究,其主要成果如下:电梯交通流产生的理论和方法是深入研究电梯群控策略必须解决的关键问题。为了描述乘客分布,首次提出了电梯乘客起始-目标楼层(Origin-Destination, O-D)矩阵的概念。以交通流采集数据为基础,以热力学熵、信息熵理论为指导,建立极大熵模型,用以推算乘客O-D矩阵。极大熵模型的求解本质上是一个优化问题,首先提出了极大熵模型的拉格朗日乘子求解方法,该方法计算结果从实数解到整数解的推定中采用启发式搜索,得到的是次优解。为了对求解结果进行改进,经对遗传算法机制的深入思考,提出了一种启发式遗传算法求解方法,可以直接求得极大熵模型的最优解。基于乘客O-D矩阵,利用蒙特卡罗采样法进行客流仿真。随着交通测定的持续执行,利用极大熵模型可以产生持续的O-D交通流,建立乘客O-D交通流数据库,为群控系统的仿真试验奠定了数据基础。电梯交通流预测是实现电梯交通模式识别和电梯群控系统的重要组成部分。针对该问题,提出了将历史数据和最新的客流数据相结合,利用小波支持向量机(WSVM)建立了电梯交通流预测模型,预测模型采用SMO算法进行训练。通过与自回归滑动平均(ARMA)模型、BP神经网络、高斯核支持向量机(GSVM)三种方法预测结果的比较,说明了WSVM提高了预测准确性和预测精度。电梯交通模式识别是电梯群控系统的重要功能模块。为此,提出了基于高斯核粒子群K均值聚类算法的电梯交通模式识别方法。该方法以粒子群优化算法为框架,以适应度函数设计为核心。为了增强聚类算法的鲁棒性,采用高斯核距离替换欧氏距离设计适应度函数,利用M-估计的影响函数分析了聚类算法的鲁棒性。仿真实验表明该方法不需要任何先验知识就能达到准确分类的目的。粒子群K均值聚类算法需要调整的参数少,易于实现,计算速度快,且具有稳定的收敛特征,能很好满足群控系统实时性的要求。可作为电梯群控系统的一个模块,辅助电梯群控系统做出决策,以期提高电梯群控系统在各种交通状况下的服务性能。针对强化学习应用于电梯调度时存在学习速度缓慢问题,本文采用CMAC网络建立基于先验知识的电梯群强化学习系统,一个方面是利用先验知识缩小强化学习算法要探索的状态空间,加快强化学习算法的收敛速度,另一个方面利用CMAC神经网络具有较好的在线增量学习能力以及收敛性快、不存在局部极小点的特点,从这两个方面对电梯群控调度进行优化,试验结果表明其有效性。建立了电梯群控仿真环境,包括电梯群控仿真试验台(硬件设计)和电梯群控仿真系统(软件设计)。电梯群控试验台符合实际电梯群控系统的结构,可以满足电梯群控系统研究的需要。而电梯群控仿真系统为电梯群控策略的研究提供了仿真平台。在该仿真平台上进行了群控策略研究,结果表明基于交通模式识别,实现了各不同交通模式之间的合理调度。更多还原

【Abstract】 With the development of economy and the advance of science and technology, elevators as the main vertical transportation tool in the buildings have been an indispensable part in our daily life. As the size of building enlarging and the number of installed elevators increasing, elevator traffic flow takes on new changes and then the high quality of elevator group control system (EGCS) is required. Hence, on the one hand, elevator traffic flow need to be surveyed to analyze the pattern, on the other hand, it needs studying the group control strategy to handle the elevator traffic flow. The dissertation investigates some key technology of EGCS, and the main contribution is as follows:The theory and method of elevator traffic flow generation is a key problem to further study the elevator group control strategies. In order to describe the passenger distribution, the concept of the elevator passenger origin-destination (O-D) matrix is first put forward. Based on t elevator traffic data collected, in terms of the theory of thermodynamics entropy and information entropy, the maximizing entropy model is constructed to estimate the elevator passenger O-D matrix. The model is solved using Lagrange multiplier method and heuristic genetic algorithm respectively. The computational example shows that GA is significantly superior to the Lagrange multiplier method. Based on the passenger O-D matrix estimated, the Mento Carlo sampling method is utilized to simulation traffic flow. With the traffic measure executed continuously, the O-D flow is generated using the model and set up the O-D database. Elevator group simulation test is executed based on the O-D database.Elevator traffic prediction is the basis of elevator traffic pattern recognition and the important component of EGCS. To this question, wavelet support vector machines (WSVM) is presented to construct elevator traffic flow prediction model by combining the historical data and recent passenger data. The prediction model is trained by SMO algorithm. The simulation showed that the prediction result not only retains the tendency of historical data but also tracking the passenger flow change in time. Through the comparison of prediction results by BP neural network, ARMA and GSVM methods, it is showed that the WSVM can increase the prediction accuracy and preciseness.Elevator traffic pattern recognition is the important functional module. A new method of elevator traffic pattern recognition based on Gaussian kernel particle swarm optimization K-means clustering algorithm (GKPSOKCA) is proposed. In order to enhance the robustness of the clustering algorithm, fitness function is devised by Gaussian kernel metric replacing the Euclidean distance. The robustness of the clustering algorithm is analyzed using the influence function of M-estimator. Simulation shows that the GKPSOKCA can classify elevator traffic patterns precisely without any prior knowledge. Compared with IEKCA, GKPSOKCA needs less parameters to determine, can be easily implemented, and has stable convergence characteristic with good computational efficiency. Elevator traffic pattern recognition with GKPSOKCA can be used as a module in EGCS and assist EGCS in making decisions in order that EGCS improves the service performance under different traffic situations.Reinforcement learning system based on prior knowledge applying to elevator group control dispatching is significantly enhancing the convergent speed of reinforcement learning. The dissertation uses the CMAC network to build the elevator group reinforcement learning system. On the one hand, addition of prior knowledge can reduce the state space explored by reinforcement learning algorithm, and enhance the convergent speed of reinforcement learning. On the other hand, CMAC network has the good online incremental learning capability and good convergent speed, and hasn’t local minima. From the two sides, the elevator group dispatching is optimized. The simulation result verified its effectiveness.Lastly, the elevator group control simulation environment is designed, including elevator group control test-bed and elevator group control simulation system. Elevator group control test-bed conforms to the real EGCS structure and can satisfy the need for study on the EGCS. Elevator group control simulation system provides the simulation platform for the study on group control strategy. Based on the simulation platform, simulation study of group control strategy is executed and the results showed that reasonable dispatching is implemented among different traffic flow pattern based on traffic pattern recognition.更多还原