实现影响因素多源异构融合的短期负荷预测支持向量机算法

针对智能电网大数据环境下,导致电力系统负荷波动的诸多因素存在多源异构性的问题,利用多核函数来对其多源异构特性进行差异化处理和融合,能够描述影响因素的内在分布特性并应对其变化,提高负荷预测精度。选取历史负荷、气温、气压、相对湿度、降雨量、风向、风速、节假日及电价9个属性作为多源异构影响因素,利用样本特征分布法、单变量法及核矩阵秩空间差异法来选择多核函数的构成,采用双层多核学习算法,建立了并行化多核支持向量机(SVM)负荷预测算法流程,并在Hadoop集群上进行了仿真验证。仿真结果表明,多核SVM比单核SVM预测平均相对误差小,双层多核学习、基于lp范数的多核SVM模型预测精度最高。因此,多核SVM能有效处理负荷预测中的多源异构数据,经并行化处理后,能提高负荷预测的速度与精度。

Short-term Load Forecasting Support Vector Machine Algorithm Based on Multi-source Heterogeneous Fusion of Load Factors

A method to select optimal multiple kernels developed from multiple kernel function is proposed for short-term load forecasting in the big data environment of smart grid, multi-source heterogeneous load factors taken into account. The multiple kernel function is able to describe the distribution characteristics of the factors, cope with their variations and improve the accuracy of load forecasting. Load factors such as historical load, air temperature, air pressure, relative humidity, rainfall, wind direction, wind speed, holidays and electricity price are selected as multi-source heterogeneous factors. Three methods(the sample distribution method, single variable method and rank space diversity method)are adopted to establish optimal multiple kernels, and parallel multiple kernel support vector machine(SVM)load forecasting algorithm is based on double layer multi kernel learning algorithm. A Hadoop cluster is built for conducting experiments of short-term load forecasting. Experimental results show that the average relative error of multiple kernel SVM is smaller than single kernel SVM''s, and the accuracy of multiple kernel SVM model based on double layer multiple kernel learning algorithm and norm is the highest. Therefore, multiple kernel SVM can tackle the multi-source heterogeneous data in the load forecasting effectively, and the speed and accuracy of load forecasting can be improved by parallel processing. This work is supported by National Natural Science Foundation of China(No. 51407116)and National Key Technologies R&D Program(No. 2015BAA01B02).