基于渐进式增长生成对抗网络的月度源荷场景生成

Monthly Source and Load Scenario Generation Based on Progressive Growing of Generative Adversarial Nets

月度场景生成是月度时序生产模拟、制定电量计划以及细化场景分析的基础。针对月度源荷场景建模通常面临的多元高维变量拟合困难、源-荷不确定性加剧等问题,提出一种基于渐进式增长生成对抗网络的月度源荷场景生成新方法。对月度时序功率序列进行均值拆分处理,以得到多时间尺度功率序列,将源荷序列纵向拼接并结合多时间尺度特性设计相应的二维卷积结构;采用平滑渐进式增长方式及特殊训练策略,逐步生成多颗粒度的源荷场景;生成网络解构低维与高维特征,首先学习月、周下的日尺度特性,再逐渐拟合高维非线性特征,以生成小时级别的720月度场景。最后,基于实际风-光-荷数据集进行算例分析。结果表明所提算法在月度源荷场景生成的有效性,可为月度调度规划提供参考。

Monthly scenario generation is the basis for monthly time series production simulation,power planning and thinning scenario analysis.Aiming at the problems of difficulty in fitting multivariate high-dimensional variables and in-tensified source-load uncertainty that monthly source and load scenarios modelling usually faces,we put forward a new method for monthly source and load scenario generation based on progressive growing of generative adversarial nets.The monthly time series are split into mean values to obtain the power series on multi-time scale,and the source and load se-ries are spliced vertically to design the corresponding two-dimensional convolution structure in combination with multi-time scale properties.A smooth progressive growth method and some special training strategies are used to gradu-ally generate the source and load scenarios with multi-granularity.The generative network deconstructs the low-dimensional and high-dimensional features,firstly learns the daily scale characteristics under the month and week,and then gradually fits the high-dimensional nonlinear features to generate the 720 monthly scenarios at the hourly level.Finally,an example analysis is carried out based on the actual Wind-Photovoltaic-Load dataset.The simulation results verify the effectiveness of the proposed algorithm in monthly source and load scenario generation,which can be used as a reference for monthly dispatching and planning.