基于EMD和SVM的短期负荷预测

Short-term Load Forecasting Based on EMD and SVM

为了有效预测具有一定周期性和随机性的非平稳时间序列性的电力负荷,提高预测精度,提出了结合经验模式分解(EMD)和支持向量机(SVM)的短期负荷预测法。该法运用EMD将负荷序列分解成若干个不同频率的平稳分量,分解后的分量突出了原负荷的局部特征,能更明显的看出原负荷序列的周期项、随机项和趋势项;根据各个分量的变化规律,选择合适的SVM参数和核函数构造不同的支持向量机分别预测;由SVM对各分量的预测值组合得到最终预测值。仿真试验表明,此方法与单一的SVM预测法及BP神经网络预测法相比,具有较高的精度和较强的推广能力。

The power load is inherently non-stationary time series, and has a certain periodicity and randomness by itself, so it is difficult to construct the forecasting model. The traditional models are constructed on the basis of the supposition that the load series are linear and stationary, which cannot predict accurately the real non-stationary load series. In order to predict the short-term power load effectively and level up the forecast precision, a hybrid forecasting method based on Empirical Mode Decomposition （EMD） and Support Vector Machine （SVM） is presented in this paper. EMD can decompose non-stationary signals into some smooth and stationary intrinsic mode functions （1MF） with different frequency in the different scale space by the sifting process. Which is regarded as new adaptive wavelet decomposition method. SVM, a novel machine learning method based on the structural risk minimization （SRM） principle, is powerful for the problem with small sample, nonlinearity, high dimension and local minima. According to the outstanding feature of EMD algorithm, firstly, the power load time series is decomposed into a series of stationary intrinsic mode functions in different scale space via EMD sifting procedure. The local features of original load series are prominent in the intrinsic mode functions so that it is more obvious to observe the cycle, random and trend parts of the original load sequence. Secondly, according to the change regulation of each of all resulted intrinsic mode functions, the right parameter and kernel functions are chosen to build different SVM respectively to forecast each intrinsic mode functions. At last, these forecasting results of each IMF are combined with SVM to obtain final forecasting result. The simulation results show that the hybrid method based on EMD and SVM has faster speed, higher precision and greater generalization ability than that of the single SVM method and that of the BP neural network method, which proves that it is an effective method.