智能电网中的客户行为分析电力行业

发布时间：2018-04-27 03:15

本文选题：极限学习机 + 人工神经网络　；参考：《北京交通大学》2017年硕士论文

【摘要】：基于通信、控制、IT技术的智能电网系统现在已成为全球趋势。通过客户行为预测未来电网负荷(电力使用)是向智能电网的重要任务,精确的预测可以帮助公用事业公司制定合理的资源分配计划,采取控制措施来平衡供电和电力需求。在竞争激烈的电力市场中,电力负荷预测对消费者和电力生产商来说是至关重要的,既可以使消费者了解自己的用电习惯,又可以帮助生产商根据客户的消费习惯制定特定的产品,从而规划运营和防止电力风险。另外,预测在电力经济优化中也起了非常重要的作用。在本文中,我们提出了一个新的数据挖掘框架来分析客户行为,以预测未来时间智能电网中特定消费者实体的负载。然后,利用极限学习机(ELM)分析集群用户电力行为的相似度,收集用户电力负荷,将具有相似行为的用户分类到相同的模型中预测,这样可以增加模型的适应性。为了证明所提方法的有效性·我们分别从理论和实验去分析。极限学习机是一种新型的机器学习算法,其随机初始化网络节点权值和偏置的策略可以解决单层前馈神经网络训练和优化慢的问题,并可取得全局最优解。最后,我们使用山东省电力公司的运监系统数据(包括设备信息、线路信息、用户信息、负载信息等)和可能影响负荷变化的外部系统数据(如天气信息),在MATLAB平台进行了仿真实验。实验结果表明,该方法能够深入挖掘用户电力行为,通过合理的用户聚类提高负荷预测的准确性,揭示预测精度与集群数量之间的关系。随着智能电网技术的发展,先进的计量基础设施(AMI)和各种监控系统的大量部署生成并积累了大量的数据。智能电表是AMI的重要组成部分,可以在一定时间内(如每15分钟或者每60分钟)获得精确的用户消耗的电力负荷。与传统电网系统相比,智能电表收集数据频率较高,能够生成更多的数据。但是,累积的大数据一直处于搁置状态。随着机器学习算法和大数据的发展,我们可以对电力大数据进行分析,充分挖掘这些隐藏在这些数据的背后的价值。例如,基于运监系统中的设备和客户数据,结合聚类算法挖掘用户用电行为,基于智能电表数据和分类回归算法,预测未来负载的变化。[37]负载预测一直是电力系统安全发展的关键,因为它可以影响了许多有关电力系统的决策,如经济调度,自动发电控制·安全评估,维护调度和能源商业化。精确的负载预测可以在经济合理的情况下启动和停止电力系统发电机组,在维护安全稳定方面发挥重要作用,保持社会正常生产和生活,有效降低发电成本。通常情况下,按照负荷预测时间的长短,负载预测可分为三类:短期负载预测,中期负载预测,长期负载预测。其中,短期负荷预测的预测时间范围是未来1小时,一天或一周。中期负载预测的预测时间范围大概是未来一个月。长期负载预测的时间范围则是未来一年,甚至三至五年。本文主要对用户电力负荷进行短期负荷预测。负载预测对能源管理系统的实时性和安全性中起着主要作用,准确的预测有利于电力系统的规划者完成各种任务,如发电量的经济调度·燃料采购的调度等。难题是·负载预测是一项艰巨的任务,因为其变化受到许多因素的影响·如天气条件,是否是节假日,人口流动,经济状况和客户的用电习惯。不准确或错误地负载预测可能会增加运营成本。据观察,电力需求预测误差仅增长百分之一,导致英国电力系统运营成本增加了 1000万英镑。这是负载预测效用类型的严重失误。而且,糟糕的负载预测会误导了规划者,导致错误和昂贵的扩张计划。高估未来的电力负荷可能导致多余的储备的电力,对负载的低估导致提供足够电力的故障。相反,准确的预测可使公用事业提供商提前计划燃料等资源·并采取控制措施·如开启/关闭需求响应装置和修订电价等。同样地,高估未来的电力负荷可能导致多余的储备的电力。相反,对负载的低估导致提供足够电力的故障。无论计划者低估还是误判负荷,高精度的负载预测技术需要先进的技术、自适应的预测模型。虽然不同的模型在动态系统中有一些优势·但改善相关缺点的可能性是不能排除的。因此,需要开发最佳和准确的负载预测模型来改善(最小化)预测误差。通过对多种数据挖掘算法、机器学习算法的分析,我们致力于提出高精度的负载预测模型。极限学习机是新提出的机器学习算法,不仅效率高而且可以防止过拟合。因此,该项目的主要研究问题是:融合跨系统的数据,进行数据预处理;对数据进行分析,挖掘影响负载变化的强关联特征;利用极限学习机构建负载预测模型,调整参数获得精度最高的负荷预测结果。在大数据的背景下·影响负载变化的因子众多,电力负荷预测是一项复杂的工作,其呈现复杂的非线性变换。传统的电力负载预测模型,大都是线性模型,缺乏非线性映射能力。因此,以前的预测方法根本不适应大数据时代的发展。另一方面,智能电网缺乏通用的训练框架,为电力系统的其他任务提供支撑,以提高智能电网分析的效率。自1990年以来,研究人员就开始关注电力负载预测问题并提出了很多预测模型[1]。如提出的时间序列分析模型ARIMA,该模型重点在于分析负载根据时间的变化曲线,从而预测未来负荷的变化,其优点是简单·缺点是没有考虑影响负载变化的因子。使用模糊逻辑方法Fuzzy Logic分析负载的变化,该方法考虑了影响负荷变化的各种影响因子,如温度、湿度等,但其没有考虑设备、线路信息。使用人工神经网络方法构建负荷预测模型,神经网络的优化和训练较慢,而且对复杂的负载变化,较易陷入局部最优值。使用支持向量回归SVR,该方法将特征映射到核空间·可以取得全局最优·但训练效率交低。通过对已有负载预测方法的调研·本文提出基于极限机器学习ELM的负荷预测模型·并通过实际的智能电网数据去验证模型的有效性。在研究中·所涉及的消费者实体可以具有各种粒度级别。例如,它可以是一个智能电表(一个家庭),一组智能电表(一个区),一个变电站(城镇或城市)或电站(通常覆盖一个很大的地理位置区)。类似地,所讨论的电力负载的时间单位也具有不同的长度。它可以是5分钟,15分钟,1小时,1天,1周,1个月,1年等。在这项工作中,我们创建一个系统来预测每个智能电表的每日最高负载。另外·需要指出的是,我们在负载预测的研究中使用的框架和技术,也同样适用于不同消费者实体的行为预测。已有的负载预测模型,都有些许的不足,无法满足当前智能电网的需求。通过研究相关的数据挖掘和机器学习算法,我们采用极限学习机来构建负载预测模型。ELM是一种新型的极限学习机作为一类机器学习方法,以简单易用、有效的单隐层前馈神经网络学习算法,受到越来越多的研究者关注。传统的神经网络学习算法(如BP算法)需要人为设置大量的网络训练参数·并且容易产生局部最优解。极限学习机只需要设置网络的隐层节点个数·在算法执行过程中不需要调整网络的输入权值以及隐元的偏置,可以产生唯一的最优解,因此它具有学习速度快且泛化性能好的优点。由于我们预测未来一段时间的最高负载P,我们首先需要对采集到历史电力负载数据进行处理,提取每天负载最大值,作为目标属性,设搜集的天数为N。另外,从外部数据库中爬取相应日期的天气信息,主要包括日最高温、日最低温、月最高温、月最低温、是否是节假日、星期几。因此,对于历史记录中的每个目标峰值负载值·我们构建与目标相关联的特征向量。在特征分析后,我们训练开始训练极限学习机ELM模型,使用N天的特征向量,目标对来完成。在训练完ELM模型后,我们将使用得到的回归模型来预测给定日期的峰值负载值P。为此,我们需要使用相同的方式处理待预测数据·并提取特征相同的特征向量。经过ELM模型,可以得出未来第d天的负载。在训练测试完成后,我们将构建的ELM模型应用到实际的电力系统中,该模型取得了很好的效果。该模型具有重要的意义·如发电机公司·可以根据未来电力负荷的变化来合理的分配能源;另外,电力公司也可以根据负载的变化做出合理的决策。综上所述,在这项工作中·我们提出了一个准确的负载预测模型·可以为智能电网的管理者提供决策支持。在我们的方法中·我们融合跨系统的电力数据,分析影响负荷变化的强关联特征,采用极限学习机(ELM)回归算法构建高精度的预测模型。实验结果表明·我们的方法能够比现有的其他预测方法提供更准确的结果·并且在计算复杂度较低。在将来的工作红·我们打算研究基于自动特征选择的负荷回归模型,以进一步提高其准确性和自适应性。另外,我们计划用不同国家的多个智能电网负载数据测试我们的方法,并对已有的模型进行微调·以确保其通用性。最后·我们将进一步拓展我们的分析框架和预测模型,使其可适用于任何粒度级别(如个体户·街区·城镇·城市和大地理区域)的消费实体负载预测中,并调优模型以取得更好的预测结果。未来的工作还可以继续探索其他综合技术,结合极限学习机模型提高智能电网负载预测的性能。
