基于人工神经网络模型的需求管理协议的优化设计毕业论文外文翻译.docx

《基于人工神经网络模型的需求管理协议的优化设计毕业论文外文翻译.docx》由会员分享，可在线阅读，更多相关《基于人工神经网络模型的需求管理协议的优化设计毕业论文外文翻译.docx（14页珍藏版）》请在得力文库 - 分享文档赚钱的网站上搜索。

1、附录B 外文翻译文献译文：基于人工神经网络模型的需求管理协议的优化设计摘要节能系统需求的不断增加引起了人们对能源管理的各种奇思妙想。一种主要的节能方法在于对电力系统需求管理。任何需求管理计划的目标是控制用户对电力的需求从而为电力设备创建负载救援和提高系统安全性。通常情况下，需求管理方法是奖励那些愿意注册负载中断的用户，我们称这些形式的协议为激励相容的协议，协议规定给用户提供的奖励应该超过中断成本，同时也应对公司有利。各种设计方法的系统是不同的，并且过去的设计机制来源于博弈论的设计机制，此类的设计已被应用。在这个协议中，我们提出人工神经网络，通过训练以确定最佳方案。人工神经使用的学习算法是反向传

2、播学习算法，我们使用电力系统参数作为神经网络输入，而输出作为期望值。博弈论的设计原理作为从人工神经网络获得结果的目标。我们提出的神经系统是在IEEE14节点测试系统上测试的。关键字： 设计机制、博弈论、需求管理、人工神经网络、反向传播学习算法1. 引言在世界上的大多数国家，电能需求大大增加。很多的电力公司对于需求增加试图生产更多的电力，这问题产生了各种反应，实际上是资本密集型并且对环境有危害。一种可选且最佳的策略是需求侧管理（DSM），需求侧管理层通过改变用户的负荷大小和模式来改变用户用电，这是通过削峰、填谷、负荷转移、战略保护、战略负载增长和灵活的负载型来实现1。用到DSM策略的设备可以预测

3、中断加载模式且能应用到整个系统或在特定的位置。需求管理计划的一个基本要求是客户必须自愿参与这些项目，DSM策略已经使用了许多方法和设计技术，包括人工神经网络中的其他软件计算技术2-3，DSM策略也用到其他程序的结合或补充。作者对分布式发电（DG）应如何补充需求管理方案作出了解释4，经济分析获得表明电力公司不必限制单一的需求管理计划而且也可以用现有的补充需求管理方案。针对之前工作中普遍使用的需求管理，我们可以得出结论，它们获得电力公司和用户的一致好评。因此，目前最佳的需求管理协议就是我们提出人工神经网络的设计。本文安排如下:在第二节,我们介绍合同的概念。在第三个节,我们描述了我们在设计系统系统的

4、模拟场景。在第四节，对神经网络模型的需求管理协议进行详细说明。在第五节,对神经网络训练和测试的实验结果进行讨论。最后第六节，结论。2. 需求管理合同随着研究需求管理协议的深入，焦点已转向准确估计用户的停运成本5，6，7。作者证实电力设备需要一个可以准确地估计用户的停电成本的方案。这将使电力公司提供有吸引力的激励措施，从而吸引了用户自愿参与需求管理项目。这一优惠措施受到用户的认可被称为协议。协议被定义为公司和用户之间的协议，其中用户同意心甘情愿参与甩负荷，作为回报获得公司货币补偿。重要补充一点是货币的好处可以采取两种形式：现金支付或降低电价费。此外，用户可以选择摆脱他的整个负载或在甩负载有所限制

5、9。以往的需求管理协议的例子有存在过10，但它不是最佳。例如，在10被用于估计客户停运成本的调查中，最佳协议关键要求是详细的说明8，并且不能进行评估，那是因为空间约束。在接下一节中，我们描述了需求管理协议的模拟场景。3. 协议方法的模拟场景需求管理协议公式化应用到14节点测试系统，如图1，14节点有5个发电机和11个负载。该节点边际电价格（）是从MATPOWER11软件最佳功率流例子获得。任何电力公司在不同的场景模拟过程中，都有可能遇到提供日常运营，我们得到的协议值使用了博弈论的每一个场景，博弈论的输出作为神经网络的目标，而训练作为测试结果的基准值。图1 14节点测试系统模拟场景：1. 所有发

6、电机和负荷连接。2. 所有负荷和每个发电机的连接关闭。3. 所有与发电机连接的负荷值都设置为0。4. 所有发电机连接负荷的值从最大容量的1/2、1/4、1/8缩减。我们结束506个不同的独特用户案例的模拟，得到合同值x（），x（）是负荷削减，（）是每个负载网络的奖金（假设是每个负载代表用户）。博弈论设计原理的输入和输出（神经网络的目标）被传送到监测网络神经。4. 神经网络模型需求管理协议人工神经网络可被描述为一个数学模型，它是复制人体生物神经系统的结构和功能，大多数的神经网络数学模型，也被称为学习算法，最普遍和有效的学习算法是反向传播学习算法，在这个项目工作中所使用的就是反向传播学习算法。人工

