Abnormal Load Variation Forecasting in Urban Cities Based on Sample Augmentation and TimesNet

With the evolving urbanization process in modern cities, the tertiary industry load and residential load start to take up a major proportion of the total urban power load. These loads are more dependent on stochastic factors such as human behaviors and weather events, demonstrating frequent abnormal variations that deviate from the normal pattern and causing consequent large forecasting errors. In this paper, a hybrid forecasting framework is proposed focusing on improving the forecasting accuracy of the urban power load during abnormal load variation periods. First, a quantitative method is proposed to define and characterize the abnormal load variations based on the residual component decomposed from the original load series. Second, a sample augmentation method is established based on Generative Adversarial Nets to boost the limited abnormal samples to a larger quantity to assist the forecasting model’s training. Last, an advanced forecasting model, TimesNet, is introduced to capture the complex and nonlinear load patterns during abnormal load variation periods. Simulation results based on the actual load data of Chongqing, China demonstrate the effectiveness of the proposed method.