基于GARCH误差校正的遗传支持向量机日前电价预测

Day-ahead Price Forecast with Genetic-algorithm-optimized Support Vector Machines Based on GARCH Error Calibration

针对时间序列预测和智能算法预测各自的侧重点不同,结合两者优点对日前市场电价进行预测。首先建立支持向量机(SVM)模型对单一时点电价进行预测,将遗传算法(GA)嵌入SVM模型中来保证SVM参数选择最优。针对SVM-GA模型训练误差和测试误差存在一定的相关性和条件异方差性,采用广义自回归条件异方差(GARCH)模型对误差序列进行拟合。然后利用拟合好的GARCH模型对SVM-GA模型预测误差进行预测,最后根据GARCH预测结果对SVM-GA模型预测进行校正。用该方法对美国PJM电力市场2005年8月份日前电价进行连续预测,总体平均误差仅8.19%,比普通方法误差减少了将近4个百分点。

A new method incorporating the time sequence modeling and intelligent algorithm modeling is presented to forecast the day-ahead electricity price. With genetic algorithm （GA） adopted to optimize the model＇s parameters, support vector machines （SVM） model is applied to forecast the price sequence. The generalized autoregressive conditional heteroscedasticity （GARCH） models are applied to adjust the error series of price forecasted by SVM-GA models, eliminating their autocorrelations and heteroscedasticity effects. A case of forecasting the day-ahead price of PJM market in August, 2005 demonstrates the proposed method has a desirable performance with an overall mean absolute percentage error （MAPE） of 8.19 percent, which is nearly 4 percent less than data forecasted by common methods.