7、神经网络的典型应用在分类，模式识别，优化和通常非线性任务的回归。它们也被使用在电力系统中的任务，如检测和定位电能质量扰动12，13和短期负荷预测14。在本文中，我们提出神经网络的应用需求管理协议的优化设计。应用神经网络在博弈论到这项工作中有许多优点，最重要的是实时的适用性。这是因为经过培训的电力系统的神经网络会给出自发需求管理的协议值，甚至变化的系统参数或拓扑，从而保证恒定最优。A. 神经网络的输入属性在这个项目中使用博弈论的方法所选择的输入属性是相同的值，在其它的神经网络我们也选择其他关键电力系统的值作为附加输入，以提高设计系统的性能。下面分别是神经网络的输入属性的使用：1. 功率数量削减：

8、x代表为兆瓦。在每个节点，我们假设用户愿意削减所有负载或削减它们负载的一小部分，x值的范围从0到94.2兆瓦。2. “电力中断性的价值”：由表示。这通常是从现有的功率流例子得到。它代表没有提供电能到特定位置的成本，单位美元（$），值的范围从37.34美元到42.71美元。3. 用户类型:用表示。这通常是归一化在0和1之间,作为区分用户愿意摆脱负载的程度，最愿意(= 1)，至少愿意(= 0)。4. 缩减成本：用C(,X)表示，我们假设是一二次成本函数，C值的范围从0到158.22美元。5. 总线电压幅度：这是电压幅度在每个负载的节点单位，值的范围从0.981.06单位。6. 母线电压相角：每个负

9、载节点的相角，值的范围为-14.891.35度。7. 无功功率：每个负载节点的无功功率，值的范围为-3.9至19 MVAr。8. 概率函数:对功率削减量我们利用一个概率函数成反比表示，用户减少则大量功率降低,反之亦然，值得范围从0.02到0.07。9. 案件编号：我们必须确定合同案件的总数，如所述之前共有506案件。上述输入的属性被馈送到作为训练和测试阶段的神经网络，培训和测试比率各为50：506案件使用了50意思是253案件是用于训练神经网络，而其余253案件用于测试神经网络，所选择测试比等于培训比，是为避免神经网络接触到的训练数据比测试数据多，从而产生偏差的情况，在下面的章节中，我们说明数

10、据预处理使用的程序。B. 数据预处理这关键阶段包括两个任务，首先输入属性是0和1之间，其次，我们采用一个合适的输出编码（二进制编码）的协议值。标准化技术通过确定每个属性列中的最大值和由该最大值除以该属性列中的所有值，表1 显示最大值输入属性被用来表示准化在各列的属性。 表一显示最大值输入属性被用来表示准化在各列的属性输出编码存在一些困难，因为我们有两个输出，协议总输出必须是负载削减和支付激励（两路输出）。为了克服这个困难我们采用二进制输出编码。这解决了两个输出的问题并协助神经网络模型的学习。一个32位的二进制代码被用来代表输出结果，在输出层产生了32个输出神经元（1神经元的每个二进制数位）。这

11、技术主要的好处是在神经网络的输出层相应增加突触权重，且提高了神经网络的学习。总共有32个神经元用以20个神经元削减的负荷和12个神经元的激励发放。C. 神经网络训练对于神经网络训练，我们利用反向传播学习算法，在神经网络选择反向传播算法是由于在进行训练和测试中实施简单和足够输入属性的可用性。图2显示了神经网络模型的最佳需求管理协议的结构/拓扑。图2 神经网络模型的结构/拓扑由于有许多输入属性，所以神经网络的输入层是由9个神经元组成，每个输入神经元接收到一个标准化的属性数值。单隐层有25个神经元，这个数字是从几个实验中确定的，其中隐藏的神经元是从1到50发生改变。 在隐藏层提供了25个神经元是最好

12、的结果。输出层具有32个神经元是根据32的间隔相对应，包括20个神经元的负荷削减和12个神经元的激励支付。神经网络的初始值是在-0.35和+0.35之间的随机值产生。经过几个实验调试得到学习系数和动能率值，一个0.0021的误差值被视为可接受值。神经模型最后训练参数:动能率(0.73)和学习系数(0.00517)。5. 实验结果神经网络模型的实现平台是在一个2 GB的RAM、Windows XP操作系统的2.2 GHz的PC机上执行，使用的软件是MATLAB v 7.9.0 (R2009b)。神经网络经过14000反复训练实现误差收敛在372.03秒（培训时间），神经网络测试时间（运行时间）是

