基于支持向量机混合模型的短期负荷预测方法

将支持向量机专家系统混合模型应用于短期负荷预测采用方法分为2个阶段:应用神经网络中的聚类算法将历史数据分割成不相连的数据域;对每个数据域选择最佳核函数预测单个SVMs。实际数据验证表明,该方法与单个多项式核、高斯核和3次样条核的SVMs预测相比具有预测精度高、支持向量少和计算量小等优点。

Short-term Load Forecasting Based on Support Vector Machines Mixture Models

In this paper,support vector machines(SVMs) expert,a mixture structure,is presented to forecast short term load in power systems.Support vector machines(SVMs) expert has a two-stage neural network architecture of SVMs,In the first stage,self-organizing feature map(SOM) is used as a clustering algorithm to partition the whole load input data into several disjointed regions,and the training data with similar features in the input space are belonged to the same region.Then,in the second stage,the multiple SVMs,also called SVM experts,that best fit to forecast short term load in the partitioned regions are constructed by finding the most appropriate kernel function of SVMs.SVMs is an advance learning algorithm which are based on statistical learning theory(SLT) and have perfect mathematical theory basement.SVMs are robust to over-fit learning problems and white noises,and have good generalized ability to solve many kinds of time series forecasting problems.Meanwhile the optimal solution of SVMs is absolute unique and optimal in whole feasible solution space,and is superior to artificial neural networks.So,in recent years,SVMs have been used to forecast short term load in power systems in many literatures.However,the short term load data change periodically,it is not appropriate using SVM for forecasting directly,So we proposes Support vector machines expert using a mixture structure in this paper.Compared to single SVMs models with polynomial kernels,Gaussian kernels and the third order spline kernels,the proposed algorithm in this paper have three advantages,i.e.more accurate forecasting,fewer support vector machines and smaller calculation complexity.Finally,using actual short-term load data of one city in south china as training data,the experiment shows that the SVMs expert's forecasting precise is enhanced 10 times and calculating complexity is descend 100 times.