[Abstract]:The smart grid system based on communication, control and IT technology has now become a global trend. It is an important task for the smart grid to predict future power grid load through customer behavior. Accurate prediction can help utility companies to formulate reasonable resource allocation plans and take control measures to balance power supply and electricity demand. In the competitive power market, power load forecasting is very important for consumers and power producers, which can make consumers understand their own electricity use habits and help manufacturers to formulate specific products according to customer's consumption habits, so as to plan operation and prevent electricity risk. In addition, it is predicted that the power economy is superior to the power economy. In this paper, we have proposed a new data mining framework to analyze customer behavior to predict the load of specific consumer entities in the smart grid in the future. Then, we use the limit learning machine (ELM) to analyze the similarity of the power line of the cluster users and collect the power load of the user. In order to prove the validity of the proposed method, we analyze the validity of the proposed method. In order to prove the effectiveness of the proposed method, we analyze both theory and experiment. The limit learning machine is a new kind of machine learning algorithm, and its random initialization of network node weight and bias strategy can solve the single layer feedforward. In the end, we use the data (including equipment information, line information, user information, load information, etc.) and the external system data (such as weather information) that may affect load changes in the MATLAB platform. The simulation experiment on the MATLAB platform is carried out. The results show that the method can excavate the user's power behavior deeply, improve the accuracy of load forecasting by reasonable user clustering, and reveal the relationship between the prediction accuracy and the number of clusters. With the development of smart grid technology, a large number of advanced measurement infrastructure (AMI) and various kinds of monitoring systems are generated and accumulated. Data. The smart meter is an important part of AMI, which can get the power load of accurate user consumption in a certain time (such as every 15 minutes or every 60 minutes). Compared with the traditional grid system, the intelligent meter collects more data and generates more data. However, the accumulated large data has always been in a shelved state. With the development of the learning algorithm and large data, we can analyze the large data and fully excavate the value behind these data. For example, based on the equipment and customer data in the monitoring system, we use the clustering algorithm to mine the user's electricity behavior, based on the intelligence meter data and the classification regression algorithm, to predict the future load. The change of.[37] load prediction has been the key to the security development of the power system, because it can affect many decisions about power systems, such as economic scheduling, automatic generation control, security assessment, maintenance scheduling and energy commercialization. Accurate load forecasting can start and stop power system generators under economic conditions. The group plays an important role in maintaining safety and stability, maintaining the normal production and life of the society and effectively reducing the cost of generating electricity. In general, the load forecast can be divided into three categories according to the length of the load forecasting time: short-term load forecasting, mid term load forecasting, and long-term load forecasting. 1 hours, one day, or a week. The forecast time range of mid-term load forecast is about the next month. The time range of long-term load forecasting is the next year, or even three to five years. This paper mainly carries out short-term load forecasting for the user's power load. Load forecasting plays a major role in the real-time and security of the energy management system. Accurate prediction is beneficial to the planners of the power system to accomplish various tasks, such as the economic dispatch of electricity generation and the scheduling of fuel procurement. The problem is that load forecasting is a difficult task because the changes are influenced by many factors, such as weather conditions, whether it is holiday, population