基于Spark和Holt-Winters模型的短期负荷预测方法

发布时间：2018-08-23 14:38

【摘要】：短期负荷预测是电力系统经济运行与系统调度的重要前提和依据。智能电网与传感器技术的发展使得电网数据剧烈膨胀,形成了用户侧负荷大数据。电网大数据形势下的短期负荷预测要求预测方法的预测精度高,计算速度快,传统的预测方法已经无法满足。以Hadoop、Spark平台为代表的大数据处理技术的出现,使得海量数据的处理方式进入一个崭新的阶段。本文分析现阶段大数据形势下短期负荷预测的需求与面临的关键问题,从提高短期负荷预测精度与提高短期负荷预测速度两个方面为切入点,结合相关算法与数据处理,展开大数据形势短期负荷相关工作。论文首先根据短期负荷预测与实验需要,利用NREL实验室公布的负荷数据,生成了约2500万组负荷实验数据集;分析了商业用电负荷与居民用电负荷的特性,得出了两种负荷类型具有周期性与季节性的结论。然后根据负荷数据周期性与季节性的特点,建立能够对季节与周期因素进行建模的乘法Holt-Winters的短期负荷预测模型,并采用无约束优化理论中的L-BFGS内存限定拟牛顿算法实现预测模型的参数优化,以降低预测模型寻优的计算复杂度,并保证负荷预测的精度。接着将文中的短期负荷预测算法在Spark集群并行化实现,以提高计算效率,实现对海量负荷数据的预测。提出了包含2个Master节点,28个Slave节点的集群构建方案,并对集群的配置进行优化,使得Spark计算性能更好。最后对HDFS分布式文件系统存储效率、L-BFGS-Holt-Winters短期负荷预测算法的预测误差和Spark并行计算效率进行实验。结果表明,论文采用的HDFS负荷数据存储,效率优于传统的文件存储方式;论文采用的L-BFGS优化方法有良好的计算效率与优化精度;与传统负荷预测算法相比,论文采用的Holt-Winters算法有更高的预测精度;Spark并行化实现后的短期负荷方法,能够应对海量负荷数据的预测需求。在论文方法在25个计算节点的计算集群中,能够在1.5分钟内实现200万规模的负荷预测,13分钟内实现2000万规模的负荷预测,能够满足大中型城市乃至省级电力单位的负荷预测需求;论文方法能够减少实际电力系统的费用消耗,节约更多的时间,为电力系统的调度调控等操作提供保障。本文基于Spark与Holt-Winters的短期负荷预测方法,是解决大数据形势下海量短期负荷预测一种可行的方案。
[Abstract]:Short-term load forecasting is an important premise and basis for power system economic operation and system dispatching. With the development of smart grid and sensor technology, the data of power grid expand rapidly, and the big data of user side load is formed. Short-term load forecasting under the situation of power network big data requires high forecasting accuracy and fast calculation speed, and the traditional forecasting method can not be satisfied. With the emergence of big data processing technology represented by Hadoop Spark platform, the processing of massive data has entered a new stage. This paper analyzes the demand and key problems of short-term load forecasting under the current big data situation. From two aspects of improving the accuracy of short-term load forecasting and improving the speed of short-term load forecasting, the paper combines the relevant algorithms and data processing. Work on short-term load related to big data situation. Firstly, according to the demand of short-term load forecasting and experiment, using the load data published by NREL laboratory, about 25 million sets of load experimental data are generated, and the characteristics of commercial load and residential load are analyzed. It is concluded that the two types of load have periodicity and seasonality. Then, according to the characteristics of periodicity and seasonality of load data, a short-term load forecasting model based on multiplying Holt-Winters, which can model seasonal and periodic factors, is established. In order to reduce the computational complexity and ensure the accuracy of load forecasting, the L-BFGS memory-limited quasi-Newton algorithm in unconstrained optimization theory is used to optimize the parameters of the prediction model. Then the short-term load forecasting algorithm in this paper is implemented in Spark cluster to improve the computational efficiency and realize the prediction of massive load data. A cluster construction scheme with 2 Master nodes and 28 Slave nodes is proposed, and the configuration of the cluster is optimized to make the Spark computing performance better. Finally, experiments are made on the memory efficiency of HDFS distributed file system and the prediction error of L-BFGS-Holt-Winters short-term load forecasting algorithm and the efficiency of Spark parallel computing. The results show that the efficiency of the HDFS load data storage is better than that of the traditional file storage, the L-BFGS optimization method has good calculation efficiency and optimization accuracy, and compared with the traditional load forecasting algorithm, The Holt-Winters algorithm adopted in this paper has a higher prediction accuracy and can meet the demand of massive load data forecasting by using the short-term load method after the parallel implementation of Spark. In this paper, in the computing cluster of 25 computing nodes, 2 million scale load forecasting can be realized in 1.5 minutes and 20 million scale load forecasting can be realized in 13 minutes. It can meet the demand of load forecasting of large and medium-sized cities and even provincial electric power units, and the method of this paper can reduce the cost consumption of actual power system, save more time, and provide guarantee for the operation of power system regulation and control. In this paper, the short-term load forecasting method based on Spark and Holt-Winters is a feasible method to solve the massive short-term load forecasting under the situation of big data.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM715

【参考文献】