基于证据理论的多模型组合电价预测

Electricity Price Forecasting Based on Multi-models Combined by Evidential Theory

传统的组合预测方法主要依靠历史预测误差确定组合权值,预测结果受历史预测误差影响较大,没有考虑预测时点的外界影响因素。利用支持向量机,神经网络和时间序列等多种不同的预测方法,从不同侧面对电价进行预测。利用神经网络等模型对预测的历史误差和预测时点的外界影响因素进行分析建模,建立每个模型的可信度评价模型。采用DS(Dempster-Shafe)证据理论对每个模型的可信度进行分析评价和合成,确立最终的模型组合预测权值。通过该权值对相应的预测结果进行加权求和得到最终的预测结果。以加州电力市场为例,证明了该方法的有效性。

The combined weights in traditional combined method for electricity forecasting are obtained with calculating the historical forecasting errors, with no considering of the environmental factors. Five models from support vector machine （SVM）, artificial neural networks （ANN）, and time series forecasting techniques were selected to forecast the electricity price from different views. Four models from ANN and SVM were selected as experts to evaluate the credit of forecasting results of the five above models, with historical forecasting errors and environmental influence. The credit were combined to calculate the weights with Dempster-Shafer （DS） evidential theory. The final forecasting was obtained by the weighted forecasting. The experiment of California power utilities validates the proposed method.