Power system source-load forecasting based on scene generation in extreme weather

Renewable energy generation featuring high penetration rates is highly susceptible to extreme weather. Meanwhile, extreme weather events, though low in probability but high in impact, often exhibit characteristics of imbalanced sample data. To address the above issue, a source-load forecasting model based on conditional generative adversarial nets (CGAN) and a multi-gate mixture-of-experts (MMoE) multi-task learning structure is proposed. First, the extreme weather indicators, which are proposed based on meteorological data, the characteristics of source-load data, and the maximum information coefficient (MIC), are utilized as the conditional information of CGAN. Subsequently, meteorological data, source-load data, and calendar rules are employed as inputs to generate extreme weather samples via the application of a CGAN. Moreover, a Dropout layer is introduced to the CGAN to avoid the problem of model collapse. Finally, an MMoE multi-task learning structure forecasting model with bi-directional long short-term memory (BiLSTM) as the tower network unit is constructed, and the improvement of forecasting accuracy is achieved by considering the coupling between source and load. Experimental results show that the model proposed in this paper outperforms traditional data augmentation methods and forecasting models under extreme weather conditions.