Ultra-short-term multi-step power load forecasting based on generative adversarial framework

Currently, in the process of carrying out power load forecasting, the accuracy of multi-step forecasting is usually much lower than that of single-step forecasting due to the limitation of the model's feature extraction ability and the impact of error accumulation. Therefore, a multi-step forecasting method for ultra-short-term power load based on the generative adversarial framework is proposed in this paper. Compared with the traditional methods, the feature extraction and learning capability of the forecasting model is further enhanced by introducing the training mechanism of generator and discriminator playing with each other in the generative adversarial framework. Specifically, the convolutional-long short term memory neural network (CNN-LSTM) is introduced into the generator to learn complex nonlinear relationship in the time series and generate the load forecast values for multiple future time steps; the multi-layer perceptron is used as the basic structure of the discriminator to judge whether the forecast value is close to the true load value. The analysis results of practical cases show that the proposed method can effectively improve the accuracy of ultra-short-term power load multi-step prediction.