基于改进GAN模型的图像化短期负荷预测方法研究

Research on image⁃based short⁃term load forecasting method based on improved GAN

为解决短期负荷预测中长序列预测精度不足的问题,提出一种基于改进GAN(生成对抗网络)模型进行短期负荷预测的方法。该模型以图像修复算法GAN为基础,结合通道、空间注意力机制和多尺度卷积进行图像修复。该模型将一维负荷序列编码为RGB图像。将图像化负荷序列中需要预测的部分进行mask处理,将序列预测问题转换为图像缺失像素的修复问题。将修复图像恢复为负荷序列,通过XGBoost(极限梯度提升树),结合气温、节假日、时刻等特征值进行修正。通过实际算例显示,所提出方法的MAPE为2.45%,与传统预测方法相比,在精度上为最优。

In order to solve the problem of insufficient accuracy of long-term series prediction in short-term load forecasting,a short-term load forecasting method based on improved GAN(Generative Adversarial Network)model was proposed.The model is based on the image inpainting algorithm GAN,and combines channels,spatial attention mechanism and multi-scale convolution for image inpainting.The model encodes a one-dimensional load sequence into an RGB image.The part of the image-based load sequence that needs to be predicted is masked,and the sequence prediction problem is transformed into a repair problem with missing pixels in the image.The repaired image is restored to a load sequence,and then corrected by XGBoost(Extreme Gradient Boosting Tree)combined with characteristic values such as temperature,holiday,and time.The practical example shows that the MAPE of the proposed method is 2.45%,which is the best accuracy compared with the traditional prediction method.