flow, economic situation, and customer's use of electricity. Accurate or erroneous load forecasting can increase operating costs. It is observed that power demand forecast errors increase by only one percent, resulting in an increase of 10 million pounds in UK power system operation costs. This is a serious error in the type of load forecasting utility. And bad load forecasting misleads planners, causing errors and expensive expansion. Overestimating future power loads may lead to redundant reserves of electricity, and undervaluation of the load leads to sufficient power failures. On the contrary, accurate prediction can make utility providers plan fuel and other resources ahead of time, and take control measures, such as opening / closing demand response devices and revising electricity prices. In contrast, the undervaluation of the load leads to the failure to provide enough power. No matter the planners underestimate or misjudge the load, the high precision load forecasting technology requires advanced technology and adaptive prediction model. Although different models have some advantages in the dynamic system. The possibility of the shortcomings can not be excluded. Therefore, it is necessary to develop the optimal and accurate load forecasting model to improve (minimization) prediction error. By analyzing a variety of data mining algorithms and machine learning algorithms, we are committed to a high precision load forecasting model. It has high efficiency and can prevent over fitting. Therefore, the main research problem of the project is: merging the data of the cross system, preprocessing the data, analyzing the data, mining the strong correlation characteristics that affect the load change; using the limit learning mechanism to build the load forecasting model and adjusting the parameters to get the most accurate load forecasting results. According to the background, there are many factors affecting load change. Power load forecasting is a complex work, which presents complex nonlinear transformation. The traditional power load forecasting model is mostly linear model, and it lacks the ability of nonlinear mapping. Therefore, the previous prediction method is not suitable for the development of the big data age. On the other hand, intelligence The power grid lacks a general training framework to support other tasks of the power system to improve the efficiency of the smart grid analysis. Since 1990, researchers have begun to pay attention to the power load forecasting problem and put forward a number of prediction models, such as the time series analysis model ARIMA proposed by [1]., which focuses on the analysis of the load basis. The change curve of the time to predict the change of the load in the future, its advantage is that the disadvantage is that it does not consider the factors that affect the load change. The fuzzy logic method Fuzzy Logic is used to analyze the change of the load, which takes into account the influence factors of the load change, such as temperature, humidity, etc., but it does not consider the equipment and line information. The artificial neural network method is used to construct the load forecasting model. The neural network is optimized and trained slowly, and it is easier to fall into the local optimal value for the complex load change. Using support vector regression SVR, this method can map the features to the nuclear space. The method can obtain the global optimum but the training efficiency is low. In this paper, a load forecasting model based on the ultimate machine learning ELM is proposed and the validity of the model is verified through actual smart grid data. In the study, the consumer entities involved can have various granularity levels. For example, it can be a smart meter (a family), a group of smart meters (a zone), one A substation (town or city) or a power station (usually covered by a large geographic area). Similarly, the time units of the power load discussed are also of different lengths. It can be 5 minutes, 15 minutes, 1 hours, 1 days, 1 weeks, 