基于变分模态分解-BA-LSSVM算法的配电网短期负荷预测

配电台区日负荷序列呈现为既包含变化趋势、又含有波动细节的不规则曲线,该文借助变分模态分解(variational mode decomposition,VMD)将包含这些信息的原始日负荷序列分解为不同频率尺度的子序列,并结合一系列复杂的环境因素,分别利用不同的最小二乘支持向量机(least squares support vector machine,LSSVM)模型进行负荷预测,最后将基于不同频率分量的预测结果相加得到最终的日负荷预测结果。为了提高LSSVM预测能力,采用蝙蝠算法(bat algorithm,BA)对各LSSVM的参数进行寻优,同时,该文分析了影响负荷变化的环境因素,设计了一套因素归一化方法,预测过程考虑了环境因素的影响。仿真结果表明,该文提出的考虑复杂环境因素的预测思想及对历史日负荷进行VMD分解、BA优化、LSSVM预测的组合预测方法能有效提高短期日负荷预测的准确性。

Short-term load forecasting for distribution transformer based on VMD-BA-LSSVM algorithm

With the wide application of all kinds of electrical equipment in the distribution system, the power load has increased in recent years, which has a great impact on distribution network. Thus, forecasting the short-term daily load is required. Combining the advantages of VMD, LSSVM and BA, a novel VMD-BA-LSSVM short-term daily power load forecasting method was designed, and the complex environmental factors were considered in this paper. Least squares support vector machine (LSSVM) is a classical machine prediction method, which has the advantages of small sample size, powerful generalization ability and fast solution. However, with the gradual improvement of forecasting accuracy requirements, simple LSSVM can''t guarantee the accuracy of the forecasting work. The daily load sequence of the distribution transformer presents an irregular curve containing variation currents and fluctuation details. These information can be separated and predicted respectively in the prediction process, thus better prediction results can be obtained. Although the daily load sequence seems to be fluctuant and irregular, the trend component and wave components in different frequency scales can be obtained by the variational mode decomposition method (VMD). Compared with the process of recursion and screening in EEMD, VMD is characterized by its non-recursive and variable mode. VMD decomposes the original load sequence into a series of specific band-limited subsequences, which aims to decrease instability. VMD has the better capability of harmonic separation, and each subsequence has a better regularity. In this paper, the VMD was used to decompose daily load sequence of a day and yield a series of subsequences with specific frequencies. Subsequences were put into four LSSVMs for the respective forecast. Different parameters in LSSVMs were optimized by the bat algorithm (BA). Meanwhile, the affection of the complex environmental factors was studied and the normalization approach of those factors was proposed. Thus, complex environmental factors were considered in forecasting. The procedures of this prediction method were as following: Firstly, the input data of the method was the daily load data with a one-hour interval and daily environmental data with a one-day interval of the previous 14 days. The daily load sequence (1 row and 24 columns, 1×24) was decomposed by the VMD method and yielded four low-to-high frequency subsequences. Secondly, the four subsequences of the previous 14 days were combined into four 14×24 matrices. Thirdly, the normalized data of the four matrices and environmental data were put into four LSSVMs to forecast the load of the 15th day. Meanwhile, the parameters of LSSVM were optimized by BA. The last, the four LSSVMs results were summed and yielded the final prediction result. In this paper, the VMD was used to decompose nonlinear, fluctuant daily load sequence and yield subsequences with different frequency scales. Subsequences were combined and put into LSSVMs for the respective forecast. Simulation results showed that the forecasting accuracy of VMD-based forecasting method was higher than EEMD-based method. At the same time, LSSVM was used to forecast, and BA was used to optimize the uncertain parameters. The simulation results showed that compared with SVM, LSSVM had a better capability to approximate the load sequence, and got higher prediction efficiency. LSSVM had less uncertain parameters than SVM, thus the efficiency of parameter optimization was higher. Furthermore, BA had excellent capability of global optimization and rapid convergence. Simulation results showed that the proposed method was the most accurate and efficient method, compared with other five forecasting methods.