Abstract:

Against the backdrop of high penetration of renewable energy integration and enhanced demand-side flexibility,traditional deterministic load forecasting methods struggle to meet the requirements of risk assessment and decision optimization in power systems.Research on short-term load uncertainty forecasting is conducted.An improved quantile regression neural network model is constructed,adopting a tilted quantile loss function and interval post-processing method to enhance prediction accuracy and reliability.Considering the dynamic variation characteristics of load patterns,an online learning method based on elastic weight consolidation (EWC) is proposed to realize dynamic updates of model parameters.Experimental results show that the proposed quantile forecasting method outperforms traditional methods in terms of normalized average width of prediction intervals while maintaining high coverage,and the coverage is significantly improved after online learning,verifying the effectiveness and adaptability of the proposed method.

Key words: short-term load forecasting, uncertainty prediction, improved quantile regression, online learning