Abstract: 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 class Ⅱ cells, then establishes a data generation model based on original GAN, and generates sufficient "class Ⅱ cell 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 "class Ⅱ cell historical load data" and the constructed RCGAN model with parameters determined. Engineering examples show that the method is correct and effective.