13、0.8x10-4秒。表2是神经网络最后测试成功的参数和准确率，神经网络模型训练所获得的结果的如下：该训练属性准确度为94.25，受过训练的神经网络模型的进行测试的属性（神经网络属性不处理之前）精度为87.14。训练和测试平均精度结果为90.695%。表2 神经网络模型最终参数6. 结论在这项目中，呈现出一个基于神经网络的需求管理协议的优化系统设计。迄今博弈论已应用此项目中，从博弈论所获得的结果基准值是从神经网络训练得到的，提出神经网络包括两个阶段：首先，是神经网络数据处理阶段，其次是神经网络的训练阶段，数据处理阶段包括输入属性缩放或规范化，输出二进制编码。神经网络的训练阶段使用一个多层监督算法

14、，是基于反向传播学习算法，该神经网络模型训练成功要求372.03秒，用于被训练模型的测试数据（验证）要求0.8x10 -4秒，总的准确率为90.695。神经网络最主要优点是在协议方法使用博弈论，其设计在于实用性。一个成功训练的神经网络，仍然能给出最优需求管理协议的任意值，如系统参数改变或者电力系统压力，从而保证了任何时候协议值都有利于公司和用户，另一补充优点是在博弈论中所提出的神经网络模型是用较少的计算过程中得到的高准确率英文原文：A Neural Network Model for Optimal Demand Management Contract DesignAbstract The e

15、ver increasing need for energy efficient systems has led to various ingenious ideas about energy management. A major offshoot of this search for energy efficient solutions is demand management in power systems. The goal of any demand management program is to control the demand for electric power amo

16、ng customers thereby creating load relief for electric utilities and improving system security. Typically demand management contract formulations reward customers who willingly sign up for load interruption with incentives. These forms of contracts are termed incentive compatible contracts and the i

17、ncentive offered the customer should exceed interruption cost and at the same time should be beneficial to the utility. There are different systems to design these kind of contracts and in the past mechanism design from Game theory, has been used in the design of such contracts. In this work we prop

18、ose an artificial neural network which is trained to determine the optimal contract. The learning algorithm utilized by the artificial neural network is the back propagation learning algorithm where useful power system parameters serve as the neural networks input while the neural systems output is

19、the contract value. Game theorys mechanism design serves as the target for results obtained from the artificial neural network . Our proposed neural system is tested on the IEEE 14 bus test system.Keywords - Mechanism design; Game Theory; Demand Management; Artificial neural networks; back propagati

20、on learning algorithmI. INTRODUCTIONIn most nations of the world, electrical energy demand has increased tremendously. It is common for electrical utilities to react to this increase in demand by attempting to generate more electricity. The problem with this kind of response however is the fact that

21、 it is capital intensive and is often harmful to the environment. An alternative and preferred strategy is Demand Side Management (DSM). Demand side management seeks to influence customer use of electricity by altering the magnitude and pattern of the customers load. This is achieved either by peak

22、clipping, valley filling, load shifting, strategic conservation, strategic load growth and flexible load shape 1. DSM strategies are used whenever the utility foresees disturbing loading patterns and can be applied either system wide or at specific locations. A fundamental requirement of demand mana

23、gement programs is that customer must participate in these programs voluntarily. DSM strategies have been designed using a host of methods and techniques including artificial neural networks amidst other soft computing techniques 2-3. They can also be used in conjunction with or to complement other

24、programs. In 4 authors explain how Distributed Generation (DG) can complement demand management schemes. Economic analysis obtained from 4 indicate that utilities need notrestrict themselves solely to demand management programs but can complement existing demand management schemes with distributed g

25、eneration .In view of the widespread use of demand management in prior works we can conclude that they are useful for electric utilities and consumers alike. We therefore present our proposed artificial neural network designed to give optimal demand management contracts. This paper is arranged as fo

26、llows: In section two we introduce the concept of contracts. In section three, we describe the simulated scenarios we used in designing our system. In section four, the neural network model for demand management contracts is described in detail. In section five, the experimental results of training

27、and testing the neural network are discussed. Finally, in section six the paper is concluded. II. DEMAND MANAGEMENT CONTRACTSAs the research on demand management contracts have deepened, the focus has turned to accurate estimation of customers outage cost 5, 6, 7. In 8 the authors prove that an elec

28、tricity utility needs to be able to accurately estimate customers outage cost. This will enable them offer attractive incentives to customers and hence entice the customers to participate voluntarily in demand management programs. Acceptance of this offer by the customer is termed a contract. A cont

29、ract is defined as an agreement between utility and customer wherein the customer agrees to willingly shed load and in return receive monetary compensation. It is important to add that the monetary benefit can take two forms: cash payments or reduced electricity tariffs. Furthermore the customer can

