数据匮乏场景下采用生成对抗网络的空间负荷预测方法

The Method of Spatial Load Forecasting Based on the Generative Adversarial Network for Data Scarcity Scenarios

针对在历史负荷数据匮乏场景下,现有空间负荷预测方法预测结果精度较低甚至失效的问题,提出一种基于生成式对抗网络(generative adversarial networks,GAN)和RCGAN的空间负荷预测方法。该方法首先建立电力地理信息系统,并生成I类元胞和Ⅱ类元胞?然后构建基于原始GAN的数据生成模型,根据十分有限的历史负荷数据生成数量充足且兼顾负荷时空分布规律的"Ⅱ类元胞历史负荷数据",达到数据增强的目的。其次构建基于RCGAN的空间负荷预测模型。最后利用生成的"Ⅱ类元胞历史负荷数据"和确定参数的RCGAN模型实现空间负荷预测。工程实例表明该方法是正确、有效的。

Targeting at the problem lacking historical load data and issue that the forecasting method to accuracy of the existing spatial load is low or even invalid under the historical load data shortage scenario,a spatial load forecasting method based on generative adversarial networks(GAN)and recurrent convolutional generative adversarial networks(RCGAN)was proposed.The method first establishes a power geographic information system and generates class I cells and classⅡcells,then establishes a data generation model based on original GAN,and generates sufficient"classⅡcell historical load data"with taking into account load-time and space-time characteristics based on very limited historical load data.Secondly,a spatial load forecasting model based on RCGAN network was constructed.Finally,the spatial load forecasting was realized by using the generated"classⅡcell historical load data"and the constructed RCGAN model with parameters determined.Engineering examples show that the method is correct and effective.