1 months, 1 years. In this work, we create a system to predict the daily most intelligent meter. In addition, it is necessary to point out that the framework and technology we use in the study of load forecasting are also applicable to the behavior prediction of different consumer entities. The existing load forecasting models are inadequate to meet the needs of the current smart grid. By studying the related data mining and machine learning algorithms, Using the limit learning machine to construct the load forecasting model.ELM is a new kind of machine learning method as a kind of machine learning method, which is easy to use and effective single hidden layer feedforward neural network learning algorithm. More and more researchers pay attention to the learning algorithm. The traditional neural network learning algorithm (such as BP algorithm) needs to set up a large number of networks. The limit learning machine only needs to set the number of hidden layer nodes in the network. In the process of executing the algorithm, it does not need to adjust the input weights of the network and the bias of the hidden element, so it can produce the unique optimal solution. Therefore, it has the advantages of fast learning speed and good generalization performance. The highest load P for the next period of time, we first need to process the collection of historical power load data, extract the maximum daily load, as the target attribute, set up the number of days to collect the N., from the external database to climb the corresponding date of weather information, mainly including the hottest day, the lowest temperature, the highest temperature of the month, the lowest temperature of the month, Whether it is holidays, weeks. So, for each target peak load value in the history record, we build the eigenvector associated with the target. After the feature analysis, we train to train the limit learning machine ELM model, use the feature vector of N days, and finish the target. After training the ELM model, we will use it. The regression model is used to predict the peak load value P. for a given date. For this purpose, we need to use the same way to deal with the expected data and extract the feature vectors with the same characteristics. After the ELM model, we can get the load of the future day D. After the training test is completed, we apply the constructed ELM model to the actual power system, which is the model. The model has great effect. The model is of great significance. For example, the generator company can allocate the energy reasonably according to the changes in the future power load; in addition, the power company can make a reasonable decision based on the change of the load. In summary, in this work, we propose an accurate load forecasting model. We provide decision support for the managers of the smart grid. In our method, we integrate the cross system power data, analyze the strong correlation characteristics that affect the load changes, and use the ELM regression algorithm to build a high precision prediction model. The experimental results show that our method can be provided than other existing forecasting methods. More accurate results and less computational complexity. In the future work red. We intend to study the load regression model based on automatic feature selection to further improve its accuracy and adaptability. In addition, we plan to test our methods with multiple smart grid load data in different countries and to advance the existing models. Fine-tuning to ensure its versatility. Finally, we will further expand our analytical framework and prediction models to apply to consumer entity load forecasting at any level of granularity, such as personal, block, town, and large geographic areas, and optimize the model for better prediction results. Future work can be followed. Further explore other integrated technologies and combine the extreme learning machine model to improve the load forecasting performance of smart grid.

【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM76

【相似文献】