应对可再生能源电力系统中的鸭子曲线：一种基于iTransformer的多任务学习净负荷预测模型

Abstract High penetration of renewable energy leads to dramatic changes in regional load patterns, forming a duck curve phenomenon that profoundly affects the operation style of the power system. To obtain accurate prediction on the duck curve scenario, the paper proposes a day-ahead net-load forecasting method based on iTransformer and multi-task learning by considering various individual resource components, such as PV power, wind power, and active load. Firstly, the dominant features of individual prediction task are identified through a combined feature selection approach. Subsequently, iTransformer is employed as the backbone network to construct the forecasting model with robust temporal dependencies learning capabilities. Additionally, the iTransformer combined with multi-task extracts multi-correlations knowledge among external factors, individual power, and integrated power in net-load forecasting. Finally, a customized loss function tailored for the multi-task model is constructed by analyzing the periodicity and volatility of individual components. To validate the effectiveness of the proposed forecasting method, a case study is carried out on four practical datasets of different regions and scales, achieving 14.30%, 19.46%, 15.93% and 14.84% RMSE decline compared with integrated forecasting. The results show that the proposed method acquires high accuracy and stable performance on the duck curve scenario.