30、 choose to shed his entire load or have a limit on load shed 9. Past examples of demand management contracts exist 10, but they suffer from non-optimality. For example in10 surveys are used to estimate customers outage costs. Key requirements for optimal contracts are explained in detail in 8 and ca

31、nnot be reviewed here because of space constraints.In the next section we describe the simulated demand management contract scenarios.III. S IMULATED SCENARIOS FOR CONTRACT FORMULATIONDemand management contract formulations are applied to the 14 bus test system in Fig. 1. The 14 bus example has 5 ge

32、nerators and 11 loads. The Locational Marginal Price ( ) is obtained from optimal power flow routines with the MATPOWER 11 software. Simulations of different scenarios any electricity utility might encounter in the course of daily operations are provided and we obtain contract values using Game Theo

33、ry for each scenario. The output from Game Theory serves as the target for the neural network while training and also serves as a benchmark the testing results. Figure 1. IEEE 14 Bus Test System Simulated Scenarios:1. All the generators and loads are connected2. All loads are connected and each gene

34、rator is in turn switched off3. All generators are connected and each of the loads is in turn set to zero4. All generators are connected and all the loads are then curtailed to 1/2.1/4and 1/8of their maximum capacity respectively. We end up with a simulation of 506 different unique customer cases an

35、d obtain contract values x ( ) which is the load curtailed and y ( ) which is the incentive paid for each load in the network (the assumption is that each load represents a customer). The inputs to the Game theory mechanism design are also fed to the supervised neural network with the output from me

36、chanism design serving as the target to the neural network.IV. NEURAL NETWORK MODEL FOR DEMAND MANAGEMENT CONTRACTSAn artificial neural network can be described as a mathematical model that copies the structure and function of the human biological neuron system. Most neural network mathematical mode

37、ls are also termed learning algorithms and the most popular and effective learning algorithm is the back propagation learning algorithm which is the learning algorithm used in this work. Typical applications of artificial neural networks are in classification, pattern recognition, optimization, and

38、regression tasks which are usually non- linear tasks. They have also been deployed in power systems in tasks like detecting and locating power quality disturbance 12, 13 and short term load forecasting 14. In this paper we present an application of neural networks to the design of optimal demand man

39、agement contracts. Applying neural networks over game theory to this task has a number of advantages, the most important being real time applicability. This is because a trained power system neural network would give spontaneous demand management contract values even with changing system parameters

40、or topologies thus ensuring constant optimality.A. Input Attributes of Neural Network In this work the input attributes chosen are the same values we used with the game theory formulations. We also chose other key power system values as additional inputs to the neural network in other to improve the

41、 performance of the designed system. The following were the neural network input attributes used:1. Amount of power curtailed: Represented by x MW. At each of the nodes we assume that the customer is willing to curtail either all of their load or a fraction of their load. In this work, the values ra

42、nge from 0 to 94.2 MW.2. The “value of power interruptibility”: Denoted by . This is usually obtained from existing power flow routines. It represents the cost of not delivering power to a particular location and it is in U.S. dollars ($). In this work, values range from $37.34 to $42.71.3. Customer

43、 Type: Denoted by . This is usually normalized between 0 and 1 and serves as a differentiator of customer willingness to shed load. Most willing ( =1) and least willing ( =0).4. Cost of curtailing: Denoted by ) , ( x c .We assume a quadratic cost function. In this work, values range from 0 to $158.2

44、2.5. Bus voltage magnitude: This is the voltage magnitude in per unit at each load node. In this work, values range from 0.98 to 1.06 per unit.6. Bus voltage phase angle: This is the phase angle at each load node. In this work, values range from -14.89 to 1.35 degrees.7. Reactive Power: The reactive

45、 power at each load node. In this work, values range from -3.9 to 19 MVAr.8. Probability Function: We make use of a probability function thats inversely proportional to the amount of powercurtailed, the reasoning being that the probability of a customer curtailing a lot of power is low and vice vers

46、a. Inthis work, values range from 0.02 to 0.079. Case Number: The total number of cases we have to determine contracts for. As stated before there are a total of506 cases.The above input attributes are fed to the neural network for both the training and testing phases. Training to testing ratio of 5

47、0%:50% is used with the 506 cases meaning 253 cases are used for training the network while the remaining 253 cases are used for testing. An equal training to testing ratio was chosen to avoid a situation where the neural network is exposed to more training data than testing data thereby biasing the

48、 network. In the following section we explain the data pre-processing procedures used.B. Data Pre-ProcessingThis crucial phase consists of two tasks. Firstly the input attributes are normalized between 0 and 1. Secondly; we employ a suitable output coding (binary coding) for thencontract values. The normalization technique used works by determining the maximum value in each attribute column and dividing all the values in that attribute column by that maximum value. Table I shows the maximum values for the input attributes used to normalize other attributes in their respective columns